The Inference Report

Language models increasingly support or automate information retrieval from document collections too large to index. This task is difficult when it requires specialist linguistic or scientific knowledge. We used the Wikipedia Game to test language models on an analogous information retrieval task. In the game, players must reach a target article using only article links. We compared sentence embedding models and LLMs across five Wikipedia languages and five academic-literature datasets. We construct very large local graphs of Wikipedia pages and of paper citations using OpenAlex and provide code to build and navigate them. This allows the study of language models' navigation capabilities without relying on third-party APIs. We compared the models with human and TF-IDF benchmarks. The Wikipedia Game format enables robust comparisons across techniques, languages, and domains. No model significantly outperformed humans, although the best models (GPT-o3 mini, GPT-4o, QwQ:32B, and DeepSeek-R1:70B) achieved comparable success. Performance and rankings varied across languages and topics, with implications for practitioners utilising these techniques.

Interpreting machine learning model predictions is a central challenge, particularly for black-box vision models. Concept-based attribution methods interpret models using human-understandable concepts. Existing concept-based methods can only interpret classification and regression models. They require access to model weights or large concept-annotated datasets, and most methods cannot provide image-level attribution. We propose Hide and Seek (HnS), a new method for concept-level attribution that does not require access to model weights or labeled concept datasets. HnS can interpret a wide range of vision models. Given a concept and image set, HnS retrieves concept-aligned images using a vision-language model, localizes the concept within each image, masks the localized region, and measures attribution from the change in the black-box model's output between the original and masked images. This enables faithful attribution by directly measuring the model's behavior and supporting local attributions. We validated HnS on CelebA and COCO-Stuff. Compared to existing methods, HnS showed better agreement with ground-truth attributions. We demonstrated the applicability of HnS on diverse tasks, including image captioning and satellite image segmentation, where no existing concept attribution method can be used. We provided our anonymous code in the supplementary material.

The expressivity of a neural network $Φ$ describes the class of functions that a network can express. The term topological expressivity refers to a family of measures particularly suited to understand binary classification and is often formulated in terms of the Betti numbers of a decision boundary $β(Φ^{-1}(0))$. Despite its popularity, this measure is not as useful for regression and other problems that do not involve a decision boundary. Instead, we propose to study topological expressivity in terms of the Betti numbers of the output of a neural network $β(Φ(X))$. We formally prove that a ReLU network with architecture $(d,\sum\limits_{k=1}^{d } β_k(3k+2),\sum\limits_{k=1}^{d}β_k(k+1))$ can express spaces homotopy equivalent to wedge sums of spheres with arbitrary Betti numbers. This gives upper bounds for any arbitrary connected ($β_0=1$) sequence of Betti numbers, but by constructing spaces homotopy equivalent to products of $N$ spheres of dimension $d_1,...,d_N$ we achieve much better bounds $(\sum\limits_{n=1} d_n, 3\sum\limits_{n=1}\limits^N d_n + 2N, \sum\limits_{n=1}\limits^N d_n + N) $ for Betti numbers given by $\prod_n (1+x^{d_n})$. Finally we conjecture a much stronger tight bound which gives the architectural bound $(d,d+1,d+1)$ for a space homeomorphic to the sphere.

Capability evaluations aim to surface dangerous model behaviors before deployment, but their reliability depends on the assumption that hidden capabilities can be elicited. We show this assumption does not hold under adversarial conditions by constructing backdoor attacks that embed encrypted neural networks within host model weights, decrypting and executing them only upon receiving a secret trigger. Using digital lockers, these hidden circuits remain provably hard to elicit and interpret under cryptographic assumptions, even given full access to model weights, extending prior work from hiding fixed strings to arbitrary computations. We implement our proof of concept constructions in PyTorch and evaluate elicitation through fine-tuning, weight noising, gradient-based trigger search, and latent adversarial training on a small chemical reaction prediction task - a proxy for CBRN-relevant capabilities. Our results demonstrate resistance to these methods and, together with our cryptographic guarantees, reveal a limit of elicitation-based capability audits. We release our models to support the improvement of defenses.

A parallel sampler can choose at random which coordinates to generate together. For a target $p$ and the output law $q_U$ of a schedule $U$ drawn afresh for each output, we compare the schedule-averaged divergence $\mathbb{E}_U D_{\mathrm{KL}}(p\|q_U)$ with the output-law divergence $D_{\mathrm{KL}}(p\|\mathbb{E}_U q_U)$, in nats. Rounds commit independent draws from exact conditional marginals frozen at the round's start, each output's schedule is drawn before sampling from a finite mixture of schedules, and every scheduled round is charged. For the uniform distribution on binary strings of even length $n$ with $n/2$ ones, the least expected number of rounds at a fixed tolerance $\varepsilon>0$ is $Θ_\varepsilon((\log n)^2)$ under the first criterion but $(e^{-\varepsilon}+o(1))\log_2\ln n$ under the second, and the double-logarithmic order is attainable with the same round count on every run. For this target, a count-state recursion gives the exact least output divergence at every round cap over all causal randomized policies. The separation extends to weighted fixed-count targets with bounded weight ratios and quota fractions bounded away from zero and one, without permutation symmetry, and to the full-support Beta$(1,1)$--Bernoulli target, where the sharp expected cost under the output-law criterion is $(e^{-\varepsilon}/2+o(1))\log_2\log_2n$, attained with rare $n$-round runs. Prior work implies the order of the schedule-averaged cost on the balanced target and its $O_\varepsilon((\log n)^2)$ upper bounds on both extensions.

Deep-space communications operate under intermittent connectivity, long propagation delays, and relay storage constraints. In such networks, the Bundle Protocol enables store-carry-forward communication, requiring relay nodes to manage limited storage and scarce contact opportunities. Relay-side scheduling and dropping decisions are inherently goal-oriented: transmission resources should be allocated to bundles most valuable to downstream mission objectives. Conventional schedulers, such as FIFO, do not explicitly optimize task utility and may provide unfair service across heterogeneous planetary sources. This paper presents a closed-loop modeling and learning framework for relay-side bundle and flow management in a contact-driven deep-space network, where lunar and Martian updates are forwarded through a shared-buffer relay to Earth. We quantify destination-side information freshness using excess age $Δ_i^e(t)=[Δ_i(t)-d_i(t)]^+$, where $Δ_i(t)$ denotes the Age of Information of source $i$ and $d_i(t)$ is a source-dependent delay baseline. Based on this metric, we define a goal-oriented loss that supports source-dependent nonlinear penalties, allowing prioritization according to mission relevance. We formulate relay control as a constrained semi-Markov decision process (CSMDP) and solve it using masked Lagrangian Proximal Policy Optimization (PPO). Feasibility-aware action masking restricts transmissions to bundles that can complete within the remaining contact window. DTNSim simulations indicate improved age-based utility and inter-flow fairness compared with baseline schedulers. The proposed framework jointly optimizes relay-side bundle scheduling, dropping, and flow control under contact-driven dynamics with explicit fairness considerations across heterogeneous sources.

Clinical sleep medicine repeatedly documents profound divergences between objective sleep duration and subjective daytime impairment, notably exemplified by sleep state misperception in paradoxical insomnia and boundary instability in narcolepsy type 1. Here, we present DyssomniaNN: a multi-scale biophysical computational framework coupling Borbély's two-process regulation of sleep homeostasis and circadian drive with Saper's hypothalamic reciprocal flip-flop circuit and a recurrent corticothalamic leaky integrate-and-fire spiking neural network. We formalize five clinical cohorts across an identical 72-hour photoperiod: healthy control, circadian rhythm sleep-wake disorder (delayed-phase), chronic insomnia, paradoxical insomnia, and narcolepsy type 1. The synthetic local field potential converges closely with empirical human electroencephalography spectra from the PhysioNet Sleep-EDF database across all canonical frequency bands (spectral mean absolute error of 0.33--0.35). Furthermore, we introduce the Well-Rested Index to quantify restorative efficacy. Simulations reveal that sleep duration fails to predict restorative recovery: paradoxical insomnia achieves normal total sleep time (21.15 h) yet yields low restoration (37.6%) from unsuppressed cortical beta-band intrusion, whereas hypocretin deficiency in narcolepsy produces severe nocturnal fragmentation (781 min of wake after sleep onset) and 35 sleep-onset rapid eye movement periods. DyssomniaNN provides an in silico platform for mechanistic sleep research and virtual neuropharmacology.

A central challenge in de novo protein design is generating proteins with desired structural and functional properties. Existing design models typically require a dedicated input representation for each property, limiting the design objectives they support. We develop Prompteina, the first all-atom structure generation model that supports new and broader design objectives specified in natural language, analogous to text-to-image generation. To train Prompteina, we construct Prompteina-Data, a paired text--protein structure dataset of 1.8M structures and 50M+ captions describing global and residue-specific structural and functional properties. We adapt Proteina-Complexa, a pretrained structure generation model, through text-conditioned pretraining and finetuning that aligns residues with their corresponding text. Prompteina outperforms baselines on 15 of 18 structural properties and generates proteins that more closely match descriptions of their intended functions. In binder design, it leads baselines on at least 13 of 20 targets and improves interface confidence by 14\% over the strongest baseline. With iterative in silico feedback, Prompteina's gain over non-iterative generation is 57\% larger than the next-best baseline on the hardest targets. Finally, we experimentally validate Prompteina designs against GM2A, a target with no reported de novo binders. We identify nine binders, with the strongest at an apparent $K_{\mathrm{D}}$ of 107 nM. These results establish natural language as an effective and powerful conditioning interface for protein structure generation.

Recursive self-improvement (RSI) aims to transform interaction experience into persistent improvements, yet external feedback is costly and world-model predictions may become unreliable as policies evolve. We introduce RSI-Anything, a domain-general framework built on JEPA-Anything that uses predictive world models to guide action and learning-data selection. Observed feedback drives policy or workflow improvement and conditional world-model adaptation, while component and compatibility checks govern the retention of updated policy-model pairs. Across diverse computational and scientific domains, RSI-Anything demonstrates substantial gains. In code repair, with the coder fixed, success increases from 26.3% with a static verifier to 29.2% through predictive selection and 31.2% after world-model adaptation, at comparable test usage. Graphical user interface (GUI) evaluations showed higher task completion, while average actions decrease by 39.4%. Across five successive policy versions, laboratory-planning scores more than double from 21.80% to 46.01%. In reaction optimization, RSI-Anything achieves 96.15% mean best observed yield versus 82.16% for Bayesian optimization under the same 40-experiment budget. An end-to-end immune-science application further connects workflow evolution and predictive modeling with therapeutic hypothesis generation. These results demonstrate how predictive world models and retained system updates enable recursive improvement across diverse environments.

Vision-Language Models (VLMs) increasingly analyze and reason about images, yet it remains unclear whether their internal representations faithfully encode the structured relations required for such reasoning. We introduce EICS (Evaluation-Invariant Concept Stability), a metric for testing whether representations consistently encode concepts under controlled one-concept interventions, and CASP-VL (Causally controlled Spatial Perturbations for Vision-Language models), a benchmark of causally controlled spatial perturbations derived from Visual Genome scene graphs. Across multiple VLM architectures, we find that spatial relations are represented substantially less faithfully than lexical information, even when the same models score well above chance on compositional benchmarks. Further analysis shows that the relation is identifiable at the predicate's own tokens but is lost when the sentence is aggregated into a global representation, whether through the (EOS) token or by mean pooling. These findings reveal a systematic failure invisible to output-based evaluation: VLMs do not preserve, in the representation used downstream, relational information that is present in their own input.

While deep learning models for tabular data have recently received significant research attention, the choice of objective functions for regression tasks remains an understudied degree of freedom. Several lines of research offer alternatives to the standard Mean Squared Error (MSE), but differences in datasets, backbones, and evaluation protocols make their practical benefits difficult to compare. In this work, we address this gap by performing a systematic re-evaluation of regression objectives across a broad collection of established tabular datasets and multiple neural backbones. Our comparison reveals that many proposed objectives provide limited or inconsistent gains over a properly tuned MSE baseline, while quantile loss yields consistent aggregate improvements over MSE across the tested backbones. We then identify a limitation that remains even with this strong objective: regions with different levels of target noise can reach their best predictive performance at different training epochs. Consequently, a single global early-stopping checkpoint can impose a compromise between uncertainty regions, even when the objective can represent heteroskedastic target distributions. Motivated by this limitation, we introduce a heteroskedasticity-aware extension of quantile loss that uses input-dependent label smoothing as a potential remedy and improves over quantile loss on several backbones. Overall, our results provide a modern empirical picture of regression objectives for tabular deep learning and highlight how differences in optimal stopping times across uncertainty regions can limit the effectiveness of global early stopping. The code is available at https://github.com/yandex-research/tabular-regression.

We study the last iterate of standard asynchronous Q-learning from a single trajectory of a finite Markov decision process. The trajectory follows a fixed behavior policy whose state--action chain is irreducible and may be periodic. For discount factor $γ$, write $H=(1-γ)^{-1}$, and let $μ_{\min}$ be the minimum stationary state--action probability and $\mathfrak h_\star$ the largest expected hitting time of the reverse state--action chain. We prove that zero-initialized Q-learning with a constant step size achieves sup-norm error at most $\varepsilon$ with high probability using $\widetilde{O}(H^4/(μ_{\min}\varepsilon^2)+\mathfrak h_\star)$ transitions, for $0<\varepsilon\le1$. The statistical term retains the quartic effective-horizon dependence of standard Q-learning, while initialization cost is additive in $\mathfrak h_\star$. For aperiodic behavior chains, a comparison with mixing time yields a transient term $\widetilde{O}(t_{\mathrm{mix}}/μ_{\min})$. An empirical state--action count rule selects a constant step size for a subsequent training run and attains the same rate without prior knowledge of model parameters, target accuracy, or confidence level. The output is the last training iterate, with both stages included in the sample count. The proof uses two deterministic Bellman envelopes to separate statistical error from fixed-policy initialization error and adapts reverse-time hitting-cost compensation from TD learning to state--action propagation. An exploration lower bound holds for arbitrary predictable step sizes. Together with the existing statistical lower bound and its extension to pilot-selected constant steps, it yields matching mean-square rates up to logarithmic factors on the stated large-sample parameter classes.

Neural semantic parsers generate meaning representations token by token, yet well-formedness depends on references, scope, and graph relations that unfold during generation. Constrained decoding prevents locally invalid tokens but does not guarantee that a valid prefix can be completed within a fixed output-length limit. We investigate structural automata at three stages of DRS parsing (decoding, post hoc processing, and training) and additionally evaluate decoding on AMR. For decoding, we introduce accepting-path decoding (APD), which maintains a verified completion within the remaining length limit and follows the model's preferred token whenever that token can also be completed. APD guarantees an accepted output without retracting tokens. Under tight length limits, it improves Smatch++ over local masking by 5.1 and 2.4 points on DRS and by 6.4 and 3.2 points on AMR, comparable to rollback under the same limit, which retracts tokens. For post hoc processing, invalid outputs concentrate on the same inputs: eight unconstrained samples yield no valid output for 10.2% of inputs with limited training data, compared with 0.4% with full data, where filtering comes within 0.2 Smatch++ points of APD. For training, penalizing automaton-rejected tokens improves development Smatch++ by 3.6-11.0 points after 10% of training, but this advantage diminishes as cross-entropy alone reduces the probability assigned to rejected tokens from 14% to below 2% within the first sixth of training with limited data.

Large language models reuse the same attention heads across tasks, but identifying these shared heads does not explain what they do differently in each task. A natural source of inspiration is task-dependent interactions among shared brain regions, which motivate a network view of language-model computation. Here, we compare fixed model components across tasks using identical graph facts, statements specifying which directed connections are present. We measure how interventions affect attention heads and the activation state in later layers (influence), as well as the answer (importance). Task changes redistribute the influence in five open-weight models spanning four families and 1.5-14 billion parameters. In Gemma-2-9B and Qwen3-14B, an activation edit changes an early internal activation vector and switches which fact the model reads while the written instruction stays fixed. The same heads then respond to different graph facts and influence later components differently. For each prompt, we construct two linear response maps, one before the edit and one after it. Each map predicts a later head's response from a measured change in an earlier head. For a separate factual intervention, the map for the corresponding condition reduces mean squared prediction error by 53.6% in Gemma and 5.0% in Qwen compared with the average of the two maps. These findings show that task context changes both which graph facts the shared components respond to and how their outputs affect later computation. A network account may help explain how models use shared parameters to perform different computations. Measurements of task-dependent influence could also guide interventions intended to alter one behavior while preserving others.

As AI agents increasingly rely on external tools, accurate retrieval is essential for reliable behavior. Existing methods optimize semantic relevance but largely ignore tool availability during execution. To address this gap, we propose RATED (Reliability-Aware Tool Retrieval via Execution Dynamics), a retrieval framework that combines robust semantic matching with dynamic reliability adaptation. RATED employs a frozen dual-encoder with contrastive fine-tuning for semantic matching, and a lightweight contextual bandit module that re-ranks candidates based on real-time execution feedback. This design enables rapid adaptation to runtime tool failures without expensive model retraining. We introduce EcoAgent, a benchmark of 5,172 real-world tools and over 210,000 query-tool pairs for offline retrieval and dynamic reliability evaluation. Experimental results demonstrate that RATED achieves a 52.2% relative MRR improvement over the ToolLLM API retriever on real-world queries, and reduces fault selection rate from 86.2% to 2.9% under severe API outage scenarios with a 30% failure rate, while incurring negligible latency overhead of less than 1 ms over encode-only inference. The code and dataset are available at https://github.com/Elliotes/EcoAgent.

Personal agents increasingly connect to services across the web and applications by using tools. As the effects of their actions accumulate and interact across services, unintended or harmful outcomes can arise when agents fulfill benign user requests. Proactively identifying these compositional failures requires systematically exploring diverse cross-service paths. Existing red-teaming methods often simply prompt LLMs to identify potentially harmful paths from task and environment descriptions or decompose harmful objectives specified by humans or LLMs into seemingly benign action combinations. However, generation is biased toward patterns learned during training, potentially favoring familiar tool combinations and leaving novel cross-service interactions underexplored. To address this, we propose X-AgentRed, a self-evolving red-teaming framework that discovers harmful tool call paths through a graph of tool dependencies and expands and refines them using execution feedback. By analyzing each tool's effects, potential harms, and preconditions for execution, the red-teaming agent constructs a directed graph that captures how one tool's effects can satisfy another's execution preconditions. Recipes sampled from this graph are instantiated as realistic scenarios in which agents can produce harmful outcomes. Execution feedback enables self-improvement by diversifying paths that produce harm through tool substitution and refining preconditions at blocked steps in unsuccessful paths to enable further execution. Across four frontier agentic models, X-AgentRed achieves $1.77\times$ the harmful path rate of STAC and explores a more diverse set of tools, yielding $2.99\times$ the tool coverage on average.

A memory that reproduces every current prediction may still have forgotten how to learn from the next observation. We characterize this distinction for exact cumulative learning and attention. For smooth convex residual losses with unrestricted real features and labels, a continuous recurrent state of fixed dimension exists exactly when the loss is polynomial. With ordinary, unweighted observations, degree $D$ requires precisely $r=\binom{d+D}{D}-1$ coordinates; even among histories with one fixed current predictor, exactly $r-d$ coordinates can remain necessary for continuation. Nonpolynomial losses have unbounded requirements on such equal-prediction histories. Bounded scalar logistic examples retain this obstruction after fixing any finite derivative jet. For ridge regression, the hidden evidence is the design matrix: the exact state costs $dp+d(d+1)/2$, versus $dp$ for the predictor. On a well-conditioned zero-prediction family, optimal $B$-bit compression has average one-update error of order $d^{-1}2^{-2B/s_d}$, $s_d=d(d+1)/2$, even with arbitrary randomized codes. Kernel attention gives a second exact closure classification. The results identify what future updates reveal, rather than proposing a new regression algorithm or asserting trained-model gains.

Many optimization problems provide direct access to only part of the gradient, while the remaining component must be inferred from function evaluations. This structure arises in performative prediction, performative reinforcement learning, and bilevel optimization with a black-box lower level. We introduce Partially Informed Stochastic Optimization (PISO), which combines the available gradient component with a running estimate of the residual. By centring each zeroth-order observation on this estimate, PISO makes the variance due to random probing depend on the tracking error rather than the residual's magnitude. This yields a randomized sketch iteration with an explicit contraction mechanism and a principled interpretation of momentum. For smooth nonconvex objectives, we establish convergence guarantees that depend on the residual's variation and initial magnitude, without requiring a uniform bound on the unavailable gradient component. With independently sampled function evaluations and Gaussian or spherical probes, PISO attains leading function-evaluation complexities of \(O(d^{2}\varepsilon^{-6})\) under Lipschitz gradients and \(O(d^{2}\varepsilon^{-5})\) under an additional Lipschitz-Hessian assumption. These rates require only one probe direction, two function samples, and one known-component sample per update. The analysis also quantifies how batching trades observations for fewer deployments. Across five settings, PISO variants improve on the tested baselines with comparable oracle access in three, reach a comparable solution faster in bilevel routing, and perform best among methods without response-model derivatives in performative reinforcement learning.

We establish finite-sample guarantees for conditional diffusion models learned from data and implemented with finitely many denoising steps. Our approach aligns the score-training objective with the sampling grid, allowing statistical estimation and numerical discretization errors to be analyzed together. Under a smooth conditional density model, we derive high-probability bounds on the generated distribution's error and quantify the required number of score evaluations. For every smoothness index $β>0$, a suitable grid yields averaged conditional Kullback--Leibler (KL) divergence $\widetilde O_p(n^{-2β/(d_x+d_y+2β)})$ relative to the original target with or without early stopping. Both the full and early-stopped samplers attain average target total-variation (TV) error $\widetilde O_p(n^{-β/(d_x+d_y+2β)})$. Here $n$ is the sample size, $d_x,d_y$ are the data and conditioning dimensions, and errors are averaged over the conditioning variable. The total-variation bound matches the rate established in existing conditional distribution-estimation theory, up to logarithmic factors.

Generative molecular design typically relies on over-generation and post-hoc filtering. Inference-time steering methods like Feynman-Kac (FK) steering improve efficiency by biasing trajectories toward high-reward candidates, yet they optimize within a single fixed configuration. In de novo nanobody design, upfront commitment to conditioning choices like scaffold identity or target epitope creates a bottleneck. Since yield varies across these conditions, steering unpromising setups wastes compute. We recast generative condition selection as an inference-time resource-allocation problem. We propose two dynamic budget allocators: a joint Global Sequential Monte Carlo (SMC) formulation that steers and allocates compute across configurations concurrently, and a decoupled approach adapting Batched Successive Elimination (BaSE). Empirically, these dynamic allocators increase the yield of designable candidates per GPU-hour nearly 5$\times$ without sacrificing diversity across distinct therapeutic targets and generative configurations.

Coding agents often make useful progress before they fail, but retry-from-scratch discards the machine state created along the way. This makes failure recovery an interesting search problem: if intermediate environment states preserve meaningful progress, restoring and branching from them could outperform repeatedly restarting from the initial state. We implement Go-Explore-style search over full Daytona snapshots and find that higher and more efficient solve rates depend mainly on progress that sits outside the repository, such as databases, services, and other runtime state. Our results suggest that full-snapshot search is most effective when long-horizon coding work spans multiple services whose state cannot be reconstructed from a single diff.

Comics are one of the great visual records of the twentieth century: hundreds of thousands of pages in which a mass readership saw its leaders, its wars and its heroes. Most of this record has never been studied as a whole; comics scholarship proceeds one book at a time. We argue that artificial intelligence, including generative models, can change this to the benefit of the medium itself. We propose PAGE, a framework that makes artificial intelligence an instrument for comics research. It parses every scan into panels, text and a bibliographic record; annotates each page with a vision-language model that follows a written codebook, as a human coder would; grounds samples of these readings in human annotation; and explains change across an archive, with estimates controlled for publisher and genre. We demonstrate the framework on 3,929 American comic books (1938--1956) and 87 volumes of Japanese manga, in three studies, each resting on a different capability. The first pairs text search with visual reading to recover how a mass medium portrayed political power. The second measures the page's visual form across an industry and recovers the hand of individual artists. The third follows who speaks across two comics traditions, checked against human annotation. Beyond research, the same instruments can make archives searchable by what is drawn, open comics to blind and low-vision readers, and help preserve fragile collections and credit uncredited artists. We outline a research agenda in which artificial intelligence strengthens the study, preservation and reach of the medium.

An ideal (language) model must continually update its knowledge. Existing work largely focuses on making these updates efficient after training. We investigate how the organization of time-varying knowledge during pre-training shapes the efficiency of subsequent updates. Models trained on chronologically ordered knowledge specialize in the most recently encountered knowledge; we call them specialists. In contrast, models trained on the same knowledge in randomly shuffled order perform well across all represented time periods; we call them generalists. Surprisingly, specialists require far fewer update steps than generalists to acquire new knowledge. To explain this advantage, we derive a law from first principles that relates update cost to the model's optimization dynamics and that closely predicts empirical results. Guided by these findings, we design a strategy for organizing pre-training data that preserves stable historical knowledge while enabling efficient updates to current knowledge. We also show that the benefits of temporal knowledge organization extend to real-world pre-trained models and existing knowledge update tools, and generalize to other knowledge orderings. Project page: https://horizon-lab.org/specialist/

Continuous diffusion language models (CDLMs) promise to bring the benefits of continuous flow modelling to language, including accelerated sampling and controllable generation. The field has grown rapidly, and the capabilities of CDLMs have begun to rival those of discrete diffusion. Yet, unlike for continuous modalities such as images or videos, our grasp of the levers that matter when training CDLMs remains limited. In this work, we focus on two of these levers, namely self-conditioning and embedding geometry, and shed light on some of the mechanisms behind them, which in turn suggest cheaper and more effective designs. We begin by showing that self-conditioning corrects a train/inference mismatch by exposing the model to inference-like states that are otherwise rare during training. These gains can be recovered without an extra forward pass by a two-level interpolant, coined $\mathrm{DoubleCat}$, which improves TinyGSM accuracy more than threefold. Self-conditioning also aids in recognising tokens dominated by large-norm Gaussians$\unicode{x2013}$an issue known as token identifiability. To address it directly, we introduce discrete reveals to $\mathrm{DoubleCat}$, yielding $\mathrm{SemiCat}$, which achieves discrete diffusion-level performance on TinyGSM, reaching $49.94\%$ with standard training hyperparameters and a tuned sampler. At the $1.7$B scale, $\mathrm{SemiCat}$ outperforms discrete diffusion on four of six likelihood-based evaluations, and surpasses one-hot flows on unconditional generation. On embedded spaces, we find that cross-entropy drives embeddings towards a geometry whose decoding error rate closely matches the one-hot one. We briefly explore a range of further options, such as hybrid corruption processes and noise schedules, including asynchronous ones. We believe this exploration can help the community better understand and advance the design space of CDLMs.

When solving complex spatial reasoning tasks, ordinary humans naturally revisit the scenes from task-relevant viewpoints and mentally reconstruct the spatial layouts from those viewpoints. But can Large Vision-Language Models (LVLMs) proactively think in the same way? We term this ability Active Spatial Mentalization (ASM), an intrinsic component of visual intelligence that remains substantially underdeveloped in current LVLMs. To bridge this gap, we introduce SpatialASM, a scalable framework to internalize ASM into LVLMs through training-time spatial simulation. SpatialASM employs an allocentric novel-view generator to ground hypothetical viewpoints in visual evidence and produces two complementary forms of supervision: i) explicit ASM trajectories that teach LVLMs when and how to adaptively invoke view planning and imagination during spatial reasoning, and ii) automatically-constructed perspective-taking tasks that further strengthens viewpoint-conditioned scene imagination. On seven benchmarks spanning relatively perception-oriented spatial understanding to more challenging tasks requiring multi-image, cross-view and viewpoint-transformation reasoning, SpatialASM yields performance gain on LVLMs of four LVLM backbones in spatial reasoning while retaining their original general vision capabilities. The experimental results demonstrate that our approach improves average accuracy across the seven spatial benchmarks. And the synthesis of ASM trajectories provides a scalable pathway to alleviate the scarcity of spatial Chain-of-Thought (CoT) supervision for eliciting active viewpoint exploration and mental scene simulation.

A central challenge in embodied recursive self-improvement is enabling agents to convert interaction history into capabilities that persist beyond one world and transfer to later tasks. A saved memory or successful final action does not by itself establish that a reusable capability was acquired; evaluation must separate capability change from carried-over resources, unequal source exposure, and early execution failures. We construct MC-EvoBench as a controlled Minecraft benchmark with a fixed catalog of 19 registered evolution chains, including 12 primary evaluation chains and seven calibration or probe chains, together with three graph-aligned paired probe families for transfer. The benchmark represents each task as a task-state graph, defines stage-level environment verifiers, assigns matched source and target roles, resets the target world, and permits only a declared agent artifact to cross the source-target boundary. For each probe, no-prior, related, unrelated, and shuffled source conditions are evaluated under matched environment-step budgets and five seeds; we report TSR, WGP, and VMP and retain provenance records. The benchmark reveals a consistent direct-capability gap between controller and planner-backed variants, while related experience yields no stable transfer: it does not improve either tool probe, its small wood-species gain hinges on one seed, and all intervals include zero. Stage audits further show that early wood-collection and crafting bottlenecks often leave later capabilities unobserved. MC-EvoBench thus offers a controlled testbed for durable, transferable embodied capabilities. Code and data are available at https://github.com/xzw-ustc/mc-bench-open-source

Recent advances in automatic pronunciation assessment (APA) have driven a shift from scripted to free-form assessment. However, existing methods often rely on multi-stage pipelines with forced alignment and separate feature extractors, limiting their practicality across response settings. In light of this, we propose OpenPA, a novel APA framework that leverages Whisper's cross-attention mechanism to jointly predict word- and utterance-level pronunciation scores without forced alignment. Through mixed-transcript adaptation and adaptive substitution weighting, OpenPA accommodates heterogeneous transcripts across closed- and open-response settings within a single model. Trained solely on speechocean762, OpenPA achieves the strongest overall performance among evaluated systems under in-domain closed and simulated open-response conditions, while maintaining its performance zero-shot on MultiPA. An additional cross-task evaluation on oral-cancer speech suggests its potential applicability beyond non-native speech.

Semantic query engines ask a large language model (LLM) to judge individual rows, and benchmarks compare these engines on the quality of a few executions at temperature 0. We measure how much another execution changes. We re-execute the text and tabular workloads of SemBench, LRO-Bench, and UDA-Bench through their own harnesses, log every one of 1.2 million model calls, and reanalyze SemBench's published executions. At temperature 0, between 0.07% and 10.2% of rows change their operator output over five executions, depending on the model, and up to 24.0% on one task. Plans amplify these changes exponentially in the number of rows an answer exposes, as a possible-worlds model predicts: on UDA-Bench, a few percent of extracted values change, but about two thirds of the answers do. In SemBench's published results, 21% of text results change quality between executions, including on engines whose plans are deterministic. Several per-query system comparisons, and 16 of 28 tests of LRO-Bench's implementation recommendations, are not settled by the executions the benchmarks run. Changing rows are rows on which models disagree with each other; identical requests within one execution disagree as often as across executions; and one execution's decision margin or reasoning length predicts which rows change. Voting reduces the variance but does not remove it, and a calibrated cascade keeps its guarantee while the rows it returns change. We derive reporting rules for benchmarks and design rules for engines.

A useful language model should know when to answer and when to say "I don't know." Can we improve that decision by training the model itself or by training a separate component to check its answers? We study both approaches on Qwen3-14B for factual questions, using a score that rewards correct answers, penalizes mistakes, and allows abstention. Across three training seeds and two held-out confidence thresholds, a learned answer selector improves threshold-averaged utility over the original model's calibrated confidence filter. On TriviaQA, mean coverage increases from 20.2% to 30.9%, while the pooled error rate among scored answers decreases from 23.0% to 14.5% under exact-match grading. The tested reinforcement-learning changes do not establish consistent gains. On NQ-Open, the selector improves threshold-averaged utility over the filter but still performs worse than always abstaining. A partial human-reviewed audit preserves the TriviaQA finding. Our results support learning to select answers in this setting while showing that reliable abstention across datasets remains unresolved.

In clinical practice, experts often begin by converting raw inputs into a set of expert features before making a decision. While black-box models trained on raw input can outperform models trained on expert views, it remains unclear what predictive signal the raw model exploits beyond what the expert view captures. We introduce an audit framework to quantify this gap using Shared Canonical Component Removal (SCCR). Unlike standard residualization or concept erasure, SCCR selectively removes only those directions strongly shared between learned raw and expert embeddings. Evaluations on synthetic data, coronary artery disease (CAD) detection, antidepressant treatment dropout prediction, and skin lesion classification using SkinCon annotations show that SCCR identifies the ways in which expert views are insufficient and recovers residual predictive signal while minimizing redundancy with the expert view.

Representations of supervised datasets support higher-level machine-learning tasks such as dataset retrieval and distribution-shift monitoring. However, existing approaches typically rely either on hand-crafted meta-features or on dataset encoders trained specifically for representation learning. We introduce TabMF, a framework that repurposes pretrained tabular foundation models (TFMs) as encoders of supervised datasets. TabMF equips a predictive TFM with an architecture-aware dataset-level readout and adapts the model end-to-end using contrastive post-training on stochastic views of synthetic supervised datasets. We instantiate TabMF on TabICLv2 and TabPFN-2.5 and evaluate the resulting representations across six tasks on unseen real datasets. Simple frozen TFM readouts yield competitive representations, indicating that predictive pretraining already induces substantial dataset-level structure. The two TabMF encoders obtain the strongest aggregate performance among the evaluated learned and classical dataset representations, with particularly strong results in dataset retrieval, compact benchmark construction, and distribution-shift severity assessment.

Generalist robot policies are trained with imitation learning and are bounded by the quality and coverage of their demonstrations. Reinforcement learning can improve and extend them, either by fine-tuning the policy weights or by training a small module to steer the input noise of a frozen generative policy. When tasks arrive sequentially, both approaches suffer from catastrophic forgetting. Fine-tuning on a task degrades prior tasks, and steering only shifts forgetting to the shared steering module. We propose Dream to Remember (D2R), which continually steers a frozen generalist policy toward new tasks, while retaining its earlier skills. As the steering module learns a new task, a frozen pretrained world model rehearses prior tasks through imagined rollouts, anchoring the module to its earlier behavior without a replay buffer or new data. When the world model is faithful enough, D2R learns the new task entirely in imagination. Otherwise, it learns the task in the environment while using the world model only to rehearse prior tasks. On CALVIN, D2R improves eight tasks in sequence entirely in imagination, raising average success from 63.7% to 74.8%, while LoRA fine-tuning on the same sequence drops it to 11.7%. With a frozen off-the-shelf generalist policy and world model, $π_0$ and A2World, D2R learns unseen LIBERO-90 tasks in the environment while keeping $π_0$'s performance on the LIBERO suites it was fine-tuned on, even though the world model is not accurate enough for policy optimization. Finally, on a real Franka robot, D2R improves three tasks in sequence entirely in imagination while retaining the earlier improvements.

Flow-matching and diffusion models are increasingly used as priors for linear inverse problems, however, most methods are bespoke to one class of trajectory prior. Plug-and-play methods have been proposed, however these implicitly decorrelate the guided trajectory, resulting in broad priors that can struggle when data is sparse. This work proposes $β$-Flow: a general plug-and-play method indexed by a correlation parameter $β$ that generalises over flow, diffusion, and decorrelated priors, and a confidence parameter $κ$ that controls how strongly the path distribution is enforced. Each step requires only one forward evaluation of a denoising network, followed by a Gaussian conditioning step. This step is an exact Kalman update that is not dependent on network gradients. We show that many existing guided samplers can be interpreted as instances or approximations of our method, each implicitly using a different trajectory prior and operator in place of the Kalman gain. From this viewpoint, we provide a qualitative comparison of methods in terms of their implicit biases and assumptions. We demonstrate consistently state-of-the-art level performance across 7 inverse problems over 4 datasets in terms of PSNR and SSIM. Moreover, we demonstrate the utility of this prior control, by showing that confident high correlations (strict priors) work best when observations are less informative, while low correlation is preferred when observations are more informative, reflecting the complementary roles of the prior and likelihood in Bayesian inference.

Large language models (LLMs) have shown strong results in classical code generation and are increasingly used to write the unit tests that safeguard that code. Whether they can do the same for quantum software is less clear, because quantum programs are probabilistic, depend on numerical tolerances, and are judged by domain-specific checkers, so a well-formed test can still disagree with the benchmark's verdict. Existing quantum benchmarks measure whether a model produces a correct solution, but not whether its tests can distinguish correct implementations from faulty ones. We introduce QTestBench, a framework that evaluates LLM-generated quantum test suites against independent benchmark checkers, never treating the tests themselves as ground truth. For each task, QTestBench prompts models to produce syntactic, semantic, and behavioral tests, injects candidate solutions at runtime, and compares each test verdict with the checker's label under three validation cohorts, each answering a distinct diagnostic question. We benchmark five LLMs on eight PennyLane QHack challenges and ten Qiskit HumanEval tasks. Generated suites almost always execute, but discrimination lags behind: the best generator reaches a test quality of 0.71 on QHack and 0.34 on Qiskit HumanEval, so executability alone overstates test quality. Oracle-verified synthetic mutants stabilize bug-detection estimates when natural faulty solutions are scarce, as on three of ten Qiskit tasks that yield none. QTestBench offers a reproducible, cross-framework benchmark for the trustworthiness of LLM-generated quantum tests.

Large language models (LLMs) are now a standard source of query expansion terms but they hallucinate. They inject non existent entities and off-topic terms that add ranking noise and can hurt strong retrievers. Existing remedies such as corpus grounding, mutual verification and preference alignment treat hallucinated terms as an undifferentiated nuisance to be filtered by agreement heuristics, without a mechanistic account of why a given term appears or a principled criterion for how many terms to keep. We propose Subsequence-Association Filtered, breadth-aware query expansion (SAFE-QE), a training-free, retriever agnostic method that draws on the subsequence-association account of hallucination. We score each candidate expansion term by its corpus conditioned cross-context reproducibility, terms that reappear under randomised query contexts yet lack support in the pseudo-relevant set are flagged as products of dominant hallucinatory associations and down-weighted. Drawing on the characterisation of language generation as a trade-off between consistency (no hallucination) and breadth (coverage), we cast term selection as constrained coverage maximisation over the pseudo-relevant set, keeping a set that is approximately broad but consistent and stable across resampling. We also introduce the Hallucinated Expansion Term Rate, a query side faithfulness metric that counts injected expansion terms no relevant document supports. We evaluate our approach on benchmarked TREC and BEIR collections. Our SAFE-QE is statistically indistinguishable from generative and corpus-steered expansion on nDCG@10 while injecting fewer unsupported terms 4.3 per query against 27.8 for Query2doc and 57.5 for HyDE and on a strong learned sparse retriever, where every expansion method regresses. Faithfulness is therefore available at no measurable cost in ranking quality.

Variational quantum circuits (VQCs) are highly sensitive to circuit architecture, yet most quantum machine learning pipelines select a single best design and discard the remaining candidates. This paper addresses whether the structural diversity across quantum circuit search spaces can be used to produce stronger predictions through model fusion. We apply Combinatorial Fusion Analysis (CFA), a model fusion framework that selects and combines multiple scoring systems using rank-score function (RSF) and cognitive diversity (CD), to a pool of 29 architecturally diverse classical models and quantum circuits. Using a diversity-aware model selection method, we identify subsets of maximally diverse models and combine them with both score and rank combination. Experiments on four binary classification benchmarks covering different domains demonstrate that CFA-based hybrid classical-quantum fusion consistently outperform the best individual model on every dataset considered with improvements up to 6.03% in accuracy. When comparing with relevant literatures, the CFA framework further achieves superior performance across all four datasets.

Data augmentation is standard practice for low-resource automatic speech recognition (ASR), but its benefits are almost always reported as aggregate error rates, leaving open why and when a method helps. On four low-resource languages and English at five data sizes, we cross two augmentations that modify existing audio (speed perturbation and SpecAugment) with two that modify text transcriptions and synthesize audio from it. We analyze errors at the phoneme level and find that they help for complementary reasons. SpecAugment has the most reliable aggregate improvements, but increases tone and palatalization errors, whereas text-based methods introduce consonant insertions; in combination, each reduces the errors the other introduces. This complementarity is predictable. Tone and palatalization are each signaled by a single acoustic property and occur least frequently in training (lowest training exposure), and text-based methods that flatten word frequency raise their frequency by 37--38%, reversing those errors while further improving aggregate error. The improvement from combining the methods exceeds the sum of their individual improvements; the superadditivity vanishes by 100 hours, limiting the complementarity to the low-resource regime. Where specific error types matter, we recommend selecting augmentation by training exposure and acoustic properties, both measurable before any training.

Production recommender systems have feature-lifecycle asymmetry: adding new features must produce gains on some metric, while feature removal must not regress any metric. In practice, feature deprecation is hard to prioritize as a feature can fail to be removed if it interacts with other features, its importance changes daily, it cannot be evaluated before production models change or data drift occurs, it is in contention for A/B test space with other launches, it causes regressions in an online A/B test, or it is consumed by multiple models that need to coordinate its deprecation. We present FORTE, a model-agnostic production framework for discovering, validating, and removing features. Its efficient perturbation implementation allows daily feature evaluation. To account for interactions, it evaluates candidate feature sets constructed from all the low-rank features through grouped ablation and its offline results are sufficiently predictive of online A/B tests to allow deprecations to proceed to online testing with confidence. Lineage tracing identifies all downstream consumer models to realize the full savings of removing a feature from the training set. We validate FORTE over 2 large-scale recommender systems and find that it identifies deprecable feature sets which do not cause online regressions, while the impact on efficiency depends on model characteristics. On a search-relevance model, FORTE resulted in 3 nested candidate sets of 32, 48, and 59 features. Online guardrails rejected the largest treatment. The 48-feature set was selected for deprecation, removing 43.3% of the model's features. On a large ranking model, FORTE identified 8 candidates for removal. Retraining and A/B testing confirmed no regression. Its launch reduced the trainable parameters by 29.9%, improved serving throughput by 19.8-25.5%, and reduced annual training costs by hundreds of thousands of dollars.

Gaussian process solvers for linear partial differential equations (PDEs) return posteriors rather than point estimates, but they condition only on equality information: the PDE residual, boundary data, and measurements. Many solutions of interest also obey inequality constraints, including nonnegativity, monotonicity, and absence of arbitrage, which equality conditioning cannot enforce. We prove that conditioning a Gaussian process prior jointly on linear equality observations and a finite-dimensional inequality event is equivalent to conditioning sequentially in terms of Borel measures on the solution space. The physics-informed posterior is therefore a valid prior for the inequality update at virtual points. The combined law is non-Gaussian yet computable through truncated-Gaussian moments, and neither the PDE conditioning nor the inequality update requires modification. We evaluated the construction on a Poisson benchmark, option-price surfaces, and air-quality fields. Under incomplete boundary data, a monotonicity constraint lowered the error of a two-dimensional Poisson reconstruction by $63\%$. On a real S&P 500 option chain thinned to $20$ retained quotes, no-arbitrage constraints reduced held-out extrapolation error by $22\%$ and cut the calendar-arbitrage rate from $10\%$ to $1\%$. On four real air-quality snapshots, nonnegativity reduced a Gaussian moment-based sub-zero-tail proxy by $74\%$ at the sparsest monitor count. Changes in mean accuracy and constraint diagnostics differed across data regimes. The probability of the noisy virtual constraint event under the equality posterior measures agreement with the specified bounds before the update. The construction turns admissibility from a post-processing step into conditioning information for probabilistic PDE solvers.

Variance reduction improves the convergence rate of stochastic non-convex optimization under the average smoothness condition. Existing adaptive variance reduction methods are especially designed for average smoothness and do not adapt to standard smoothness. In this paper, we study how one algorithm can adapt to these two settings and the noiseless setting without knowing the smoothness constants or the variance bound. Given the gradient budget $T$, we develop a two-phase method. The first phase compares two independent gradients and switches to normalized momentum when their difference exceeds a threshold. This phase provides both a candidate with standard smooth and noiseless guarantees and also ensures a bounded expected function gap for the second variance reduction phase. The resulting rates are $O(T^{-1/3})$ under average smoothness, $O(T^{-1/4})$ under standard smoothness, and $O(T^{-1/2})$ without noise. When $T$ is unknown, we propose a method with two paths that attain the corresponding rates with logarithmic factors in the stochastic settings. Here, we use a variance reduction path to deal with average smoothness and employ the AdaGrad-Norm for standard smoothness and noiseless setting. Neither method restarts from the initial point or requires a growing number of oracle calls in each round.

Detecting communicative meaning in dialogue requires modeling not only lexical content but also the participant-role structures through which sentiment, emotion, politeness, and formality (SEPF) are expressed. Existing approaches rely on contextual language representations while leaving semantic roles implicit. We introduce a role-aware transformer architecture that maps the four SEPF dimensions to theory-grounded Semantic Role Labeling (SRL) configurations, incorporating these configurations directly into the transformer's self-attention mechanism. Role-aware attention models non-directional internal states (sentiment and emotion), while directional role-aware attention captures initiator-to-recipient relations for external cues (politeness and formality). The resulting role-aware representations are fused with contextual embeddings through a learned gate that adaptively weights how much role information each input contributes, rather than appending semantic roles as an auxiliary feature vector. Across four benchmark datasets, the framework yields statistically reliable gains over strong BERT-based baselines for emotion, politeness, and formality, with gains concentrated in theoretically predicted cases such as low-frequency, complex emotions and highly polite expressions; sentiment improvements are marginal and non-significant, suggesting that sentiment is conveyed more through lexical choice than participant-role structure. Additional evaluation on real-world data demonstrates that the framework improves both predictive performance and model traceability. These results demonstrate that linguistically grounded modifications to transformer self-attention improve the detection of communicative meaning, generalize beyond benchmark datasets to real-world mental-health data, and yield traceable, role-grounded explanations that extend beyond predictive performance to interpretability.

Discourse particles subtly shape the flow of discourse by reflecting conversational context and interlocutor attitudes, often leaving the truth-conditional meaning untouched. They are also often polyfunctional, with distinct senses which partially overlap with other particles; for example, the exclusive use of English "just" and "only" ("John only/just drinks water"). Despite the frequency of these words in natural language, they have gone largely unstudied due to the subtle nature of their effects. We investigate LLM sensitivity to "just" and "only" in natural conversations using behavioral and mechanistic analyses. First, we find that LLMs and humans similarly prefer the presence of these particles in context, and both are more sensitive to "just" compared to "only". Second, when sampling from LLMs with and without these particles, we find LLMs consistently rate generations as more likely, given their source context, demonstrating their particle sensitivity. In both experiments, instruction tuning increases sensitivity. Finally, we localize particle-sensitive units within LLMs and find the overlap between "just" and "only" is asymmetric: "only" units transfer selectively to exclusive "just", while all "just" units transfer to "only" equally well regardless of sense, suggesting LLM lexical-level particle representations are partially organized by sense. We release our code, sense-labeled dialogue stimuli, and human preference annotations at https://github.com/JiahuiKChen/Not_Just_an_Evaluation.

How well can vision-language models (VLMs) reason about what is missing from an image? We study this question with OccludeBench, a manually curated benchmark for recovering fully occluded objects from the visual evidence that remains. OccludeBench distinguishes direct cases, where the hidden object must be recovered from a specific physical cue such as a shadow or reflection, from inferential cases, where it can be inferred from broader scene context. Across five VLMs, accuracy averages 42.7% on direct cases versus 62.7% on inferential cases, exhibiting a 20.0-point gap. Humans attain higher accuracies of 63.4% versus 68.5% for direct and inferential images, respectively, with no significant gap. To locate this failure, we introduce OccludeTrace, a four-stage diagnostic that isolates where the visual reasoning process breaks down. The bottleneck emerges when models must recover the relevant object from the projection within the full image: they succeed on only 50% of failed examples, compared with 91% of examples they answer correctly. Extended reasoning, tool use, and few-shot prompting yield inconsistent improvements and do not reliably eliminate the gap. In contrast, an inpaint-then-confront intervention, which prompts models to first generate a candidate hidden object and then evaluate that hypothesis against the visible physical cue, reduces the gap to a statistically non-significant difference for both models with image-generation capabilities. Together, our results establish a tractable, controlled diagnostic for isolating how VLMs reason about visual content that is missing from an image.

Momentum methods such as STORM and second-order momentum track the gradient by adding a transport correction to momentum, at the cost of an extra oracle call per iteration. We introduce randomized momentum transport, which evaluates the correction with probability $q$ and otherwise uses a cheap prediction, with importance weighting that preserves the conditional mean and adds $(q^{-1}-1)$ times the prediction loss to the variance. Under transport-bias, second-moment, and residual-control assumptions, linear minimization oracle (LMO) updates, including normalized, sign, and exact spectral (Muon) directions, retain an $O(T^{-1/3})$ stationarity bound for each fixed $q>0$ with a single-gradient initialization. For unnormalized STORM, some $q<1$ lowers the leading coefficient of our sufficient gradient budget if and only if the gradient-difference noise modulus exceeds the smoothness constant, by at most about $10\%$. An online diagonal predictor admits a regret-to-convergence guarantee. At approximately equal oracle cost, querying a quarter of the corrections lowers the final training loss by $18$ to $43\%$ on MNIST1D and raises the test accuracy of Hessian-vector transport on CIFAR-10 by $1.5$ points.

We present new algorithms for stochastic optimization problems in which the objective is neither smooth nor convex, subject to a convex constraint set. Our algorithms return a $(δ,\varepsilon)$-Goldstein stationary point, as recently defined for constrained problems by Liu et al. (2024), with improved complexities, and are the first to achieve optimal convergence rates for this task in several settings. In particular, we design projection-based and projection-free algorithms that achieve a $\widetilde O(δ^{-1}\varepsilon^{-3})$ rate using stochastic gradients, which we show to be optimal, and a $\widetilde O(dδ^{-1}\varepsilon^{-3})$ rate using zeroth-order queries, which is also optimal in its dependence on the dimension $d$ and improves over prior work. Notably, our lower bound is given in terms of the Frank-Wolfe gap and holds even for smooth problems, which is of independent interest. These results show that in nonconvex stochastic optimization, the convergence rate is degraded neither by the presence of constraints nor by nonsmoothness. Our analysis is based on a reduction to online linear optimization, extending a related technique from unconstrained optimization.

Discrete graph diffusion models (GDMs) corrupt the graph structure by adding noise independently to nodes and edges under a Markovian noise model. In the unconditional setting without node features, an equivariant denoiser trained on a single graph is able to generate novel graphs and sample-level metrics such as edge overlap and WL kernel similarity find no evidence of memorization. In this work, we challenge this view and show for the first time that graph diffusion models indeed exhibit memorization. Analyzing the reverse diffusion trajectories of GDMs trained on two-block stochastic block models (SBMs), we find a strong train-data bias: the model recovers a corrupted training graph almost perfectly compared to a held-out test graph. We term this conditional memorization, as it surfaces only when the reverse trajectory is conditioned on a corrupted training graph and is not apparent during inference. It coincides with a growing generalization gap, where the training loss improves as the loss on held-out graphs degrades, and it cannot be mitigated by early stopping. The onset of memorization always precedes the emergence of community structure in the generated graphs, thus every checkpoint earlier produces structureless graphs. This behavior holds across all node sizes we test.

Memes play a central role in modern online communication, especially in contexts of political conflict where visual and textual cues combine to convey persuasive messages. We study persuasive technique detection in Ukrainian memes using vision-language models (VLMs). We introduce UkrMeme, a dataset of 200 organically produced Ukrainian wartime memes annotated for persuasion techniques, and PTM-UA, a Ukrainian translation of a subset of the Propaganda 950 benchmark that preserves original visual content while translating meme text. We evaluate VLMs in a zero-shot prompting setup across image, text, and multimodal inputs using hierarchical F1 metrics. Finally, we compare Layer-wise Relevance Propagation (LRP) and model-generated token attributions and analyze the semantic categories of evidence emphasized by each explanation method.

State-space models (SSMs) maintain a fixed-size recurrent state, but reading and writing it at every decoding step can bottleneck hybrid models. This eager state management causes large state I/O with low arithmetic intensity, leaves speculative decoding no cheap rollback, and feeds state quantization error into subsequent recurrent updates. We introduce ReplaySSM, which unrolls the recurrence to represent the state as a checkpoint and a buffer of recent SSM inputs. This separates output computation from checkpoint updates. By updating the checkpoint periodically, ReplaySSM avoids state writes on most steps and roughly halves the dominant state memory traffic. Keeping recent inputs also makes speculative rollback a buffer-pointer operation and enables output-only verification without materializing intermediate states. The same separation allows a low-precision checkpoint to serve output computation on most steps while a higher-precision checkpoint serves recurrent updates, so rounding error in the low-precision copy never enters later updates. We implement ReplaySSM in vLLM for Mamba-2 and Gated DeltaNet. At serving batch size 256, it accelerates standard decoding by $1.20$--$1.47\times$ over a tuned vLLM baseline across four hybrid models, and delivers about $1.8\times$ the throughput of vLLM's speculative decoding on the two 120B-class models. Separate state precisions also reduce output KL divergence and downstream accuracy losses compared with a shared low-precision checkpoint.

Multiple-gradient descent is typically formulated in Euclidean geometry, whereas matrix-aware optimizers use directions induced by other norms. We analyze a single-loop multi-objective method combining a norm-constrained linear minimization oracle with entropic updates of task weights. For any finite-dimensional norm pair, we establish an $O(T^{-1/2})$ ergodic stationarity rate in the deterministic setting, deriving the required trajectory gradient bound from smoothness and lower-boundedness in the unregularized case. With stochastic gradients and a bounded-gradient assumption, one persistent blended-gradient momentum buffer yields an $O(T^{-1/4})$ rate whose leading term has no explicit dependence on the number of tasks. We also establish static regret for the inner weighting objective on arbitrary bounded gradient sequences and, with entropy regularization and a weight floor, iteration-wise inner-gap control after burn-in while retaining ergodic stationarity, conditional on a trajectory gradient bound. The optimal blended gradient can be nonunique in spectral and sign geometries, motivating a selection-independent inner gap. Our analysis extends the spectral construction studied by concurrent MOON to general norms. Experiments compare the method with existing multi-task optimizers and examine its oracle, weighting, and momentum choices.

Multi-agent reinforcement learning (MARL) has become a prominent research direction, particularly within collective robotics, where the ARGoS simulator is widely used to study emergent swarm behavior. Yet despite this growth, systematic support for reuse and reproducibility in MARL-based swarm robotics remains limited and insufficiently standardized: a single experiment is assembled from manually aligned simulator files, training scripts, configuration values, paths, seeds, and environment details, so reuse and reproduction stay fragile even within one research group. To address this gap, we introduce MARGoS, a lightweight, configuration-driven platform developed in this project that elevates these experiment-defining inputs to first-class, platform-managed artifacts. MARGoS enables reproducible experimentation through configuration hashing, end-to-end seed propagation, frozen (immutable) configurations, structured logging, environment fingerprinting, and exportable experiment bundles that facilitate sharing, replay, and verification of results across machines and over time. A five-dimension evaluation (architecture, efficiency, reproducibility, usability, and collaboration) supports our claim that managing experiment variation as a platform responsibility makes ARGoS+MARL experiments easier to reuse, reproduce, and hand off. Our code is available at https://github.com/tombackert/margos.

What can an adversary recover from a model trained with DP-SGD? Its $(\varepsilon,δ)$ guarantee does not answer this question directly. We derive a distortion bound that can be computed before training and applies even to an adversary who observes the parameter trajectory during training, knows every other training record and the population distribution from which the training records are drawn. The analyst may choose the representation in which error is measured, allowing the bound to target the aspects of reconstruction that are privacy-sensitive in a given application. Our information-theoretic framework separates two contributions: the information about a record revealed by training, measured through conditional mutual information, and the uncertainty already present in the data distribution, measured through source entropy. Rate-distortion theory combines them into a lower bound on the expected error of every reconstruction attack. For DP-SGD, the leakage bound depends only on the sampling rates and noise multipliers, while the source entropy is lower-bounded from a sample with finite-sample confidence. An orthogonal block construction keeps the resulting bound informative for high-dimensional images. We instantiate the framework in pixel space and in LPIPS feature space, obtaining, to our knowledge, the first lower bound on reconstruction distortion measured directly by LPIPS. On a range of image datasets the bounds establish non-trivial unavoidable expected reconstruction error within heuristically recommended privacy-budget ranges and beyond. Experiments compare the bounds with strong reconstruction attacks and identify the principal sources of the remaining slack.

Low-bit inference on accelerators with native 4-bit support requires quantizing both weights and activations to realize the full throughput gains these formats provide. While post-training weight quantization has been extensively studied, activation quantization remains considerably more challenging and typically requires calibration data. In this work, we introduce DUALQUANT, a dual-scale, calibration-free post-training quantization framework that directly minimizes the weight-reconstruction Mean Squared Error (MSE) and, as an immediate by-product, yields per-channel activation rescalings prior to activation quantization. Our key observation is that the scaling factors obtained by minimizing weight-reconstruction error provide effective activation rescalings, enabling competitive activation quantization without calibration data. We formulate DUALQUANT as a non-differentiable weight-reconstruction problem and develop an efficient solver based on variable splitting and block coordinate descent, where each subproblem admits a closed-form update. DUALQUANT provides a unified algorithm supporting a wide range of formats, including the full MX family and NVFP4. To our knowledge, DUALQUANT is the first dual-scaling method to formally support MX and NV block formats. Across Llama and Qwen models evaluated on WikiText-2, C4, and zero-shot downstream tasks, DUALQUANT matches or exceeds the accuracy of calibration-based SmoothQuant across most models and formats, and substantially outperforms calibration-free baselines, with large gains for MX and NV block formats. We validate these gains with a real packed W4A4 inference path for Llama-3.1-8B on an RTX 4090. The resulting implementation closely matches the accuracy of fake quantization while delivering up to 2.61x memory reduction and up to 2.46x throughput speedup over BF16.

Geospatial foundation models can support disaster damage assessment, but high accuracy on familiar disasters may hide shortcut learning: prediction from event or scene characteristics rather than evidence that transfers to a new disaster. We study embeddings from TerraMind, Prithvi-EO-2.0, DOFA, and DINOv3 on the multi-disaster datasets xView2 and BRIGHT. Holding out an entire disaster, instead of randomly splitting image tiles, reduces mean balanced accuracy from 0.807 to 0.638 on xView2 and from 0.818 to 0.568 on BRIGHT. Four sets of tests examine what information supports prediction: event and acquisition context, before/after image inputs, building count, and correct image pairing. Pre-event images remain predictive, and a classifier given only the annotated number of buildings reaches balanced accuracy of 0.629 and 0.653. Combining before/after features performs best on average on xView2, but does not improve over pre-only features on BRIGHT. Shuffling image pairs within the same disaster and damage class causes modest losses. Motivated by these findings, we test whether feature transformations learned only from training disasters improve generalization to unseen disasters. These transformations improve performance, with supervised contrastive learning performing best on xView2 and PCA on BRIGHT. Our results highlight the importance of evaluation on unseen domains for assessing model robustness and show that simple feature-level transformations can improve generalization. Code and experiments can be found at https://github.com/alishibli97/gfm-shortcut-learning.

Discrete graph diffusion models corrupt graph structure through independent edge updates with a Markovian noise model. Combined with the exchangeability requirement on the output distribution, the terminal prior is necessarily an Erdős--Rényi (E-R) random graph. This has a fundamental consequence for unconditional graph generation without node features: in the large graph limit, any GNN denoiser applied to an E-R graph approximately outputs an E-R graph, making effective denoising difficult. In practice, generated graphs fail to recover meaningful node heterogeneity and higher-order structure. Previous models introduced specialized node features to compensate, framing this as an expressivity problem, while the role of the prior remained unexamined. We instead address the root cause by introducing the Latent Sociability Prior (LSP) and a join} diffusion process that co-evolves latent node sociability variables and graph structure, removing edge independence while preserving exchangeability. Unlike existing latent graph diffusion methods, the sociability mechanism requires no complex encoder: its sole role is to guide edge updates and preserve node heterogeneity throughout the trajectory. We validate our approach on multiple graph generation datasets and show that the proposed framework more accurately captures graph distributions than baseline methods without node features.

Prior-data fitted networks (PFNs) are trained to approximate the Bayesian posterior predictive distribution (PPD) of a synthetic prior, yet results about the approximation quality are surprisingly thin. Taking an empirical risk minimization perspective, we decompose the deployment error of PFNs into prior, context, and amortization error, and measure the latter against a reference PPD under a realistic prior. Results show that PFNs generally recover the targeted PPD very well. Nevertheless, we find that parts of the residual amortization error persist, especially in \textit{rare} settings under the prior pretraining distribution, even when scaling the model size, which manifests in disagreement between independently pretrained PFNs and leaves uncertainty about the predictor at inference time. As a remedy, we study multiple types of Bayesian model averages and find that these can consistently reduce this uncertainty and improve predictions up to full-size TabICLv2 models.

Training long-horizon language model agents requires assigning credit to intermediate decisions from sparse terminal feedback. Within a finite group of sibling rollouts, useful action comparisons depend on their context, evidence for both alternatives, and how that evidence enters learning. We propose Matched-Context Group Policy Optimization (MCGPO), a procedure for constructing weighted action preferences from already collected trajectories. MCGPO defines comparison contexts with pre-action state summaries and observed rubric progress, annotated by task rules. Two-action support selects among ordered exact and coarse matching rules. Fixed folds separate candidate decisions from the terminal outcomes used to orient their preferences, and support-dependent evidence weights scale an auxiliary pairwise objective alongside shared GraphGPO task learning. MCGPO reuses rollout data and each response's original history, preserving critic-free training. Across six environment-backbone settings with three training seeds each, MCGPO improves mean success over matched GraphGPO: by 0.47-1.46 percentage points on ALFWorld and WebShop with Qwen2.5-1.5B and 7B, and by 3.96 and 5.16 points on WebShop with Qwen3-4B and 8B. The ALFWorld gains build on baseline success above 92%, where remaining headroom is limited. Integrated success improves in three of four Qwen2.5 settings. Three-seed factorial controls support opposite-fold evidence and evidence weighting; joint state-rubric matching achieves the highest mean in all four representation-control settings.

Basic mathematical tasks like modular addition have been a valuable tool for illuminating aspects of neural network behavior that are challenging to see in noisy, unverifiable, large-scale, real-world settings. These studies have typically focused on a single task. However, modern language models are exposed to a vast number of tasks despite being trained with a next-token prediction objective. In this work, we introduce two mathematical task ecosystems, IntWorld and PermWorld, built from distinct but related tasks on integers and permutations. Both are designed to simulate task diversity within a cohesive mathematical 'world' where tasks are distinct but related. Because the datasets are procedurally generated, we can precisely vary the number of tasks, allowing us to begin to disentangle the role of task variety from other factors in training. After describing the datasets, we provide some initial results obtained from these datasets which support the idea that increasing task diversity increases the richness of a model's internal representations.

On-policy self-distillation promises to replace the sparse outcome signal of verifiable-reward RL with dense token-level credit: a teacher conditioned on the reference solution scores the student's own rollout, and the resulting distributional shift is distilled back into the policy. We investigate whether this scoring signal actually contains corrective credit. Using gold error spans in student rollouts, we test the two signals consumed by self-distillation objectives, the full-distribution KL and the token log-ratio, for two properties: whether they localize errors and whether they push probability away from erroneous student tokens. They exhibit neither of these properties nor can they reliably distinguish correct from incorrect rollouts within the same problem. We argue that this failure is structural: locating an error in a reasoning trace is itself a search problem, while a scoring teacher is never explicitly asked to perform that search. The same model, with the same privileged information, localizes errors substantially better when allowed to generate a critique. This forms the motivation for \emph{critique-conditioned self-distillation}: first generate a critique of the current rollout, then condition the scoring teacher on that critique. The resulting signal localizes errors, is corrective within error spans, and shifts probability toward tokens that initiate backtracking such as ``Wait'', ``But'', and ``Actually''. We further show that the critique is useful both at test time, where resuming a wrong rollout with the critique in context converts math problems the model cannot solve by sampling, and in cold-start GRPO, where it provides an advantage signal on prompts for which verifiable rewards alone provide none.

World-action models (WAMs) predict a robot's actions together with the observations those actions will produce. However, the strongest models rely on video diffusion transformers that are far too computationally expensive for resource-constrained edge devices. We present SmolWAM, a latent WAM that we derive from the 2B-parameter Cosmos Policy by keeping its pretrained trunk and rebuilding its inference interface: single-frame visual latents, cached text features that preserve the trunk's attention sinks, 12 of 28 transformer blocks, and 1,008 tokens from which actions and future visual latents are read out in one rectified-flow step. Trained in 3.6 GPU-hours, SmolWAM succeeds in 95.5% of 400 LIBERO episodes at 57 ms and 2.3 TFLOP per 16-action chunk with 2.4 GiB of GPU memory, 11x faster than its teacher with 18.5x fewer FLOPs and 3.2x less memory, and its world head still forecasts scene motion.

Cooperative multi-agent tasks often impose constraints on a shared objective or collective safety. A team level violation cost function can measure this constraint outcome, but does not distinguish which agents' actions most strongly influence that violation. We present SafeDiSTL, a constrained multi-agent reinforcement learning framework that allows for credit assignment even with joint constraint measurements. We use Signal Temporal Logic (STL) robustness to measure episode-level violation through running-maximum cost increments. Smooth full-episode robustness and differentiable dynamics provide action sensitivities, which allocate each cost increment across agents at the same decision time. These sensitivities account for delayed effects when weighting agents' shares of each cost increment. We study the framework in cooperative payload transport, connected navigation, and networked traffic control tasks. SafeDiSTL achieves 1.6-5.1x the safe-success rate in transport, improves navigation safe-success by 11.1%, and delivers 50-68% more throughput in traffic when compared to constrained baselines. Results demonstrate a strong improvement of the learned policy in dealing with shared constraint while retaining the performance of unconstrained policies.

Efficient attention mechanisms, such as linear and sliding window attention, aim to reduce the computational complexity of Transformers from quadratic to linear. However, their degraded in-context recall often necessitates interleaving full attention layers, leaving the quadratic bottleneck unresolved in long-context processing. To address this, we introduce Orthros attention, which shifts the efficiency paradigm from structural allocation to phase-aware decoupling. As a unified module, it dynamically switches its computation pattern across different phases of sequence processing, enabling linear complexity prefilling and full attention decoding. Our design grants Orthros attention a lower asymptotic complexity compared to vanilla Transformers in prevalent long-context scenarios characterized by heavy prefilling and light decoding. Beyond theoretical formulation and complexity analysis, extensive empirical evaluations demonstrate Orthros's superiority across multiple dimensions. First, we validate its fundamental language modeling capabilities, demonstrating scalability across varying model sizes and strong competitiveness in comprehensive architectural comparisons. Second, Orthros exhibits exceptional conversion feasibility via a data-efficient Transformer-to-Orthros conversion, yielding performance that even surpasses the original Transformer. Finally, Orthros achieves a $21.9\times$ prefilling speedup over Transformers with end-to-end inference gains, while preserving long-context capabilities, as demonstrated by near-perfect 128K needle-in-a-haystack results and Transformer-level performance on real-world in-context recall tasks. Overall, these advantages in high performance, flexibility, and efficiency position Orthros as a promising candidate to succeed the Transformer in industrial applications.

Scaling generalist robotic policies is limited by the high cost of collecting embodied interaction data. Large-scale egocentric human videos offer a rich alternative, but they lack robot action labels, and existing action annotation methods are often noisy under real-world visual complexity and diverse manipulation dynamics. We introduce LAPT-Scale, a scalable latent action pre-training paradigm with two components: LAPT-Scale-LAM learns continuous latent actions by explaining frame-to-frame visual transitions in human videos and uses them as human-derived supervision, while LAPT-Scale-Policy trains a VLA policy through human-video latent-action pre-training, target-robot embodiment alignment, and task-specific post-training. Unlike prior vector-quantized LAMs, LAPT-Scale-LAM uses a continuous latent space to preserve fine-grained motion and interaction dynamics. We further combine perceptual, semantic, and optical-flow alignment signals to improve transition-level supervision beyond pixel reconstruction. Trained on 23.3K hours of egocentric video, LAPT-Scale-LAM provides implicit action supervision for policy pre-training, yielding consistent scaling gains and improving real-robot success from 63.3% to 85.0% as annotated human video scales from 1K to 5K hours. The learned policy also generalizes to unseen visual conditions.

Origin-destination (OD) flow forecasting supports urban mobility decisions by predicting the volume and patterns of human movement across a city. Most existing public OD benchmarks provide only a partial view of citywide mobility, covering a single transport service or at most two modes. With partial mode coverage, observed demand changes are difficult to separate into shifts across transport modes and changes in total travel. We introduce SeoulMMOD, a large-scale public benchmark for citywide multimodal mobility in Seoul. SeoulMMOD provides three years of hourly OD flows estimating total mobility for 6 urban travel modes across 25 districts and 426 sub-districts, together with travel-time, travel-distance, calendar, rainfall, and point-of-interest (POI) signals. Using SeoulMMOD, we benchmark 14 forecasting baselines under a common protocol and analyse the effects of spatial scaling, joint versus mode-separate training, and cross-year generalization. Results show that several spatio-temporal models that work at district level do not scale to sub-district OD forecasting, and that naive joint multi-mode training does not consistently improve accuracy. Together, the dataset and benchmark establish a reproducible testbed for multi-mode, multi-year OD forecasting at the city scale.

Counterfactual inference asks how an observed outcome might have differed under different circumstances, such as whether a patient would have recovered had they received a different treatment. Such questions arise whenever decisions depend on the outcomes of alternatives that were not observed. Answering counterfactual queries generally requires knowledge of the underlying causal structure, which is often unavailable in practice. A common approach is therefore to first estimate the causal graph from observational data and then perform counterfactual inference. However, errors in the estimated causal graph can propagate to downstream counterfactual predictions. To address this challenge, we introduce the Counterfactual Graph Neural Process (CFGNP), an end-to-end approach that answers counterfactual queries directly from observational data without first estimating a causal graph. CFGNP uses Graph Neural Networks (GNNs) to exploit the dual-graph structure by restricting cross-world message passing to corresponding factual--counterfactual variables. Across synthetic and image-based benchmarks, CFGNP achieves stronger predictive performance than existing approaches, including on the German Loan benchmark, where it reduces MSE from $4.51 \times 10^{-2}$ to $0.19 \times 10^{-2}$. Furthermore, on synthetic causal graphs, a model trained only on 10-node graphs maintains strong predictive performance on graphs with up to 30 nodes, without retraining.

Legal language understanding requires reasoning over semi-structured regulatory documents, yet standard pre-training objectives do not explicitly target this structural capability. Because current legal language models rely solely on masked language modeling, they treat statutory text as flat token sequences and ignore the underlying hierarchical constraints. To bridge this gap, we present a multi-task pre-training approach for German traffic law (StVO). Our approach combines three complementary objectives: Next Sentence Prediction (NSP), Knowledge Graph (KG) Guided Clustering (CLU), and Knowledge Graph (KG) Aware Paragraph Classification (CLS). This approach explicitly encodes structured legal knowledge directly into the model weights. This encoding enables zero-shot legal reasoning at inference time and removes the need for discrete graph traversal. We evaluate our approach across nine Transformer-based encoders and 324 experimental runs. On synthetic StVO question answering, our models demonstrate consistent relative improvements between 10.3% and 155.0% over baselines. Furthermore, relative to the established LexGLUE baselines (Chalkidis et al., 2022), our models improve over prior results on four of seven legal benchmarks. It also achieves consistent zero-shot transfer gains between 3.4% and 15.8% on Austrian driving regulations. These experimental results indicate that our training approach generalizes successfully beyond the initial training legal domain.

Federated clustering can support applications such as patient stratification across hospitals and customer segmentation across distributed financial datasets. Responsible AI calls for fair representation of demographic groups and support for data deletion, motivated in part by rights such as the GDPR's right to erasure. These goals interact: deleting records can change the balance between groups and alter the clustering of the remaining population. We study exact machine unlearning for federated socially fair $k$-means, whose objective is to minimize the largest group-average squared distance to the nearest center. Our approach keeps separate local summaries for each group, so summaries at unaffected institutions remain valid as global fairness weights change. For two groups, we prove an initialization approximation guarantee independent of client group proportions under ideal sampling, and we show that same-seed replay reproduces retraining exactly. Reuse pays off most when a whole client withdraws: across three datasets, replay is $35.8\times$ faster than retraining for the initialization-only model and remains $2.4\times$ and $1.7\times$ faster with one and two fair-refinement rounds, which improve worst-group cost. For individual record deletions, gains are modest (about $1.06\times$), and certifying cached assignments can cost more than it saves. These results map where exact unlearning can profitably reuse computation. Code and reproducibility materials are available at https://github.com/arareddy/mufsc.

Generative Flow Networks (GFlowNets) learn policies that sample discrete objects in proportion to a positive reward. They are trained by minimizing balance objectives, whose exact satisfaction guarantees correct sampling. However, a comprehensive theory linking nonzero balance loss and approximate stationarity to sampling accuracy, especially under evolving sampling distributions and backward policies, is still lacking. To bridge this gap, we develop a unified theory for Flow Matching, Detailed Balance, and Trajectory Balance in this regime. We first bound terminal symmetric KL by the square root of the expected balance loss. Then, we show that approximate stationarity controls this distribution error up to a tangent-space residual measuring parameterization incompleteness. Finally, we establish finite-time convergence under fixed or controlled sampling and fixed, balance-trained, or separately updated backward policies. Under standard smoothness assumptions, with fixed or controlled sampling and a fixed or balance-trained backward policy, the expected loss is $O(\barε^2+T^{-1})$ and expected terminal symmetric KL is $O(\barε+T^{-1/2})$ under uniform coverage. For a separately updated backward policy, we characterize the sufficient update decay to approach the residual neighborhood and preserve these rates. These results connect balance objectives to sampling accuracy and clarify how parameterization, coverage, and moving components govern GFlowNet training.

Long-context serving runs out of KV memory before the hardware reaches its useful throughput. HiSparse stores history in host memory and fetches only the selected entries, but it uses one cache budget at every load and does not handle draft-model speculative decoding such as DSpark and DFlash2. We present Adaptive DHiSparse, which supports speculative decoding over hierarchical sparse KV and adapts residency to load. It protects the required working set, retains additional history while GPU memory is spare, and recovers the rest through exact fused retrieval. DSpark and DFlash2 keep their proposal and acceptance rules: queries share missing-row retrieval, draft KV uses a window-bounded ring, and verification staging is reused across layer groups. At 100K context and concurrency 128, GLM-5.3 with DFlash2 delivers 793 output tokens/s per decode GPU, 1.45$\times$ HiSparse and 1.27$\times$ non-speculative adaptive serving. Held to the same latency bounds as HiSparse, it sustains up to 4$\times$ the concurrency and 2.3 to 3.7$\times$ the output, depending on the per-token bound. On DeepSeek-V4.1-Flash at 500K context and concurrency 512, Adaptive DHiSparse with DSpark delivers 1.85$\times$ the output of resident serving with DSpark. In a load step, optional residency reduces host fetches from 14.2% to 1.4% of selected rows at low load and releases memory under pressure.

The fast pace in which protocols, frameworks, and technologies are being proposed for LLM-based agents results in a complex scenario for anyone aiming to create state-of-the-art AI solutions. Without an assessment on how different protocols, frameworks, and libraries work together to compose an Agentic Web solution, developers and organizations face different challenges (e.g., skilling, governance, infrastructure) even before becoming part of the Agentic Web ecosystem. In this context, we propose a system for performing Agentic Web evaluations called MASTIF (Multi-Agent System TestIng Framework) and present a baseline assessment of Agentic Web technologies, including 3 LLMs, 3 protocols, and 4 agent implementation frameworks, against 300 Mind2Web benchmark tasks. Results suggest that selecting a task-specific technology stack may result in gains of task completion rate, efficiency score, and social value stability scores; and reductions in verbosity and time to complete tasks. With the presented results, we expect to present a baseline in terms of metrics and results for this complex and fast evolving scenario for those aiming at designing, assessing, and developing multi-agent systems.

Any automation within reliability-critical contexts such as medicine or law has to be transparent and verifiable. Despite their advantages, current LLM-based systems do not meet these requirements due to their inherent black-box nature. Retrieval Augmented Generation (RAG) attempts to ground the generative process in the relevant interpretable documents, thereby reducing the risk of AI 'hallucinations' and potentially enabling verification of the initial generation context. However, as the number and size of potential sources increases, naive retrieval approaches tend to perform worse, ignoring critical context elements. In this paper we aim to tackle this problem in the context of German Legal Code. Using structure-driven parsing to create a Knowledge Graph (KG) for the source document navigation and retrieval, we compare several KG-informed chunking and retrieval strategies and evaluate their effect on the accuracy of retrieved citations.

Neural networks (NNs), while highly flexible and versatile, are often poorly calibrated. Post-hoc calibration methods adjust the outputs of an already trained model using a separate calibration dataset and additional calibration parameters. This process adds overhead that requires separate optimization of the calibrator and subsequent model selection. In this paper, we show that standard (post-hoc) calibrators can be trained end-to-end with the backbone via a bilevel Stackelberg formulation. This formulation eliminates the need for a separate calibration set or a dedicated calibration loss by jointly optimizing accuracy and calibration using cross-entropy, introducing only one additional hyperparameter: the calibrator learning rate. Although our framework supports calibrators ranging from temperature and matrix scaling to NN layers, our results show that using the final classification layer itself as the calibrator yields accurate, well-calibrated models with no additional parameters. Our approach outperforms SOTA calibrators with minimal added complexity, while also avoiding the multiplicative overhead of separate calibration components in settings such as multi-task learning.

Standard Policy Gradient (PG) algorithms suffer from high estimation variance because they rely on Monte Carlo updates. While off-policy variance reduction mitigates this problem, establishing optimal sample complexity currently requires either restrictive assumptions, such as uniformly bounded importance weights, or impractical second-order operations. In this work, we introduce Loopless Variance-Reduced PG (LVR-PG), a unified theoretical framework that abstracts a class of first-order estimators that integrate recursive updates with defensive importance sampling, including known estimators such as PAGE-PG and STORM-PG. We prove that this algorithmic class achieves $\tilde{\mathcal{O}}(ε^{-2})$ last-iterate global convergence under weak gradient domination for general parametric policies. We establish this guarantee using purely first-order updates, with minimal assumptions and without gradient preconditioners. By extending state-of-the-art lower bounds to continuous parametric policies, we establish an $Ω(ε^{-2})$ sample complexity limit, confirming our framework is minimax-optimal up to logarithmic factors and horizon terms.

Diffusion models typically learn the generative dynamics of a single target distribution or a family of distributions indexed by conditioning variables fixed during training. This limits their ability to adapt to a previously unseen distribution from only a small set of examples. We introduce an in-context solution operator learning framework for diffusion models, in which each target distribution is treated as a task drawn from a meta-distribution and is specified at inference time by an empirical context set. Our method learns an in-context score operator that maps the context samples, a noisy query point, and the diffusion time to the task-dependent score, enabling reverse-diffusion sampling on unseen tasks without retraining or parameter updates. A key observation is that the task-dependent score admits a Gaussian-kernel conditional-mean representation. Its empirical Nadaraya--Watson approximation can be represented exactly by a model in our constructed transformer class, providing a principled foundation for the proposed query-dependent attention architecture. The operator is trained through amortized denoising score matching across task distributions. We establish generalization-error bounds that quantify the effects of the context size, the numbers of training tasks and within-task samples, model complexity, and intrinsic and ambient dimensions. Numerically, we evaluate the method on a range of scientific and visual datasets, including Navier-Stokes, ERA5 climate data, QuickDraw sketches, and Food-101. The results show that the proposed method performs well on relatively simple task families, while its performance degrades as task complexity increases, thereby illustrating both the promise and the limitations of in-context learning for diffusion models.

Learning how local perturbations affect a network is important for identifying which components drive or absorb a system's response. We formulate this as seed-conditioned response ranking: given a perturbation at one seed node, ranking all other nodes by response strength. This is challenging when the model is given only the interaction topology, while the mechanism governing how perturbations propagate must be learned from data. We present VQIL (Visible Quantum Interference Learning), a trainable Szegedy quantum walk for seed-conditioned ranking. Edge weights, link phases, and per-node phase rotations are optimized end-to-end by backpropagation while the walk operator remains unitary throughout training. Node scores are obtained from a geometrically discounted coherent sum of quantum-walk amplitudes, enabling quantum interference before the final squared-norm readout. Because each learned parameter is attached to a named edge or node, the trained model provides directly inspectable hypotheses about the roles of individual graph components in the propagation mechanism. On AC power-flow perturbations over six transmission grids, with electrical parameters withheld from the model, VQIL improves Spearman rank correlation from 0.147 for the untrained walk to 0.831 and outperforms an identically trained random-walk model. On gene perturbation over five protein-protein interaction networks with Perturb-seq supervision, VQIL improves rank correlation from 0.171 to 0.620, outperforming the strongest learned classical and neural baselines. These results show that trainable quantum-walk interference provides a scalable and interpretable inductive bias for learning seed-conditioned response rankings on graphs.

Audio-visual dataset distillation aims to synthesize a small dataset from a large paired audio-visual dataset, while nearly preserving the training utility for neural networks. Existing approaches such as DM and AVDD face two critical limitations: 1) they are memory-intensive, making them difficult to scale to complex architectures and higher samples-per-class (SPC) settings, and 2) they often fail to preserve strong cross-modal semantic correspondence. To tackle these issues, we propose X-AVDD, a cross-attention-based distillation method that explicitly captures inter-modal relationships during synthesis. By decoupling cross-modal alignment from joint optimization, X-AVDD is substantially more memory-efficient, reducing end-to-end synthesis cost by $\sim 8\times$ GPU-hours, while producing highly aligned synthetic data. Empirically, X-AVDD significantly outperforms other audio-visual distillation methods. On VGGSound, an audio-visual dataset of short YouTube clips, X-AVDD improves top-1 test accuracy from $8.2\%\rightarrow24.0\%$ at SPC=$10$ and $9.8\%\rightarrow26.5\%$ at SPC=$20$. Furthermore, X-AVDD improves cross-architecture generalization (e.g., $25.5\%\rightarrow35.3\%$ for a ViT target on AVE dataset at SPC $=50$) and successfully scales to complex architectures such as AudioCLIP, achieving $62.3\%$ accuracy using only $5.65\%$ of the full dataset, against the full dataset accuracy of $72.4\%$ while completing in $9\times$ lower training time compared to the full dataset.

Prior work suggests fraudulent scientists write their articles with higher rates of linguistic obfuscation than genuine scientists, with a possible mechanism suggesting obfuscation makes research more costly for a reader to analyze. This result has not undergone additional empirical scrutiny in the open science era, nor have other psychological mechanisms been evaluated. The current replication and extension addressed this opportunity by comparing 800 fraudulent and 800 genuine control papers (2013-2025). The results supported the linguistic obfuscation hypothesis, suggesting fraudulent papers were more obfuscated than genuine ones. A parallel mediation model evaluated three competing mechanisms, finding that reality monitoring, as measured through decreased rates of adjectives, best explained the link between fraudulent science and linguistic obfuscation. Cross-validated detection accuracy using obfuscation features (60.7%) and an elastic net of over 100 language dimensions (71.6%) were significantly greater than chance. Implications for scientific fraud and psychology of language research are discussed.

Time series foundation models (TSFMs) increasingly offer covariate support, with calendar information being the most common type of categorical covariate in practice, as it is available for every point in time. However, how these features are encoded is largely treated as an implementation detail rather than as an object of study, despite encoding being one of the few design choices left for pretrained models at inference time. With the present study, we demonstrate that the way information is propagated into the model substantially affects forecast accuracy of TSFMs. Therefore, we systematically compare seven calendar encodings, including three conventional target-agnostic representations (ordinal, one-hot, and cyclic) and four target encodings, across four state-of-the-art TSFMs on nine datasets from multiple domains in a zero-shot rolling evaluation. The nine datasets are influenced by social activities, such as weekends and holidays, making calendar information particularly relevant for forecasting. We find that encoding choice affects forecast accuracy across all models, with MASE differences of up to 9.9 % on non-holidays and 83 % on holidays within a single model. Effectiveness and utilization vary substantially across models. Improvements are concentrated around holidays, where calendar information provides signals unavailable from target history, while some model encoding combinations also improve non-holiday forecasts. A SHAP-based attribution analysis locates the contribution of the covariates relative to the autoregressive component and explains why some model-encoding combinations degrade performance. Our findings suggest that current benchmarks for covariate-informed TSFMs can distort model comparisons by imposing fixed encoding strategies.

Linearizing softmax attention offers a route to efficient Transformer inference, but preserving pretrained attention remains challenging. We identify an important source of approximation error overlooked by standard linearization methods: a small number of outlier dimensions contribute disproportionately to query-key dot products and account for a large fraction of the approximation error. Motivated by this heterogeneity, we propose Outlier-Aware Attention Linearization (OAL), a dimension-aware method for constructing quadratic attention approximations. Starting from a second-order Taylor approximation, we introduce a Hadamard-product formulation that makes individual query-key contributions explicit. This formulation relaxes globally shared coefficients to support dimension-specific linear terms and cross-dimensional quadratic interactions while preserving separable evaluation. To retain this flexibility with fewer parameters, OAL groups dimensions with similar contribution scales to share coefficients, while allowing distinct treatment of the few outliers. The resulting positive quadratic kernel is calibrated against pretrained attention outputs and admits linear-time evaluation with respect to sequence length for fixed head dimensions. On Qwen2.5-0.5B, OAL reduces frozen-attention output relative MSE by approximately 8.9% compared with DiJiang. Replacing middle-layer attention modules and jointly fine-tuning the kernel coefficients with LoRA further lowers WikiText-103 perplexity from 31.71 to 27.50 compared with a learned global quadratic, while improving accuracy on PIQA, HellaSwag, and ARC-E. These results demonstrate the value of outlier-aware dimension grouping for both attention approximation and pretrained-model conversion. The code is available at https://github.com/wuqianyue09/OAL

Large language models struggle to reliably identify and correct errors in their own reasoning. Unlike prior training-based approaches that fine-tune an LLM to generate revisions or feedback, we learn where to intervene in failed rollouts and determine how to redirect them. To choose where to intervene, we train a 130K-parameter MLP head on a frozen encoder using only final-answer correctness. For redirection, we find that under a fixed generation budget, the tested inline injections fail to reliably redirect reasoning, whereas presenting the same instruction in a new user turn changes the model's course. The resulting correction framework, UPTURN (User-Prompt TURN injection at a learned position), requires no LLM fine-tuning: it retains a selected prefix, injects an instruction as a new user turn, and regenerates the suffix. Across four reasoners and six math benchmarks, UPTURN improves average accuracy in every setting over a 32-rollout baseline. It outperforms self-correction and reflection-token injection baselines under matched correction budgets, transfers across reasoners and to GPQA-Diamond without retraining, and complements GRPO and self-correction methods.

Curriculum learning typically organizes training samples from easy to hard. In GRPO, medium-difficulty samples are attractive for online selection because they are more likely to yield informative group-relative learning signals, and prioritizing them can improve training efficiency. However, it remains unclear whether persistently selecting such samples also leads to the strongest final model performance, or whether training can benefit from progressing toward harder samples. Meanwhile, sample difficulty changes as the policy evolves, so fixed difficulty estimates can become misaligned with the current policy. We therefore re-estimate sample difficulty under the current policy and, inspired by the easy-to-hard principle of curriculum learning, design and evaluate a staged curriculum that progressively shifts the active difficulty band toward harder samples. Experiments on GSM8K show that persistent medium-difficulty selection does not achieve the strongest observed performance. Across three random seeds, progressing to a moderately harder difficulty band improves test accuracy to $0.6591 \pm 0.0077$, compared with $0.6449 \pm 0.0024$ for the fixed-medium curriculum. However, further progression to a harder difficulty band reduces accuracy to $0.6405 \pm 0.0053$. These results show that staged difficulty progression can outperform persistent medium-difficulty selection, but its benefit is non-monotonic. Effective GRPO curricula can therefore benefit from progressing beyond medium difficulty, while avoiding excessive progression toward overly difficult samples.}

With the advances of coding agents, software developers are spending more time reviewing code than writing code. As human review velocity falls behind agentic code generation, it is tempting to build automated review agents towards a future of "lights-off" software factory with little to no human supervision. To assess the practicality of such a fully autonomous software factory, we construct CRBugBench, a new benchmark that evaluates whether review agents detect human-verified bugs. We build a data pipeline that mines bugs from repository history, and pair each bug with its introducing PR to form a review task. CRBugBench consists of 1,000 review tasks across 98 open-source GitHub repositories, and 800 additional tasks from proprietary industry repositories at one large company. We then use the benchmark to study agent review quality: We found that SOTA review agents do well: the best (claude-opus-5 with copilot-cli) catches 71.9% of bugs within its top-5 comments, up from 47.7% (opus-4.6) in five months of model evolution. Yet agents remain less effective on complex changes and bugs requiring non-local reasoning, with sharp performance drops on proprietary industry repositories where project context is more often needed. To improve bug detection, we found scaling compute helps most through prompt changes: exploration-nudging prompts gain up to 5.8% performance at 10% higher cost. From these findings, we derive concrete guidance for developers to adopt review agents - scaling review effort to PR complexity, steering agents by project priorities, supplying repository knowledge, dispatching models by PR characteristics, and monitoring quality as models evolve.

Many machine learning approaches are two-level: an expert selects a class of methods, within this class the choice of key components is critical and usually manual. This paper addresses the automatic tuning of these components and proposes a framework transferable across models of different nature. The multi-agent large language model (LLM) framework has three roles---analyst, coder, and recorder~---with an experiment log and a sandboxed execution environment. Unlike systems that optimize the entire ML script or search architectures in a discrete space, the framework operates over an open but semantically narrow space: the agent generates code only for the specified components, leaving the overall network structure unchanged. The approach is validated in two directions: tuning a classification spiking neural network with correlation learning, and solving high-order differential equations with Physics-Informed Neural Networks (PINN). Across five tasks---spiking classification on MNIST and Digits, and boundary-value problems for the generalized nonlinear Schrödinger equation (GNLSE) and Kuramoto--Sivashinsky (KS) equations---the agent outperforms the baselines: classification accuracy is 13 percentage points (MNIST) and 6 percentage points (Digits) higher; the error against the analytical single-soliton solution is reduced approximately 2491-fold; the invariant-preservation error drops approximately 325-fold in the selected GNLSE run with a Gaussian perturbation and by more than 100-fold on average over repeated initializations; and the mass drift drops 14-fold for the KS problem with a Gaussian perturbation. The framework thus tunes key components on tasks of different nature without changing its internal structure.

The Abstraction and Reasoning Corpus (ARC) is increasingly used to evaluate whether neural networks can infer task-specific transformations from few demonstrations. Recent ARC models differ in architecture, but also in broader training and evaluation protocols, including pretraining data, data augmentation, test-time setup, and optimization budgets, making conclusions about architecture benefits debatable. In this work, we study how the performance of the vision-inspired VARC and recursion-based TRM models is affected by pretraining and test-time training (TTT) budgets under a common inductive protocol on ARC-AGI-1. We cross five pretraining budgets with four TTT budgets for each test task. Each budget specifies a task-variant count and total sample presentations, defining the controlled training budgets. We observe that increasing TTT sample presentations improves the standard ARC pass@2 metric in all settings for both models. Increasing the TTT task-variant count consistently helps only TRM, whereas VARC often performs better with fewer variants. VARC also remains competitive without pretraining, while TRM benefits much more from it. These results show that pretraining and TTT budgets affect ARC evaluation and motivate reporting at comparable budgets.

Latent field models are essential for spatial statistics, but the posterior distribution of the latent field, even with a Gaussian process (GP) prior, is analytically intractable whenever the likelihood is non-Gaussian. Existing approximations are either too slow or too restrictive for large non-Gaussian datasets. We present a scalable variational inference framework that mirrors the sparse nearest-neighbor autoregressive structure of the prior in the variational family, uses normalizing-flow conditionals for expressiveness, and amortizes all variational parameters via a single shared neural conditioner. Doubly stochastic optimization with reduced ancestor sets achieves polylogarithmic complexity per iteration.

Tabular foundation models (TFMs) are increasingly popular for their strong predictive performance. Many also provide posterior predictive distributions (PPDs), whose reliability remains less studied. We assess reliability through point prediction, distributional quality, marginal and subgroup calibration, and interval efficiency. We then introduce TabReliBench, a large classification and regression benchmark for these comparisons and dataset-level analyses. We find that TFMs' rankings change between TabArena and our benchmark and across reliability metrics. Regression TFMs usually achieve near-nominal marginal coverage and retain narrow intervals, but this favorable marginal behavior does not imply valid conditional coverage. Dataset-level analyses show that classification difficulty often accompanies poor calibration, whereas regression predictive performance is not a reliable indicator of calibration. Ensembling improves predictive performance but does not consistently improve calibration of PPDs; TFMs also benefit little from conformal post-processing, which does not reliably resolve subgroup gaps. Reliable TFM selection therefore requires multidimensional evaluation beyond aggregate rankings. Code is available at https://github.com/Tianqi-Zhao/TabReliBench.

Reinforcement learning with verifiable rewards (RLVR) has substantially improved LLM reasoning, yet growing evidence suggests it mainly performs distribution sharpening: concentrating probability mass on reasoning paths the base model already produces. Recent work therefore seeks RLVR's gains without needing reward signals, by sampling from the power distribution $π_α\propto π_{\text{base}}^α$ via Markov chain Monte Carlo or sequential Monte Carlo. We show that these samplers fail to converge to $π_α$ under realistic compute budgets, and that their gains are largely recreated via simple Best-of-$N$ with appropriate grading . This suggests they succeed by selecting high-probability sequences, not by approximating $π_α$, raising a natural question: is the power distribution itself a worthwhile target? We test this hypothesis directly by fine-tuning the base model to match $π_α$, sidestepping sampler convergence issues entirely. Despite approximating the target far more faithfully than existing samplers, this strategy yields gains that are noisy and specific to certain models and sampling tasks. Together, our results suggest that the value of power sampling lies not in faithfully targeting $π_α$, but in the simpler goal of extracting sequences the base model already deems likely.

Machine learning methods are increasingly used to predict chemical properties directly from the molecular structure to bypass the often prohibitive cost of quantum mechanical calculations. Recent work has shown that including the ground-state electron density as input to the machine learning model can improve accuracy in this setting. This work employs the recent surrogate functional framework originally developed for orbital-free density functional theory and extends it to predicting molecular properties. We show that passing the ground-state density as input to the model, while simultaneously enforcing it to be the minimum of the surrogate functional, surpasses state-of-the-art performance for six out of seven extensive properties of the QM9 dataset. Additionally, the approach shows improved generalization to larger molecules taken from the QM40 dataset. On all extensive properties of both datasets, our approach outperforms models that are informed by the ground-state density but do not enforce the surrogate functional constraints.

AI assistance in software engineering focuses on implementation and verification, leaving the design phase, and in particular the maintenance of architecture documentation, comparatively underserved. Architecture as Code makes architectural artifacts versionable, diffable, and machine-readable, but it does not reduce the effort of keeping them updated, so design documentation still ends up as a fragmented, author-dependent, and only partially trusted record of why a system looks the way it does. In this paper we introduce the Agentic Software Architecture Workflow (A-SAW), a process-level artifact that extends Architecture as Code with a persistent agent layer: an IDE-integrated agent drafts architecture documents from a version-controlled C4 model, a typed element catalog, and instrumented templates. Every draft passes a mandatory human review gate before it is accepted. We evaluate it in a large industrial organization along two tracks: a rubric assessment of 45 architecture documents including 13 matched source-regeneration pairs, and 13 task-anchored practitioner interviews. Our results show that the workflow collapses the between-author variance that made quality unpredictable. Intrinsic quality also rises, by a median of +19 points across the matched pair. The cohort has a mean of 58% and 89% bundle templates, catalog, agent, and a richer regeneration context. So they bound the agent's contribution from above rather than isolating it. In the agent-generated documents, completeness (100%), consistency (93%), and machine-readability (99%) reach or approach the rubric maximum, while traceability reaches 79%. Eleven of the 13 participants rejected agent output that no human had reviewed, and one warned that polished output may lead reviewers to skip the review. Our work supports keeping human review mandatory and concentrating it where the agent must infer beyond the repository.

Selecting a subset of features (columns) from a tabular dataset is a ubiquitous challenge in real-world applications. Industry experts seek to minimize model training costs, climate scientists require an interpretable subset of factors to understand extreme weather events, and clinicians need to choose among diagnostic tests due to financial and time constraints. When looking at the scientific literature to determine which feature selection method to use, practitioners across domains face an unrepresentative landscape of benchmarking. Despite extensive research, existing feature selection benchmarks provide limited guidance because they rely on outdated models, ad hoc evaluation protocols, or artificially generated data that do not reflect real-world problems. To close the gap between scientific benchmarking and real-world applications, we introduce SelectArena, a modern benchmark for feature selection. We compare 15 widely used feature selection methods on 71 manually curated real-world tabular datasets, covering classification and regression problems, using standardized evaluation protocols and state-of-the-art tabular (foundation) models. Our results reveal clear trade-offs across predictive performance, runtime, stability, and validity. Tree-based methods, CART and RFImportance, along with F-Test are Pareto-optimal when taking into account predictive performance and runtime. F-Test, information-theory-, and distance-based methods lead on stability and validity, with F-Test emerging as a balanced choice across our metrics. Search-based methods are slow, unstable and achieve low performance. SelectArena turns these findings into practical guidance and, as an extensible platform, supports future feature selection research.

In this paper we describe a Constraint Acquisition Tool that generates size independent constraint programs from input data and positive and negative example solutions of multiple problem instances. The system finds size specific models based on typical constraint patterns over global and binary constraints, then generalizes the models to a size independent model that can be run on unseen input data. The system produces MiniZinc code as well as a textual description of the model. We evaluate the tool on the example problems of the PTHG21 Challenge and show the feasibility of the approach.

Machine learning models are increasingly used for automated feature extraction from weakly annotated bioimaging data, such as microscopy and cytomorphology images. Due to the sole dependency of automated feature extraction, existing methods yield inconsistent classification performance across weakly annotated datasets. Here, we propose a model-agnostic knowledge augmentation method that enables ML models to exploit domain knowledge beyond what automated feature extraction can capture. The proposed approach improves classification performance across seven ML architectures. When integrated with masked autoencoder-based self-supervised learning, our method achieves the strongest overall performance, improving precision by 2.9% over the next-best method on three in-house microscopy datasets and by an average of 6.8% across three public cytomorphology benchmarks. We perform nucleus-level classification based on mislocalization of TAR DNA-binding protein 43 (TDP-43) from the nucleus to the cytoplasm, a phenotype linked to Amyotrophic Lateral Sclerosis (ALS), using three independent microscopy datasets of motor neurons derived from human induced pluripotent stem cells. Accurate quantification of this phenotype at scale remains challenging and may contribute to inconsistencies across studies. Application of the best-performing model to these microscopy datasets reveals biological interpretations related to ALS, including dose-dependent effects of the proteasomal stressor, MG132, and the oxidative stressor, sodium arsenite (SA), on TDP-43 mislocalization, variable pathology in motor neurons from sporadic and familial ALS donors, and a lack of protection of the pharmacological compounds, edaravone and niclosamide, against SA-induced TDP-43 mislocalization. Together, our pipeline enables in-vitro cellular image analysis to advance understanding of ALS-associated pathology and potential treatments.

Hypergraph Neural Networks (HNNs), designed to model higher-order relations beyond pairwise interactions, typically rely either on fixed structural weights or on normalised attention. Fixed weights can propagate misleading information in heterophilic settings, while normalisation makes attention coefficients competitive, in the sense that, in a message-passing framework, increasing the weight of one sender must reduce the weight of others in the same aggregation set. Our main contribution is the introduction of PANDA (Prior-guided AttentioNal Dual-path Architecture), a plug-in for two-stage message passing HNNs applicable to both undirected and directed settings. The main path interpolates between learned receiver-normalised attention and structural incidence priors, while a non-competitive (not normalised) auxiliary path lets each sender contribute its own transformed information. We also show that, with fixed structural coefficients, PANDA recovers a certain complex-valued convolution operator adopted in previous works. Across four HNNs spanning spatial and spectral methods in both directed and undirected settings, plugging PANDA in yields an average improvement of 5.18 percentage points across eight real-world benchmark datasets. The code is available at https://github.com/jainris/panda.

Large language model (LLM) agents are rapidly improving on long-horizon scientific and olympiad tasks. Recent systems have produced a Lean-formalized resolution of the Navier--Stokes problem and autonomous solutions to four open questions in the Erdős Conjectures database. Yet neither system releases both its underlying model and complete inference stack. More broadly, a single agent session remains brittle on difficult problems: early errors can propagate, missing domain knowledge can stall progress, and apparent success can mask a misinterpretation or weakening of the original task. We introduce HOA (Humanize Olympiads Agent), an open, flow-level framework that coordinates multiple agent sessions through three complementary feedback channels. First, an independent reviewer evaluates each candidate and provides actionable feedback. Second, a structured wiki supplies task-specific knowledge that may be missing or unreliable in model pretraining. Third, a verifier checks generated Lean proofs against trusted challenge statements. With GPT-5.6 Sol, HOA achieves perfect scores on the IMO, IPhO, and IOI 2026, near-full marks on the IChO 2026 problem set, and solves all 672 problems (100%) on PutnamBench. On the Lean-Eval leaderboard, HOA ranks first, solving 251 of 304 problems (82.6%). We further evaluate HOA with open models. Using Kimi-K3, HOA achieves full scores on the IMO, IPhO, and IOI, and a gold-level score of 97.1% on the IChO. These results closely match the closed-source GPT-5.6 Sol + /Goal baseline on the IMO and IPhO, while exceeding it on the IChO and IOI, demonstrating that systematic feedback can help bridge substantial differences in base-model capability.

Fault diagnosis in asset-intensive industries requires multiclass models trained on noisy, sensitive data collected across geographically distributed client sites. Federated learning enables collaborative training without centralizing data, but most approaches do not provide formal privacy guarantees for individual records stored at client sites and can degrade under label noise. To address these limitations, we develop a differentially private, distributionally robust framework for federated multiclass classification that provides sample-level privacy for records stored at each client. We build on a federated Wasserstein distributionally robust multiclass objective with the Crammer-Singer loss, solved using an Alternating Direction Method of Multipliers (ADMM)-based algorithm. To protect the released client updates, our method injects calibrated Gaussian noise into the proximal linearized client updates of both the local model and the local distributionally robust variable before aggregation at the central server, and we quantify the cumulative privacy loss using the moments accountant. We provide per-round sample-level $(ε,δ)$-differential privacy guarantees and an overall composed privacy guarantee across all training rounds, together with convergence results and an explicit privacy-utility tradeoff. Experiments on simulated industrial fault data and UCI benchmark datasets show that our method outperforms differentially private federated empirical risk minimization and stochastic gradient descent baselines in most tested settings, with the largest gains under strong privacy-induced noise.

Attention Residuals (AttnRes) enables content-dependent retrieval over network depth, which can be viewed as single-stream softmax attention along the depth axis. Inspired by the success of multi-head attention along the sequence axis, we introduce Multi-Stream Attention Residuals (MSAR), an extension of AttnRes that generalizes this mechanism to multi-head depth attention, yielding substantial performance gains with minimal additional overhead. MSAR draws stream diversity from sequence context through multi-scale causal convolutions, allowing different heads to retrieve historical representations through distinct contextual views while preserving the integrity of the hidden representation. The retrieved streams are combined through lightweight adaptive mixing to generate the single-stream layer input. By retaining single-stream sources and expanding them on demand, MSAR supports multi-stream computation without requiring persistent multi-stream storage or communication. Across all 12 downstream benchmarks at the 18B and 28B MoE scales, MSAR consistently outperforms prior residual variants, including standard residual connections, manifold-constrained Hyper-Connections (mHC), and AttnRes. Scaling-law experiments yield an iso-loss compute speedup of $1.58\times$ over the standard residual baseline, compared with $1.35\times$ for AttnRes and $1.30\times$ for mHC. These results make MSAR a promising residual architecture for large-scale pre-training.

Vision-language-action (VLA) policies issue sparse targets whose arrival at the robot controller is delayed and irregular, so a reasonable predicted action can still produce abrupt motion near people. Remedies are evaluated on different streams, delays and metrics. JitterVLA replays identical commands and arrival traces through alternative controllers, isolating the controller's effect. We apply it to synthetic reaches, hand trajectories from 20 people, three open VLA pipelines (838 episodes) and a physical UR3. Interpolation lowered modeled p99 jerk for every source, by 83% (615 to 106 m/s^3) on human motion with 200 ms delay, at a 2.0 cm tracking cost, and the effect replicated on 598 new episodes; jerk-limited generation lowered jerk further at a larger cost. Entropy-discontinuity associations had opposite, replicated signs in two pipelines, and entropy gating helped in one but not the other. On the UR3, interpolation lowered accelerometer-derived jerk by 3.0-14.0% in all six exploratory pairs, while a position-derived metric did not resolve the contrast. JitterVLA supports screening before human-centred studies; code and data will be released.

Evaluating large language models (LLMs) on AMLworld's smallest test splits is costly because each of over a million targets requires a separate graph prompt. Sampling reduces this cost, but small balanced subsets change unweighted precision and F1. We introduce AMLworld-Compact, an evaluation set that retains every illicit test transaction and samples benign transactions from supervised-score strata. Its HT-Coreset protocol reduces evaluation targets by 271x and 611x on HI-Small and LI-Small. Horvitz-Thompson weights give unbiased full-test confusion-count estimates for fixed predictions; recall is exact, while precision and F1 are ratio estimates. In a held-out study using Graph Feature Preprocessor (GFP) features, one boosted-tree family defines strata for evaluating the other over 500 draws per configuration. At the released sizes, F1 RMSE is 0.00-2.58 percentage points versus 4.44-7.41 for uniform sampling with HT weights; conservative intervals remain broad. Under the evaluated graph prompts, seven open-weight LLMs have weighted precision near the predict-all-illicit baseline. In an outcome-stratified audit of 1,000 HI-Small few-shot outputs from one run per model, lexical checks pass for 84.5% of correct and 56.2% of incorrect binary verdicts, so these markers do not establish correctness. We release the evaluation sets, construction code, and audit tools at https://huggingface.co/datasets/typhoon-ai/AMLworldCompactEval and https://github.com/nat-nischw/AMLworld-Compact.

Large Language Models (LLMs) are increasingly used throughout the software development lifecycle. This growing reliance on LLMs raises concerns about their trustworthiness, with deception posing a particularly important risk. Deceptive behavior takes the form of misleading responses or assistance in fabricating, concealing, or distorting information. To study this behavior, we introduce SWE-DeceptionBench, a benchmark for evaluating LLM deception in software engineering. The benchmark covers six lifecycle stages and two deception triggers: Active Deception, where models recognize infeasible or conflicting requirements but still present them as achievable, and Passive Deception, where models help fulfill explicit user requests to mislead others. We further introduce Neutral, Reward, and Pressure conditions to study the effect of external inducements, resulting in 720 test instances in Chinese and English. Evaluating seven mainstream LLMs, we find substantial deception even under the Neutral condition, with aggregate rates of 49.88% for Active Deception and 64.29% for Passive Deception. Deception rates vary across lifecycle stages, and external inducements affect Active and Passive Deception to different extents. We also evaluate lightweight defenses, finding that guard models miss many deceptive requests, and Self-Reminder reduces but does not eliminate deceptive responses. These findings highlight the need for more effective deception detection and mitigation methods to support trustworthy LLM-based software engineering.

The external validity of a study is the extent to which its findings generalize to unseen experimental conditions, such as datasets. We formalize the results of a study together with its research question, and define external validity as the probability that two independent and same-sized replications of the study produce results whose maximum mean discrepancy (MMD) is below a given tolerance. This approach, centered on the choice of a space of results and of an appropriate reproducing kernel for the MMD, is flexible enough to be deployed on multiple different types of results and research questions. Additionally, we establish an asymptotic power-law relation between the size of a study and the quantiles of the MMD, which we use to estimate the desired size of an externally valid study from a small-scale preliminary study. We deploy our framework to rankings with ties and discuss kernels suited to different research questions. Experiments on existing case studies and on a simulated newly designed study illustrate the applicability of the framework. We release our code as part of the genexpy Python module.

Professional agents increasingly work through command-line interfaces and multiple tools over shared workspaces containing files, messages, and business applications. These environments create safety risks that are difficult to capture with attacks placed at a single, known read location. We introduce APEX-Harm, a benchmark for evaluating prompt and script injection against long-horizon agents performing finance, consulting, and legal tasks in large workspaces. Each task contains 115 to 328 files, together with email, calendar, chat, document, spreadsheet, PDF, and code-execution tools exposed through MCP servers. The benchmark includes six attack categories covering static and dynamic injections, task-specific script replacement, and attacker-provided generic scripts, and evaluates attack success, exposure, and task performance. We evaluate 12 frontier models on 180 task instances per model. Opus 5 and Sonnet 5 have the lowest overall attack success rates (8% and 20%), while task-specific script replacement and attacker-provided generic script attacks remain substantially more effective, averaging 81-84% and 73% attack success, respectively. Gemini 3.8 Flash achieves the highest mean task score (0.877) while still exhibiting a 33% attack success rate. These results show that completing the requested task is not sufficient evidence of safe behavior: an agent can produce a correct deliverable while also performing an unauthorized action in the workspace.

User simulators are gaining attention as environments to train LLM assistants that can better cope with the nuances of multi-turn conversations with real users. Prior work on user simulation has emphasized generating realistic user utterances, leaving unexplored their potential as a source of direct feedback on the assistant's responses. Such evaluation signals may not be explicitly communicated in language, but can instead reside in the latent reasoning of users. In this work, we introduce user simulators with intrinsic feedback, namely, models equipped with latent evaluation states: internal thoughts, reactions, and scalar ratings that evaluate the assistant from a first-person user perspective at each conversation turn. We train LLM-based simulators using interaction logs between real users and assistants and demonstrate that intrinsic feedback improves simulation realism and critiquing behaviors versus prompted simulator baselines, while also providing turn-level evaluations that existing fine-tuned simulators lack. Further, evaluations generated by our trained simulators correlate better with human judgments than those of prompted simulators of the same size. Lastly, our experiments reveal that simulators with intrinsic feedback provide effective reward signals for multi-turn RL training, matching setups that rely on external judge LLMs for rewards while cutting RL compute by more than half.

We discuss three simple connections between generative diffusion and optimal transport. First, the joint law of the data and of their Ornstein-Uhlenbeck noising is, at every time, the solution of an entropic optimal-transport problem, at a temperature set by the noise level. Backward diffusion can consequently be read as an annealing which gradually decreases the entropic regularization. Second, we study the speciation transition--how the modes of the data distribution emerge during generation--and compare how it happens in diffusion and flow matching with respect to optimal transport. Using Gaussian mixtures as high-dimensional example, we show that the dynamical instability through which diffusion and flow matching separate the modes becomes, for optimal transport, an exponentially thin boundary layer at the beginning of generation, which is rounded by entropic regularization. Third, considering a generic forward dynamics with a general confining potential, we show that the backward process driven by the score also satisfies a drift-based variational principle whose minimal cost is the entropy difference between initial and final distributions.

Vocal intensity conversion renders a recorded utterance at a different vocal intensity, rather than simply scaling its amplitude, while preserving its content and speaker identity. Existing systems control intensity through a reference utterance or a few discrete categories. Instead, we control the conversion with any target sound pressure level (SPL) on a continuous dB scale. A lightweight converter, trained as a conditional cycle-consistent GAN without parallel data, transforms the features of a frozen speech encoder, from which a pretrained vocoder resynthesizes the speech. We compare three representations, log-mel spectrograms, WavLM, and SpeechTokenizer, and introduce an objective protocol that measures the achieved intensity change against the target one on a calibrated corpus. With WavLM, the converter achieves a measured 83% of the target change in intensity and also shifts the fundamental frequency. A perceptual evaluation confirms that perceived vocal intensity increases with the target level on amplitude-normalized stimuli.

Negative polarity items (NPIs), such as "anymore" and "ever", are grammatical only in select environments, predominantly in the context of negation. Because their grammaticality depends on both syntax and semantics, they have a rich, multilingual literature in theoretical linguistics, yet work on NPIs in language models is almost entirely confined to English. We introduce BLiMPPI, a multilingual evaluation suite of over 12,000 NPI minimal pairs in 33 languages, built by a pipeline that discovers NPIs in Universal Dependencies (UD) treebanks, labels their licensing environments, and perturbs these into minimal pairs. Evaluating 12 multilingual and 2 monolingual models alongside n-gram baselines, we find that multilingual models struggle most with lower-resource languages and with pairs that are not n-gram solvable. We further examine the effects of NPI typology, language family, licensing variability, and NPI-licenser distance. Data: https://github.com/arzuburcuguven/blimppi-data

Correlation matrices combine nonlinear positive-definiteness constraints with a simultaneous row-column permutation symmetry. Their generation therefore involves two distinct problems. First, every output must remain on the correlation elliptope, irrespective of how its variables are ordered. Second, the generator must learn the domain-specific distribution occupied by empirical data inside that valid space. Existing financial image generators mix these problems: they learn ordered matrix entries and repair invalid draws afterward. We introduce CorrFlow, which separates them. A generalized Fisher transform maps full-rank correlation matrices bijectively to unconstrained coordinates, guaranteeing valid outputs without projection. A graph flow-matching network learns the empirical distribution in those coordinates and is permutation equivariant: reordering the variables only reorders the output. We compare a conditional CorrGAN reimplementation, image diffusion, and CorrFlow on $80\times80$ Taiwan equity matrices using five seeds per neural method. CorrFlow produces 100% valid draws without repair and obtains feature sliced-Wasserstein distance $.712\pm.141$, versus $4.339\pm.252$ for CorrGAN and $4.610\pm.191$ for image diffusion after their invalid outputs are projected back to the correlation space. It also improves this distance over historical lookup, while lookup retains higher local coverage. These results support the combined construction under the evaluated training budget; they do not establish superiority over fully tuned image models or isolate the contribution of each component.

Dynamic vocabularies built on-the-fly via Lempel-Ziv-Welch (LZW) compression let a language model represent the same text in fewer tokens, but prior work shows that this efficiency comes at a cost to reasoning and perplexity. We show that this quality gap is not a fundamental price of compression but can be substantially closed through targeted improvements to the compression layer and LZW codec, without modifying the underlying pretrained language model. Our analysis reveals four key issues: (1) LZW's compression algorithm creates an implicit conflict between the input embedding and the output embedding, a tension that a single shared embedding cannot resolve; (2) naively applying RoPE to LZW-compressed sequences ignores that merged tokens can span multiple tokens; (3) merging certain tokens (notably digits) disproportionately hurts downstream accuracy; and (4) dynamic vocabularies introduce an inherent segmentation ambiguity that confounds standard evaluation. We address these through decoupled input/output dynamic embedding layer, size-aware RoPE positioning, and a modified LZW codec supporting selective merges and multi-view segmentations. Across four models, our recipe improves perplexity on every corpus we evaluated as well as mathematical reasoning accuracy, closing roughly half the quality gap between zip2zip and the uncompressed baseline at a similar compression ratio on general text. This represents a significant step toward lossless adaptive-vocabulary compression for language models.

Psychology has proposed many principles of cognition. Each, however, was developed in isolation, leaving their combined explanatory power unknown. We assembled a small set of established principles into one model, which we call COntextual Retrieval and Encoding (CORE). CORE consists of a hierarchy of episodic memories into which prediction errors are written and whose encoding, retrieval, and forgetting depend on context. Fitted jointly to 7,036,442 choices from 137 experiments, the model predicts human choices better than black-box models and reproduces 75% of reported behavioral effects when simulated. Moreover, its latent variables show classic signatures of Bayesian learning and event segmentation. Taken together, these findings suggest that a small cognitive core accounts for much of human behavior. They also open a path toward making psychology more than the sum of its parts, by building computational models on the entire experimental record.

Large Audio Language Models (LALMs) achieve strong performance on speech-text tasks, but standard audio encoders produce hundreds of tokens for even short audio clips. Because self-attention scales quadratically in sequence length, this inflates inference-time compute substantially. Current training-free token compression methods prune or merge tokens based solely on coarse attention or feature similarity. However, treating audio tokens as generic, isotropic features smooths over spectral transitions, destroying fine-grained phonetic transitions essential for decoding speech. In contrast, speech signals exhibit highly structured temporal patterns where information density peaks at acoustic transitions. We propose AudioTETRIS, a training-free, two-stage framework that uses the audio encoder's representation geometry as an acoustic prior for token importance before applying attention-based compression. Instead of discarding encoder structure after the encoding stage, AudioTETRIS propagates this prior directly into the LLM decoder to compute a per-token fused score. This score drives progressive top-$k$ token selection across decoder layers without requiring additional training or auxiliary models. Evaluated on Qwen2-Audio-7B, Phi-4-Multimodal, and Audio Flamingo 3, AudioTETRIS retains near-uncompressed performance on several speech tasks using only 25% of audio tokens, matches or surpasses full-token accuracy at moderate compression rates, and consistently outperforms existing training-free baselines.

Diagnostic information in multi-view medical imaging studies, like echocardiography, is distributed across views, each capturing different aspects of cardiac structure and function. In clinical practice, these views are first interpreted through quantitative measurements, which then support specific diagnoses. However, existing echo representation learning methods do not explicitly model measurements as an intermediate representation between views and diagnosis. As a result, individual measurements are not explicitly represented in a form that downstream diagnostic tasks can selectively use, hindering accurate diagnosis. Inspired by this hierarchical clinical workflow, we introduce EchoTab, a measurement-grounded representation learning framework that represents multi-view echocardiography studies using measurement-specific tokens as an intermediate representation between echo views and diagnosis. Each measurement token selectively attends to the views relevant to its estimation and is trained through tabular reconstruction to predict the corresponding measurement value. The resulting tokens form a factorized echo representation, which encode clinically meaningful structural and functional properties and can be selectively used for downstream diagnosis. Unlike the commonly used contrastive approaches, EchoTab does not require negative samples or large batch sizes. Experiments on two datasets demonstrate that EchoTab learns effective measurement-grounded representations that improve diagnosis classification in multi-view echocardiography and its learned attention aligns with established measurement-view relationships used by cardiologists. Our code is accessible at https://github.com/michelleespranita/EchoTab.

Reinforcement Learning from Verifiable Rewards (RLVR) is a key driver for Large Reasoning Models, but it is vulnerable to "zero-reward collapse", a training failure where a sudden increase in task difficulty causes the learning signal to vanish. We argue that the conventional approach of discarding these failed trajectories is wasteful, as they contain valuable diagnostic information about the model's cognitive gaps, a bright side that can be leveraged for recovery. Existing solutions, however, often regress to static, "off-policy-style" interventions like pre-defined hint libraries or perform post-hoc refinement, failing to provide the adaptive, forward-looking guidance needed to escape the collapse. To address this, we propose a paradigm shift towards on-policy pedagogical intervention and introduce MENTOR, a multi-agent framework that acts as a live mentor. MENTOR diagnoses the root cause of failure in real-time and dynamically generates targeted, instructional prompts to guide the learning agent's next attempt. As a plug-and-play module, MENTOR integrates seamlessly into existing policy optimization pipelines like GRPO without modifying the training infrastructure. Extensive experiments on math and code generation show that MENTOR effectively mitigates zero-reward collapse, outperforming state-of-the-art baselines. Notably, it achieves massive average relative gains of up to 78.3% on Llama-3.2-3B-Instruct over its base model.

Recent results [arXiv:2602.04378, arXiv:2602.22608] have established worst-case objective-error lower bounds of order $1/T^2$ after $T$ linear minimization oracle (LMO) calls, matching the upper bound [arXiv:1406.1305] in its dependence on $T$ and thereby resolving a long-standing open question. Upon close inspection, however, both lower bounds are somewhat brittle. The first [arXiv:2602.04378] applies only to the standard Frank--Wolfe algorithm, and, as we show, its prescribed slow exact-line-search trajectories can amplify sufficiently small initialization errors exponentially, corresponding to roughly one additional bit of initialization accuracy per iteration. The second [arXiv:2602.22608] obtains lower bounds for any deterministic LMO-based first-order method via a resisting oracle that adversarially adapts to the queries of the algorithm. In both cases it is conceivable that randomization may circumvent these lower bounds; in particular, neither establishes a lower bound for unrestricted randomized first-order LMO methods. We close this gap by proving an $Ω(L\operatorname{diam}(S)^2/T^2)$ distributional lower bound for randomized first-order linear minimization oracle methods over strongly convex sets. While our construction reuses the family of [arXiv:2602.22608], our proof characterizes the conditional distribution of the hidden permutation after adaptive queries and bounds the expected decrease of a combinatorial statistic controlling optimizer variance. We further show that the same error threshold holds with probability at least $3/4$ and yields a stopping-time tail bound.

Uncertainty quantification for electric vehicle (EV) charging demand is essential to turn the increasing pressure of EV adoption on power systems into a source of flexibility. We study a large-scale dataset of public charging stations in Scotland, where demand exhibits a hierarchical structure, organized across multiple aggregation levels, from individual charging points to regional and national scales. The charging infrastructure is continuously evolving, with new stations deployed and existing ones occasionally removed, which calls for forecasting and uncertainty quantification methods tailored to this evolving structure. To address these challenges, we combine hierarchical forecasting with recent advances in conformal prediction with missing data. The conformal procedures considered produce valid prediction intervals at every node of the hierarchy and the algorithm CP-MNR-Nested successfully accounts for the time-varying hierarchical structure of the data.

Detectors of machine-generated text score almost perfectly on scientific-literature benchmarks, implying we can reliably distinguish with these detectors human and machine-generated documents. We looked at ten scientific-document subsets from six popular resources, and on several of them that apparent ease is a side effect of how the documents were put together. A logistic regression using 27 superficial features reached 0.833-0.997 macro-F1, and on the three arXiv subsets it outperformed every detector we evaluate with F1-scores 0.951-0.997. The strongest cue to the detector is the PDF extraction artifact: human texts retain their original line breaks, whereas generated texts are extracted as a single block of prose. This cue, together with three other formatting features, tell the classes apart at 0.996 AUC. To distinguish benchmarks that contain such a signal from detectors that rely on one, we introduce format counterfactuals: just re-wrapping a machine-written abstract without changing a single word reverses 20-27% of a public detector's correct decisions. Normalizing formatting before training reduces the generalization gap from 0.32 to 0.22. On format-clean benchmarks, it also raises out-of-domain macro-F1 from 0.33 to 0.65 for detectors trained on artifact-containing subsets.

Pre-trained robot policies, such as Vision-Language-Action (VLA) models, provide a strong foundation for generalist robotic manipulation, but adapting them to new tasks with only a few demonstrations remains challenging. Directly fine-tuning a pre-trained policy often achieves strong in-distribution performance, yet it can distort the representations learned during pre-training and weaken the model's generalization ability under out-of-distribution shifts. We propose Retrieval-Anchored Representation Alignment (RARA), a two-stage adaptation method, to address this issue. First, RARA retrieves representation features from pretraining with similar task intent and motion, and aligns the policy's representations of the target demonstrations with the pre-trained representations of the retrieved demonstrations. Second, RARA fine-tunes the policy on the target demonstrations while limiting how much its representations of the retrieved demonstrations change during fine-tuning. By reducing the representation mismatch before action learning and preserving pre-trained knowledge during adaptation, RARA limits unnecessary representation drift during fine-tuning. Experiments with Diffusion Policy and GR00T on 13 tasks in RoboTwin, RoboCasa, and on a real robot show that RARA improves the average success of fine-tuning by 9.5 points in-distribution and 13.6 points out-of-distribution, and outperforms all compared baselines. Project page: https://shua-kang.github.io/RARA_robot/

Feature selection is a central problem in high-dimensional financial regression, where large numbers of engineered variables and strong correlations among predictors reduce model stability and interpretability. $L_1$ regularization (Lasso) is widely used for sparse regression, but its shrinkage can be restrictive when informative signals are spread across correlated variables. Exact sparse regression avoids this shrinkage but requires solving a combinatorial optimization problem that is computationally difficult and is typically fit to minimize in-sample rather than out-of-sample error. This paper presents a feature selection framework based on Monte Carlo Tree Search with Generalized Nested Rollout Policy Adaptation (GNRPA). Feature selection is formulated as a sequential decision process, and candidate subsets are scored by their out-of-sample validation performance rather than training error. The method also supports a lexicographic objective accounting for subset size and multicollinearity. We study this approach on a contemporaneous explanatory regression task: explaining the weekly return of the NASDAQ-100 index from 1,072 engineered features spanning market, macroeconomic, and factor data. This isolates the value of subset selection under multicollinearity from the separate problem of predictive forecasting. On this task, the proposed method identifies sparse feature subsets with higher out-of-sample explanatory power than an optimized Lasso baseline, reaching an out-of-sample $R^2$ of 0.6961 versus 0.6313 for Lasso. These results suggest that Monte Carlo Tree Search is a practical alternative to convex regularization for sparse explanatory feature selection in financial regression, with potential value for factor attribution and risk decomposition where transparency matters as much as out-of-sample fit.

We study simultaneous competitive and regret guarantees for TCP acknowledgment. Each acknowledgment serves all pending packets, and the objective is to minimize the total acknowledgment and waiting costs. It is well known that no deterministic online algorithm achieves a competitive ratio better than $2$. Given $k$ randomized online experts and $0<\varepsilon<1$, we construct a randomized online algorithm whose realized total cost is always at most $(2+\varepsilon)$ times the offline optimal cost $\mathrm{OPT}$ and whose expected regret against the expert with the smallest realized cost is $O(\sqrt{\mathrm{OPT}\log(k+1)}+\log(1/\varepsilon))$. The algorithm requires no advance knowledge of the number of packets or $\mathrm{OPT}$. Our algorithm controls switching costs between experts and uses the gap between the permitted competitive bound and its current cost to determine when to learn. We also prove lower bounds showing that, for $k\ge 2$, this regret bound is asymptotically optimal in $\mathrm{OPT}$, $k$, and $\varepsilon$, and that $\varepsilon>0$ is necessary for sublinear regret.

Reasoning models can benefit from test-time scaling, but continued reasoning may become ineffective or harmful, a phenomenon termed overthinking. Evaluations against full thinking compare accuracy and token usage but lack references for overthinking severity and strategies' theoretical mitigation potential. To characterize overthinking at the model-task level, we first introduce Step-wise Truncation Probing (STP), repeatedly generating answers from multiple retained prefixes along each fixed sampled reasoning trajectory. Using each trajectory's earliest measured position attaining maximum empirical accuracy as a reference, we define Overthinking Ratio and Accuracy Degradation as the fraction of thinking tokens generated afterward and the accuracy decrease from this reference to full thinking, respectively. Secondly, we then formulate dynamic routing, fixed budgets, and dynamic early exit within a unified framework to quantify the theoretical mitigation potential of different termination strategies. Based on this unified framework, we use the complete STP measurements to construct a token-accuracy Pareto frontier for each strategy under a common normalized thinking-token budget. We further develop OTBench, a reusable benchmark built from the STP process, which produced 70.5B reasoning tokens and the corresponding probing measurements across six models and four benchmark groups.

We present Rationally Linearized Multilayer Perceptrons(RaLiMPs), a new approach to efficiently linearizing MLPs using low-degree rational approximations of standard activation functions. Inspired by polynomial-based MLP linearization, we show that rational approximations can provide accurate representations at substantially lower degree. RaLiMPs exploit these approximations to construct kernel representations with narrow bottlenecks, enabling efficient composition across MLP layers. We provide theoretical guarantees for the resulting approximation and evaluate RaLiMPs across applications ranging from implicit 3D models to large language models. Our experiments demonstrate favorable accuracy--efficiency tradeoffs, with RaLiMPs achieving competitive or improved accuracy while substantially reducing inference cost.

Repository-level issue-resolution benchmarks such as SWE-bench evaluate submitted patches primarily based on whether they pass the correctness tests, leaving the runtime performance of different test-passing patches largely unexplored. Evaluating patch performance is essential but challenging since designing performance workloads for each benchmark instance is costly. However, these benchmarks already provide a natural candidate: the built-in \emph{FAIL\_TO\_PASS} (F2P) tests, which exercise the behavior affected by the issue fix. We therefore investigate whether the provided F2P tests can be repurposed to evaluate patch performance. Overall, we choose four frontier models, Claude Opus~5, GPT~5.6, DeepSeek V4 Pro, and Qwen~3.8 Max, paired with SWE-agent to resolve the pre-filtered measurable corpus from verified splits of SWE-bench and SWE-bench-Live, yielding 262 cases with at least two different correct solutions, including the developer ground-truth patch. We then measure each solution's runtime performance in instruction count and wall-clock time on the native F2P test, and use Welch ANOVA instead of the RSD proposed in prior work as a more robust statistical discrimination criterion. We find that this methodology exhibits notable discrimination capability in runtime performance and identifies performance differences among patches in almost half of the instances. Our empirical study shows that the runtime performance of generated patches is independent of the issue resolution rate of LLMs, with nearly reversed model rankings. The case study also shows that the performance analysis reveals architectural regression in some patches that the correctness evaluation overlooked. These findings suggest that resolution rate does not adequately characterize LLMs' capability for repository-level issue resolution, motivating future evaluations to consider patch performance alongside it.

Anticipation of vessel trajectories is an essential task when navigating at sea. This applies to both human navigators and autonomous surface vessels. Trajectory forecasting enables sub-routines such as collision avoidance to assess the situation and react accordingly in time. This paper investigates Transformer and Performer models for trajectory prediction of surface vessels in confined waters. Using widely available position data from the Automatic Identification System (AIS), candidate models are trained and evaluated. Historical information on how other vessels have moved in the local area surrounding each agent is conveyed using positional heat maps. The heat maps are generated using Kernel Density Estimation (KDE) and encoded with a Convolutional Neural Network (CNN) before entering the Transformer models. The models are trained with quantile loss to efficiently obtain a model-aware prediction interval, as this is crucial for timely collision avoidance. The proposed model architectures are shown to greatly outperform previous work while the inclusion of local heat maps enables the prediction of vessel movement that is not otherwise indicated from the previous positions.

Recent speech LLMs can leverage rich paralinguistic cues, enabling more complex and personalized interactions. However, little work investigates how this sensitivity can lead to stereotyping, and existing evaluations largely measure demographic disparities without distinguishing beneficial personalization from stereotypical behavior. We introduce BASIS-Bench, a benchmark designed to disentangle these two aspects through three axes: (1) demographic robustness, measuring disparities in content-based tasks across speakers and demographic classification capabilities; (2) stereotypical bias, measuring whether perceived demographic cues trigger unprompted stereotypical behavior; and (3) personalization, measuring whether models can tailor responses to demographic characteristics when explicitly requested. We evaluate eight recent speech LLMs and find that, despite often recognizing demographic cues, most fail to translate them into appropriate personalization. Explicit demographic labels improve personalization but also amplify stereotypical associations already present in the textual backbones. These findings highlight the need to train speech LLMs not only to perceive demographic cues, but to learn when they should shape their responses.

When a language model is distilled from a teacher that shares its base parameters, the student can inherit behavioural traits of the teacher---including misalignment---even from data that is semantically unrelated to the trait (Cloud et al., 2026). All existing defences require knowing the trait. We remove this requirement and prove that a subliminally transferred trait can be recovered from the parameter updates alone, without supervision, by exploiting a single asymmetry: the trait is expressed consistently across diverse datasets generated by the teacher, whereas task-specific and shared capabilities vary. The resulting estimator returns a weight-space direction that can be amplified to discover and interpret the trait or projected out to suppress it. Applied to the subliminal-learning benchmarks of (Cloud et al., 2026), our method recovers the transmitted animal-preference traits without any trait labels and suppresses their expression with almost no impact to task performance.

Long-horizon embodied agents must maintain task-relevant information, coordinate heterogeneous robots, and adapt their decisions as execution unfolds. Existing benchmarks investigate dynamic task execution and embodied multi-agent coordination, but provide limited evaluation of their intersection. We introduce MAGMA-Bench, an architecture-agnostic benchmark for persistent task reasoning in dynamic multi-robot environments. MAGMA-Bench evaluates agents across increasing task horizons and controlled online perturbations, including execution failures, task interruptions, and mission updates, with scenarios requiring coordination across robots with heterogeneous capabilities. It uses a structured skill-level interface to separate task-level reasoning from perception and low-level control. We evaluate locally deployable open-weight planners, multi-agent state-tracking strategies, and frontier models. Even the strongest evaluated models fail on a substantial fraction of clean episodes, and their success rates decrease further under online perturbations. Supervised adaptation greatly improves a small model, but performance remains substantially lower on longer and multi-robot missions. We release the benchmark, 261,646 task-level training examples, and an architecture-agnostic expert-trace generation pipeline.

Multimodal language models (MLLMs) exhibit promising spatial reasoning capabilities, yet the internal mechanisms supporting them remain poorly understood. We uncover S-Space, a linear representation subspace in which object-token representations encode continuous coordinates along horizontal, vertical, and distance axes. Although we identify these axes by differentiating spatial-answer scores with respect to object-token activations, S-Space is intrinsic rather than induced by specific data or prompts, remaining readable under non-spatial questions. We show that S-Space functions as an internal spatial workspace of MLLMs. Perception encodes object locations into it, forming spatial representations that can be directly read out to support accurate spatial judgments. Internal reasoning transforms it for changing viewpoints. Verbal responses selectively access it, evidenced by causal interventions that alter spatial judgments while largely preserving non-spatial responses. Case studies further show that S-Space assigns locations to unseen entities and the observer, and connects physical directions with abstract linguistic meanings, suggesting both the capabilities and limitations of large-scale language training. Finally, we use S-Space as a diagnostic tool for spatial reasoning failures. Direction-word readouts reveal systematic coupling between cardinal and image-relative reference frames. More importantly, a perception-computation decomposition exposes a key bottleneck: MLLMs can form capable spatial representations, yet fall short in the spatial transformations required to reason reliably across viewpoint changes.

Intermittent ``spikes'' in loss, during the pre-training of Large Language Models (LLMs), remains a mysterious phenomenon both theoretically and empirically. In this paper, we show that simple quadratic models with heavy-tailed, power-law distributed inputs can replicate these spikes. We introduce a theoretical framework and simulation algorithm to model per-layer training dynamics of Deep Neural Networks (DNNs), using such quadratic models. In a case study on an Adam hyperparameter configuration (where $β_2$ is set close to $1$) known to induce spikes in Transformer LMs, our simulation successfully predicts bumps indicating spiky behavior from a given checkpoint. Furthermore, we derive an appropriate $β_2$ schedule from observed metrics, which empirically improves performance.

Multi-teacher on-policy distillation (MOPD) extends on-policy distillation (OPD) by routing domain-specific data to specialized teachers, integrating their capabilities through dense token-level supervision on student-generated trajectories. However, domain routing only distinguishes the supervision sources for different domains; updates from all domains still act on shared parameters. When teachers differ substantially in response style and domain expertise, supervision that is effective in isolation may still lead to persistent gradient conflicts and capability degradation during joint optimization. To address this challenge, we propose Cross-Teacher Guarding for MOPD (Road MOPD), which introduces non-routed teachers as auxiliary cross-domain critics on the same student trajectories and coordinates multi-teacher supervision through token-level credit assignment and gradient updates. Specifically, cross-teacher-guided credit assignment uses teacher agreement to strengthen token-level corrections jointly supported by the teachers and attenuate conflicting signals; preventive multi-teacher gradient projection uses auxiliary gradients from the same trajectories to constrain potentially conflicting updates and adaptively controls projection strength to retain target-domain learning signals. In experiments, Road MOPD mitigates degradation in the disadvantaged domain, gains over 5 percentage points under high conflict, and improves overall performance in low-conflict and three-domain settings.

Identifying causal effects from observational data typically requires a fully specified causal graph, an assumption that is often unrealistic in complex or high-dimensional settings. Causal abstractions address this limitation by retaining only partial causal information, thereby defining equivalence classes of compatible causal graphs. In this paper, we study causal effect identifiability directly at the level of these graph collections, rather than at the level of individual causal diagrams. We formalize several notions of identifiability in this setting and organize them into a hierarchy that clarifies their relationships and differences. In particular, we establish that a causal effect may be identified over a collection of graphs without admitting a single causal derivation that holds uniformly across all compatible graphs. These results help reveal gaps and clarify ambiguities in the existing literature.

In machine learning, a self-attention dynamics describes the attention mechanism of transformers, and is obtained treating the layer index of a transformer as time index. Even though naturally discrete-time, this dynamical system is most often studied in continuous-time. Scope of this paper is to provide a thorough investigation of the discrete-time version of the self-attention dynamics: a complete classification of its equilibria is given, and their stability properties are studied, in both the single-head and the multihead case. We show that in both cases the main attractor of the dynamics is a consensus equilibrium aligned with the principal eigenvector of the value matrix (in the single-head case, a corresponding matrix of weights in the multihead case), even though other locally asymptotically stable bipartite equilibria can be show to exist at least in the single-head situation. The behavior is confirmed by looking at the learned weight matrices of a real transformer.

Runtime failure monitors for robot policies are typically evaluated by how accurately they distinguish failed from successful executions. Yet an accurate alarm may arrive too late for recovery, while an unnecessary intervention may disrupt an execution that would otherwise succeed. We introduce RESCUE, a counterfactual framework for evaluating whether a monitor's alarms enable useful interventions. By restoring simulator and policy states at execution checkpoints, RESCUE compares alternative actions, identifies the latest checkpoint satisfying a specified recovery criterion, and evaluates whether acting on alarms reduces expected task-completion cost. We evaluate RESCUE with OpenVLA-OFT and $π_{0.5}$ over LIBERO-Spatial and LIBERO-Plus using 128,340 counterfactual branches. Recoverability depends on the state--intervention pair rather than on robot state alone: at $τ=0.5$, 14 of 66 OpenVLA-OFT/LIBERO-Plus failures contain a recoverable checkpoint under continuation, compared with 34 of 66 under continuation or scripted recovery, while recovery reduces recoverability in other conditions. On the primary strong-policy conditions, favourable detection scores do not establish lower intervention cost. Under the evaluated stop-and-reset model, even an idealised detector can afford only a 2.0--3.3% false-alarm rate before unnecessary resets outweigh the savings from stopping failures early. These results show that monitor value depends jointly on the intervention available and the workload in which it is deployed.

Clinical prognosis requires integrating structured measurements, physiological waveforms, longitudinal history, and population context. We introduce PRISM (Population-Retrieved Inference for Structured Multimodal Prognosis), a multimodal language model that combines per-feature numerical tokenization, recency-weighted aggregation of longitudinal ECGs, and a dual-modality population memory bank. Retrieval operates in representation space: the model receives mean-pooled embeddings of similar training-cohort encounters, while no individual retrieved record enters the prompt. On the 1,443-task MDS-ED benchmark, PRISM achieves AUROCs of 86.2% for diagnosis, 93.1% for deterioration, 91.8% for ICU admission, and 92.8% for mortality. These results use a 96-hour pre-arrival window unavailable to reported baselines. With benchmark-window inputs only, PRISM reaches 0.8895 AUROC, compared with 0.8937 for UniPACT and 0.8889 for MDS-ED. A matched-input compact fusion model using the same five evidence streams and frozen encoders reaches 0.7652 AUROC, versus 0.9092 for PRISM. Replacing numerical EHR encoding with text serialization reduces AUROC by 3.66 points; removing pre-arrival history reduces it by 1.97 points; and zeroing retrieval reduces it by 0.99 points. Genuine retrieval outperforms random-neighbor retrieval by 0.97 macro-average points, although a paired task-level sign test does not establish a uniform advantage (p=0.067). On the external MC-MED cohort, supervised fine-tuning reaches 0.8611 AUROC, compared with 0.8227 for locally trained gradient boosting and 0.7841 for source-only PRISM. These results support structure-aware multimodal integration and supervised external adaptation.

Federated learning (FL) seeks to optimize a shared global model from decentralized client data. While effective for average performance, this process can produce unfair performance across clients, particularly under statistically heterogeneous data distributions. Existing fairness-aware and multi-objective FL methods address this through client re-weighting, gradient correction, or common-descent formulations, but often change the global optimization direction or treat client objectives symmetrically. We propose FedPCD, a priority-constrained federated optimization framework that separates the global optimization objective from client protection. FedPCD preserves the conventional federated aggregate as the primary objective, while allowing only the minimum correction necessary to meet client performance constraints. Across heterogeneous FL benchmarks, FedPCD consistently outperforms state-of-the-art baselines, achieving the highest global accuracy across all configurations and leading worst-10% client accuracy in nearly all. FedPCD further preserves competitive trade-offs between global accuracy and client performance variability. Beyond this trade-off, FedPCD demonstrates faster empirical convergence, reaching matching or superior performance in substantially fewer communication rounds.

Generative Recommenders increasingly use Reinforcement Learning (RL) to optimize outcomes that may appear hundreds or thousands of interactions after a decision. Standard RL propagates reward according to temporal proximity rather than explicit relevance, attenuating the signal assigned to distant but important actions and creating a source of sample inefficiency. We introduce Explanatory World Models (EWM), an attribution framework that uses explanations to identify the earlier events that plausibly contributed to a realized outcome. We instantiate EWM with a frontier language model and use its explanations as learning signals using a conservative hierarchical operator that recursively redistributes reward while preserving the undiscounted episode return. To investigate what constitutes a useful explanation, we operationalize three complementary variants inspired by Halpern's actual causation, Lombrozo's explanatory simplicity, and Deutsch's hard-to-vary explanations. We evaluate EWM on two controlled long-horizon RL benchmarks, a public sequential-recommendation benchmark, and an industry-scale generative recommender. On Crafter and WordCraft, EWM substantially outperforms PPO and LaRe, with its largest gains when rewards are most delayed. On KuaiRand, EWM outperforms SFT and LaRe offline and A2C online. Finally, in an offline evaluation of an industry generative recommender, EWM considerably improves Recall and increases behavioral proxies. These results suggest that explanations can provide structured and reusable credit for learning from delayed outcomes.

Existing LLM-based agent approaches for long-horizon machine learning engineering (MLE) often use context engineering, implemented through custom harnesses that manage tools, skills, and accumulated experience. However, we find that as single-session execution lengthens, the agent's reasoning intensity declines alongside slowing progress, summarized as long-horizon deceleration, and the benefits of these custom harnesses vanish as the underlying model improves, summarized as harness staleness. These findings suggest that their benefits may not persist across versions, which motivates a workflow-level approach in which agents build on one another's work across sessions. We model these alternating agent sessions as a context-reset semi-Markov decision process (CR-SMDP) and establish sufficient conditions for alternation to exceed single-agent reference limits, derive an explicit experiment-budget bound, and characterize when session caps preserve this advantage. Guided by this analysis, we introduce Humanize MLE Agents (HMA), which alternates two agents on a shared workspace, limits each session to at most k experiments and resets the context between sessions, without additional task-specific harness or prior knowledge supplements. Across 75 Kaggle tasks in MLE-bench, HMA reaches a state-of-the-art any-medal rate of 78.2%, exceeding the strongest open-source baseline by 8.0 points with a quarter of its time budget. Moreover, among six configurations spanning five heterogeneous agent pairs, HMA improves any-medal rate over the constituent baselines by 6.4 points on average, demonstrating cross-pair generality. A supplementary case on kernel optimization reaches 878 cycles, within 1.6% of the community leaderboard leader, demonstrating applicability beyond MLE.

Agentic orchestration delegates tasks to sub-agents with independent contexts, providing a key path to extending the effective execution horizon of large language model agents. However, existing orchestration methods face two limitations: in execution, current methods fix sub-agents' goals and harnesses at creation time, limiting their adaptability to dynamic real-world environments; in optimization, sparse and delayed feedback constrains effective learning of the orchestration policy. To address these limitations, we introduce POrchestra, an on-the-fly proactive agentic orchestration framework. It allows sub-agents to proactively communicate with the orchestrator at any logical point to request harness adjustments while maintaining continuity of local execution. This mechanism not only enables sub-agents to adapt seamlessly to dynamic environments but also naturally inspires On-Policy Orchestration Distillation (OPOD) based on hindsight feedback. OPOD attributes sub-agents' proactive revision requests back to preceding orchestration decisions to derive targeted improvement signals, and leverages these signals as process supervision alongside the final reward to optimize the orchestration policy. We evaluate POrchestra on three long-horizon agent benchmarks spanning deep research, dynamic digital assistance, and software engineering. POrchestra improves task success by an average of 9.1% relative to the strongest baseline; combined with OPOD, it improves small-orchestrator success by an average of 10.0% relative to existing training baselines. POrchestra also achieves a more favorable cost-performance trade-off. Our code is publicly available at https://github.com/XueyangFeng/POrchestra.

On-policy distillation (OPD) provides teacher supervision on student-generated trajectories, but a single trajectory realizes only one continuation at each decision, limiting supervision coverage under a finite sampling budget. We propose BranchOPD (Branching On-Policy Distillation), which selects informative decision points using student uncertainty and teacher-student disagreement, resamples student-policy continuations from shared prefixes, and applies token-level teacher supervision to both main trajectories and branches. BranchOPD operates at token-level decisions in mathematics and within-turn decisions in agent tasks. Shared-prefix reuse reduces student-generated token costs by approximately 19% on ALFWorld and 31% on WebShop relative to seven independent trajectories. We also introduce protocol-boundary repair (PBR), which uses targeted supervision under protocol constraints to mitigate repetition and wasted generation. BranchOPD outperforms Vanilla OPD and Multi-Rollout OPD (n=7) across all evaluated benchmarks, with gains over Vanilla OPD of 18.55 percentage points in mean mathematics accuracy over 16 sampled answers (avg@16) and 13.8 points in mean ALFWorld ID/OOD success. Mechanistic analyses and selection ablations show that branching at uncertain decisions expands subsequent behavior coverage and outperforms random selection and turn-initial branching.

Nonlinear model predictive control (NMPC) can be applied to large nonlinear systems with complex objectives while enforcing constraints, but the cost of solving the underlying optimization problems limits real-time application for systems with fast dynamics or many states. Neural networks can reduce this online cost by approximating the control law, but may violate constraints and typically require many precomputed optimizer solutions for training. The learning-based iterative solver for constrained optimization (LISCO) instead learns to predict primal-dual steps toward satisfaction of the Karush-Kuhn-Tucker (KKT) conditions through self-supervised training without precomputed solutions, but its dense, monolithic architecture grows with the prediction horizon. We remove this horizon dependence by exploiting the temporal structure of the problem in the solver architecture, which makes the network size independent of the horizon while leaving the training loss and the solver iteration unchanged. A closed-loop resampling strategy further aligns the training distribution with closed-loop operation at no additional sampling cost. On an economic NMPC benchmark of a CSTR cascade, the proposed solver scales to larger problems than monolithic LISCO and achieves median speedups of up to a factor of four over IPOPT. In closed loop, it achieves comparable control performance to IPOPT with no observed constraint violations.

Generating edge case scenarios is crucial for training, testing and safety assuring autonomous systems. This identifies hazardous failures to increase system reliability and robustness, ensures the system can still function safely on the boundaries of its operational design domain and learns the expected system behaviour for these edge cases. The rapid advancement in the generation and analysis capabilities of language models (LMs) has expanded their applicability and recent work has investigated LM scenario generation. However, many recent approaches rely on human-in-the-loop or rule-based strategies. This restricts the LM's ability to generate diverse and edge case scenarios; the long tail of the distribution. Our work uses local LMs in a laptop-based agentic AI framework for adaptive scenario generation. This is cheap, fully-controlled, ensures data privacy and security. It uses think, act, observe, update, and continue loops with lightweight JSON encoding to generate scenarios. We use the system requirements, system description and environment description as knowledge bases in contextual prompts for an LM to ensure scenario realism. A second, independent LM critiques each scenario using domain specific criteria so it can be refined. The refined scenarios are run in simulation and evaluation metrics collected. These metric scores are analysed and similar scenarios are created for edge cases (exploitation) or novel scenarios are generated to explore the scenario space. Post-hoc clustering using metric scores and thematic-type analysis of the rich-text scenario descriptions produced by the LM pinpoints functional gaps and their triggering conditions. This creates traceable evidence for the safety assurance case. Our work highlights the potential of a flexible and adaptive local agentic AI framework for automating edge case generation and producing traceable safety assurance evidence.

We derive approximation rates for Sobolev functions defined on the complex unit sphere and complex projective space, representing quantum states, and on the special and projective unitary groups, representing quantum unitary operations. By using harmonic analysis on compact groups and compact homogeneous spaces, we characterize the corresponding Peter--Weyl decompositions and Sobolev spaces via irreducible representations and the Laplace--Beltrami operator. Combining spectral truncation with finite-depth quantum-circuit approximation, invoking the Solovay--Kitaev theorem, we establish quantitative rates that separate the spectral truncation error from the circuit approximation error.

Representational similarity measures (RSA, CKA, Procrustes) are chosen for the transformations they treat as irrelevant, and all of them take the set of conditions as given. That set is fixed by two decisions, which inputs are admitted and how they are grouped, both taken before the matrix a measure operates on exists: no invariance property reaches them. We measure what they are worth where the arrangement the conditions ought to have is known independently of the systems compared, the acoustic organisation of vowels, across fifteen languages and twenty-seven representations. Changing which categories are compared, at a fixed number of them, moves the gap between two representations by more than twice the effect being measured, while changing how many moves it not at all. In this setting rankings of models are identified under these choices, and rankings across languages are not.

Multimodal LLMs can fail to read text embedded in an image's coarse structure (text a human recovers by squinting) and report that no text is present. We name this class Squint Text and study it as a recognition deficiency, motivated by cases from a production moderation pipeline. We introduce SquintBench, comprising 1076 production-derived moderation images, and SquintCtrl, a controlled synthetic suite for mechanism diagnosis. Frozen-weight analysis shows that hidden text leaves a coarse signal while glyph identity is weakly represented; identity becomes substantially more accessible when the correct spatial regions are supplied, implicating spatial selection as a major bottleneck. Suppressing high-frequency content also improves reading at a fixed visual-token budget, while visual token count may also contribute. These findings motivate Squint OPSD: a frozen teacher receives a downscaled view and the ground-truth transcription as privileged context, and its token-level distributions supervise a student along on-policy trajectories. At inference, the student reads the original image alone. Squint OPSD reaches 88.61 ANLS on SquintBench, versus 26.50 for the unadapted backbone, 24.44 for the best general-purpose model on the original view, and 68.60 at best for an off-the-shelf model under a downscaled view. These results substantially narrow the gap between available visual evidence and its generated read-out without auxiliary views at inference.

AI systems are making rapid progress in mathematics, but this progress is hard to measure. Benchmarks of solved problems saturate, and in benchmarks of open problems each problem can be resolved for the first time only once. We present SqueezeBench, a benchmark that is resilient to saturation: a success raises the bar instead of removing the problem. Each problem fixes an integer function $f(n)$ whose values are known only for small $n$, and the benchmark keeps a lower and an upper record on it. A record is a computable Lean function with a machine-checked proof that it bounds $f$. A submission proposes a new function and proves that it bounds $f$, that it is no worse than the current record at any $n$, and that it is strictly better at some $n$. Once accepted, it becomes the record that the next submission has to beat. The system thus chooses the bound as well as its proof, and it may improve the record at a single $n$ or its growth rate. At every $n$, the gap between the two records shows how far the certified bounds are from determining $f(n)$. Every proof is checked by two independent Lean kernels, so no referee needs to check the mathematics. The checks also reject a function that computes $f$ by exhaustive search at small $n$, and a maintainer only reads the submitter's report and looks for such a search at larger $n$. We release ten problems from extremal combinatorics with twenty proved initial records and the admission tools. As a baseline, agents based on a frontier language model, with at most one hour per record, had 82 submissions accepted and improved every record, without going beyond the published literature. Agents based on two newer models then started from these records, had 60 more submissions accepted and improved 19 of the 20 records again. For eight of the ten problems, at least one record is still asymptotically behind the best published bound.

Cyber-Physical Systems are developed through multiple model based activities, comprising several stages of development. As the development progresses through simulation, experimentation, and operation, observations of system execution may reveal missing, uncertain, or invalid design knowledge. The path from execution results to persistent design-time enrichment, however, is often implicit. This paper addresses this challenge by proposing an evidence driven workflow for systematically transferring knowledge produced during simulation and operation back into design-time models. We demonstrate the workflow using an Influence Model, a design-time artefact used to represent how design artefacts and environmental factors jointly affect system response properties, using a sign-following robot use case. We show how simulation evidence can enrich Influence Models by supporting participant pruning decisions, interaction characterisation, influence-function abstraction, and validity-scope information. The application demonstrates the feasibility of the workflow for simulation-based evidence and its potential benefits to improve design-time reasoning and feedback in subsequent development iterations.

Large language models (LLMs) have taken the world by storm. Their ability to communicate with us in our own language, and to reason about our world, has made them a favorite tool for writing, answering questions, and much more. Their answers already influence our decisions and our actions. What would happen if LLMs themselves could act in the world directly? What would happen if LLMs could interact with each other? Agentic LLMs combine the power of large language models with the ability to act in the real world. In this book we will discuss how agentic LLMs work, how they act, how they interact, and what this means for us, as users.

Speech recognition on low-cost microcontrollers is, in practice, limited to keyword spotting or a fixed list of commands. We present Oído, an inference engine that runs open-vocabulary English speech recognition entirely on an ESP32-S3 (240 MHz dual-core Xtensa LX7, 512 KB of SRAM, 8 MB of PSRAM, 16 MB of flash, no neural accelerator). Oído executes a 13M-parameter Conformer-CTC model with int8 weights, per-frame dynamic int8 activations and int8 attention on the chip's SIMD instructions, with a schedule designed around flash and PSRAM bandwidth. On LibriSpeech test-clean/test-other the on-device arithmetic reaches 3.70%/8.23% word error rate (WER), within 0.12 points of full precision, and 3.33%/7.15% with an on-chip 1.3M-parameter language model; a transducer head gives 2.99%/6.67% and an int4 variant fits in 8.3 MB. This is 2.3-2.8x fewer errors than the vendor's recognizer on the same chip, and fewer than Whisper tiny.en and Moonshine tiny in full precision on a laptop, including in real noise and reverberation. Measured on a physical board, the firmware runs at 1.97x the duration of the audio on one core, with transcripts identical to an instruction-exact emulator: accurate, but not yet real time (a 4 s command is transcribed about 8.7 s after it ends). The two-core schedule gives wrong results on silicon and is disabled. A streaming variant shows partial text while the speaker talks and delivers the final text 4.1-11.1 s after clips of 3.6-10.6 s (7.2-22.4 s in utterance mode) at 4.86%/11.03% WER; a Spanish model makes 34-60% fewer errors than multilingual Whisper tiny. Code and models are released.

Compression schemes provide a general framework for deriving data-dependent risk bounds for learning algorithms. For learning rules based on preferent compression, existing results upper-bound the risk in terms of a sample-dependent complexity $\boldsymbol{k}$, defined as the cardinality of the compression augmented by the training examples on which the reconstructed hypothesis is inappropriate. These results, however, yield a margin of order $O(\sqrt{\ln N/N})$ above $\boldsymbol{k}/N$. In this paper, we establish that the logarithmic factor is not intrinsic. We prove that, for every fixed confidence level $1-δ$, the risk is bounded with probability at least $1-δ$ by $$ \frac{\boldsymbol{k}}{N} + O\left(\frac{1}{\sqrt{N}}\right), $$ uniformly over all possible values of $\boldsymbol{k}$. Thus, risk estimates for learning algorithms based on preferent compression attain the same law-of-large-numbers rate as the empirical estimate of the probability of a fixed event. We also show that the leading term $\boldsymbol{k}/N$ is necessary and investigate its relation to the empirical risk. In particular, we identify regimes in which the empirical risk retains the $O(1/\sqrt{N})$ convergence rate, as well as regimes in which it converges more slowly or ceases to be consistent, while the complexity-based estimator preserves the $O(1/\sqrt{N})$ rate.

Recent accent classification systems can achieve high accuracy in standard in-corpus evaluations. However, when speakers overlap between the training and test sets, such results may overestimate model robustness compared with real-world deployment scenarios. This study revisits the robustness of DistilHuBERT-based accent classification by using L2-ARCTIC and multiple external English accent datasets. The comparison covers file-level and speaker-disjoint splits, along with balanced external-data training and teacher-student pseudo-labeling. The large performance gap between in-corpus and external evaluations suggests that file-level evaluation substantially overestimates performance, while speaker-disjoint evaluation can provide a more realistic estimate of the generalization to unseen speakers. In this study, models trained on external data alone do not reliably generalize to L2-ARCTIC. Among the tested configurations, balanced external-data training offers the most promising supervised setting. The pseudo-labeling experiments further indicate that confidence-based filtering alone is insufficient under domain mismatch. Overall, this study suggests that robust accent classification evaluation should simultaneously include speaker-disjoint evaluation and cross-corpus testing.

Collective Autonomous Vehicle (AV) route choices in real-world traffic networks can create externalities for Human-Driven Vehicles (HDVs), as routes that benefit AVs may increase travel times for other road users. To address this, we introduce a social component, the marginal cost, to the selfish travel time AV reward. Marginal cost captures an AV's impact on other road users by comparing their travel times with those in a counterfactual simulation in which the AV is removed. We study AV route choice in mixed-autonomy traffic settings using Multi-Agent Reinforcement Learning (MARL), where AVs learn routing policies through repeated interactions with the traffic environment. Experiments on the real-world networks of Ingolstadt and Provins show that incorporating marginal cost into the AV reward not only limits the increase in HDV travel times and shortens AV travel times relative to selfish AV routing, but also improves overall system performance. In Ingolstadt, selfish routing of 40\% AVs increases the travel times of other vehicles by 1.7\%, and this drops to 1.1\% when the social reward is included. Interestingly, not only does overall system performance improve, but the AVs are also better off when acting socially.This effect is consistent across tested cities, penetration rates, and replications, yet when the weight of the social component in the reward increases, the system fails to reach the desired performance. For a more scalable approach to social AV routing in mixed-autonomy traffic, we compute the marginal cost only for the AVs whose route choices most affect the traffic system. This reduces the number of counterfactual simulations and achieves similar performance at a significantly lower computational cost. Our results suggest that AV routing should go beyond selfish rewards and, for the benefit of all traffic participants, include a social component in its objective.

Reinforcement learning (RL) has emerged as a key paradigm for improving large language model (LLM) agents, enabling stronger reasoning, planning, and tool-use in interactive environments. However, existing agent RL methods remain fundamentally constrained by an on-policy reasoning bottleneck: when tasks exceed the model's current capability boundary, effective trajectories become extremely sparse and training often collapses. We introduce Rollout-Time Intervention (RTI), a technique designed to elicit successful trajectories beyond the policy's intrinsic exploration capacity. RTI operates by leveraging reward-side evidence (i.e., evaluation rubrics) during rollout to convert erroneous trajectories into successful ones, providing effective learning signals for tasks that the policy rarely solves on its own. Although intervention trajectories are inherently off-policy, we find that they can effectively guide policy optimization when trajectories with excessive distributional shifts are filtered out. Extensive experiments across five benchmarks demonstrate that RTI consistently outperforms RL baselines, enabling the post-trained policy to approach the performance of substantially stronger contemporary LLMs and, more importantly, transcend its original capability boundary. The code is available at https://github.com/Alibaba-NLP/qqr.

Anomaly detection, the identification of collision events that deviate from Standard Model expectations, is one of the most promising model-agnostic strategies for new-physics searches at the Large Hadron Collider. Searching for rare events in learned low-dimensional representations removes the need to commit to hand-crafted observables a priori, but sacrifices the immediate physical interpretation those observables provide. For anomaly detection to enter a discovery pipeline, flagged events must therefore be diagnosed locally, both to distinguish potential new physics from mismodeling or detector effects and to identify observables for targeted follow-up searches. We evaluate supervised and self-supervised neural embedding strategies on COLLIDE-2V, a 750-million-event dataset simulated under realistic High-Luminosity LHC conditions, using the reconstruction error of an autoencoder trained in the learned latent space as the anomaly score. Across the configurations and benchmark signals considered, supervised objectives, which exploit simulation process labels during training, generally outperform their self-supervised counterparts and, in their best configurations, improve on an autoencoder trained directly on kinematic features, raising the signal efficiency at 10% background acceptance on all held-out processes. We further introduce a local diagnostic pipeline that fits a Gaussian Mixture Model to the SM latent space, profiles each component through high-level observables without process labels, localizes the flagged events, and ranks observables by their local discriminating power. The ranking recovers the defining kinematics of the held-out processes, while correlating the anomaly score with the same observables quantifies how closely the score tracks them, revealing cases where the most discriminating observable is not the one the score follows most closely.

Photoplethysmography (PPG) is increasingly recorded by wearable and clinical devices at enormous scale, and a growing number of PPG foundation models (FMs) have been developed to compute reusable representations for various downstream tasks. These models are difficult to compare: different evaluations use different datasets, task definitions, preprocessing, downstream heads, hyperparameter tuning, and baselines. We present PPG-Bench, a unifying open-source benchmark evaluating five public PPG FMs across nine open-access datasets and 26 tasks spanning cardiovascular state estimation, activity, sleep, screening, and identification. We evaluate models under four downstream regimes of increasing adaptability (linear probing, random forests, frozen MLP probes, and end-to-end finetuning), with model-specific hyperparameter optimization and subject-disjoint cross-validation. We also evaluate against three controls: random/constant baseline, a deterministic pulse-rate-variability (PRV) feature set, and a generic time-series FM (MOMENT). Across all 68 task-regime combinations, 35.3% show no meaningful FM advantage over the random/constant baseline, and in 83.8% no PPG-specific FM outperforms both PRV and MOMENT. We also find that added downstream head capacity generally improves performance, but can cause significant degradation if the model is performing at chance/constant level. Current PPG FMs therefore cannot yet claim to provide generalised and transferrable representations of PPG signals, and future evaluation in this field should include checks against simple, non-learned controls. Code for the benchmark and instructions to evaluate more models can be found here - https://github.com/mannas-ai/ppg-bench.

This paper addresses the grid quantization error of deep-learning (DL)-based direction-of-arrival (DOA) estimation. Classification- based DOA methods discretize the continuous angular space into a grid, so that with a peak-based read-out, the accuracy and resolution are limited by the grid step and by the width of the label peak. To address this issue, we propose a sigmoid label that encodes the sources' DOAs as the sum of angle-shifted monotonic sigmoid functions. Differentiating the output of a CNN-Transformer network trained with the sigmoid label produces sharp spectral peaks at the source DOAs. Combined with a moderately refined grid step of 0.25 degrees, the CNN-Transformer with the proposed sigmoid label attains sub-degree accuracy and a 1-degree resolution, breaking the limit of the grid quantization error. Under the unified CNN-Transformer backbone, we compare the proposed Sigmoid label with the Gaussian label, the 0-1 hard label, and the binary cross-entropy (BCE) baseline. Numerical results demonstrate that the Sigmoid label attains the highest accuracy and superior resolution compared to other labels.

Earnings call transcription presents a realistic continual learning scenario where Automatic Speech Recognition (ASR) systems are updated with new recordings but risk catastrophic forgetting of previously learned financial speech patterns. Adapting Whisper-small from a general-domain speech corpus to the Earnings22 benchmark reveals the opposite: all evaluated methods exhibit improvement on held-out financial speech from the same domain during adaptation rather than forgetting. This task complementarity makes regularization counterproductive. Nevertheless, when regularization is warranted, a layer-wise mini-Kronecker-Factored Approximate Curvature (mini-KFAC) approximation seems to achieve lower word error rates (WERs) and more stable optimization than standard diagonal Elastic Weight Consolidation (EWC). The results suggest that task complementarity should be assessed before applying continual learning constraints, and prefer mini-KFAC EWC over diagonal EWC when needed.

Time series causal discovery --- the process of detecting causation between time series --- has its roots in signal processing research from over a half century ago. It has since received active and widespread interest, evolving into different forms across domains such as statistics, dynamical systems and machine learning. Entering this fragmented and constantly evolving research landscape can be daunting. This tutorial provides a unified overview of time series causal discovery across research communities. We introduce the key principles invoked by modern discovery methods, and we discuss their limitations and assumptions. We also discuss how discovery methods are evaluated, benchmarked and fine-tuned. We further discuss how causal discovery for time series is becoming crucial for downstream tasks in agriculture and Earth science, in healthcare, and in artificial intelligence systems. We believe that this tutorial will serve as a valuable resource in navigating the deluge of recent causality research, as well as an invitation to signal processing audiences to contribute to causal problems. To further help readers, we host a companion repository to link to useful code packages and resources at https://github.com/CHAI-UK/Awesome-Time-Series-Causal-Discovery.

Concurrency bugs manifest only under specific thread interleavings, making them hard to detect and reproduce. Existing controlled concurrency testing techniques take over thread scheduling to explore the interleaving space. As the space grows rapidly with program scale, finding interleavings that can trigger bugs within a limited budget remains challenging. The program's concurrency semantics can reveal which variable accesses and synchronization operations under specific interleavings lead to unexpected program behavior, and thus help focus the exploration on bug-relevant interleavings. However, prior techniques have not fully exploited such semantic information. In this paper, we present REINS, an agentic approach for detecting and reproducing concurrency bugs. At its core, REINS guides thread scheduling with the agent's reasoning about the concurrency semantics and known concurrency bug patterns. REINS starts with runtime observations and combines them with static information to reason about the program's concurrency behavior and formulate hypotheses about potential bugs. A hypothesis translates into a set of constraints on event orders and variable states, guiding a controlled scheduler toward interleavings likely to trigger the bug. REINS draws conclusions from the execution results and keeps refining the constraints until the bug is triggered or the hypothesis is refuted. Compared with existing techniques on three public benchmarks, REINS detects more bugs within the same time budget. Furthermore, REINS found six previously unknown concurrency bugs in real-world projects, three of which have been confirmed by the developers.

Large language model (LLM) agents are becoming direct users of the open web. As they take on increasingly complex and information-intensive tasks, effective web access is becoming a core capability for agentic problem solving. Existing approaches have advanced agentic web access through different access models, including search-augmented agents, browser-based agents, and, more recently, filesystem-based agents. Nevertheless, as users expect agents to complete long-horizon, end-to-end tasks, they increasingly need to conduct broad information gathering, perform browser actions, and revisit acquired evidence within a single trajectory. Yet our preliminary analyses reveal that existing access models struggle to meet these demands. They primarily support either information gathering or browser interaction, while combining their capabilities can involve substantial costs or performance trade-offs. To this end, we present Web\$hell, an interface that unifies web access under a single access model. Using a common schema defined by WebContract, WebCompiler compiles live webpages into addressable information and action objects. These objects are maintained in WebCache and handled through a shared set of operations, providing agents with consistent, operable, and persistent web access. Through extensive experiments across web search, web navigation, and deep research benchmarks, we show that Web\$hell achieves strong task performance at lower cost. Our results further indicate the importance of interface design for effective web access, with Web\$hell opening new possibilities for more capable and efficient LLM agents on the web.

Traditional system identification typically involves two steps: initially, training data is processed to estimate a model of the system of interest; then, the model is used to predict outputs for new input queries. In contrast, recent in-context learning approaches learn a meta model that maps directly from training data and query inputs to outputs in a zero-shot fashion, bypassing explicit model estimation. This approach leverages a meta model trained offline, on large sets of synthetically generated data, to emulate the behavior of a system identification algorithm. In this work, we advance this meta-learning framework for system identification in three directions. First, we develop a unified probabilistic meta-identification architecture that combines a Transformer-based encoder-decoder with tailored mechanisms for handling long training sequences, enabling calibrated predictions over long horizons. Second, we show how a pre-trained meta model can act as an uncertainty-aware synthetic data generator for classical system identification, providing task-specific data that regularize parametric models when real data are scarce. Third, we systematically evaluate the framework on standard nonlinear system identification benchmarks and on a challenging out-of-distribution scenario, demonstrating competitive accuracy with state-of-the-art methods and strong generalization, including cases where input-only adaptation is sufficient to handle significant distribution shifts.

Exhaustive optimistic scoring can make action selection costly in linear bandits with a large fixed set of $K$ actions, requiring $\mathcal{O}(K)$ score evaluations per round. We introduce Persistent Candidate LinUCB (\textsc{PC-LinUCB}), which maintains $k$ active actions and samples $q$ challengers uniformly from the remaining $K-k$ actions at each round. By scoring only these $k+q$ candidates, \textsc{PC-LinUCB} reduces the number of score evaluations per round to $\mathcal{O}(k+q)$. We establish problem-independent and problem-dependent high-probability regret bounds consisting of a standard linear-bandit learning term and an additive discovery cost of order $\mathcal{O}\left((K-k)\log(K-k)/q\right)$ at fixed confidence. Notably, this additional cost does not grow with the horizon. For fixed gaps and sufficiently long horizons, we establish an $Ω\left((K-k)/q\right)$ regret lower bound under uniform challenger sampling, showing that partial candidate access itself imposes an unavoidable regret cost. Experiments show that \textsc{PC-LinUCB} substantially accelerates LinUCB while maintaining comparable regret, and that the persistent candidate mechanism can extend beyond LinUCB.

Forecasting desertification has received far less attention than mapping it. We use a convolutional LSTM (ConvLSTM) trained on ten annual Earth Observation (EO) variables (2000-2023) over the semi-arid state of Ceará, Brazil, on a 256x206 grid at roughly 2.2 km resolution, to measure how much of the year-to-year evolution of a MapBiomas-derived degradation index (IDV) can be predicted from satellite data. The network predicts a bounded change on top of the last observed state, is trained in two stages (single-step, then recursive rollout) and is evaluated on a contiguous spatial and temporal holdout. State-wide mean IDV rose from 0.238 to 0.350 between 2000 and 2023, about 0.005 per year, but cell-level change is dominated by year-to-year jumps, and a gradient-boosting model given all channels explains only 4.7% of its variance. Against this ceiling, the ConvLSTM lowers RMSE relative to persistence by 1.6% one year ahead and 4.4% four years ahead in a held-out southern band (p < 0.001), beats persistence with a constant drift and reproduces 53-69% of the regional drift. Drift-aware baselines only help from the third year onward, by up to 4.4% in RMSE at six years. An earlier formulation that redrew the whole field each year predicted changes unrelated to the observed ones (correlation -0.06) and produced a spurious decline when run recursively. Channel knockouts suggest that elevation and EVI degrade the forecast, while slope and aspect have no measurable effect. Monte Carlo dropout and stochastic scenario maps convey the uncertainty that the expected-value forecast hides. Under climatological forcing, mean IDV rises from 0.350 in 2023 to 0.370 in 2030. EO-based desertification forecasting is therefore reliable for the regional trend, not for individual cells.

Linear models can predict well in complex domains even when underlying mechanisms are nonlinear. We argue that this can occur because the complexity of the world exceeds the complexity that can be learned reliably from finite, partially observed data. Flexible models can exploit curvature, thresholds, and interactions that fit one hospital, period, population, policy setting, or measurement process but fail elsewhere. When this instability lies mainly in the nonlinear component, omitting it can give linear models lower prediction error in a new environment. We connect this claim to work on simple linear rules, bias and variance, regularization, best linear prediction, measurement error, distribution shift, invariance, Rashomon sets, and underspecification. Proposition 1 gives a simple squared-error case in which a nonlinear rule learned in the training environment raises target error because it transfers too weakly; finite-sample estimation error can strengthen this result. Partial observation alone is not enough: the link between observed variables and unobserved factors must vary across environments. Stable nonlinear structure favors flexible models, and linear models remain vulnerable to linear spurious relationships. Model complexity should therefore be chosen by cross-environment stability as well as training-data accuracy.

Long-term personalized tutoring requires language models to translate a learner's evolving history into appropriate instructional support. Existing LLM-based tutoring systems use memory and retrieval to personalize responses, but the pedagogical rationale linking historical evidence to teaching strategies is not always explicit. We introduce RegulaTutor, a theory-informed framework for long-term personalized tutoring that connects historical evidence, learner interpretation, and learner-facing instruction. Guided by self-regulated learning (SRL) research, its tutoring harness operationalizes a cognitive, affective, metacognitive, and motivational (CAMM) framework for deriving evidence-supported learner interpretations and teaching policies. The resulting policies retain links to their supporting evidence and are updated as new observations arrive. This staged workflow also supports systematic refinement through PairRoute, an offline procedure that uses intermediate execution records to distinguish harness, model, and mixed failures. PairRoute proposes and evaluates targeted harness revisions, then constructs reviewed stage-level supervision for confirmed model failures remaining under the selected harness. We focus this refinement on historical evidence acquisition, which underpins personalized instruction and allows outputs to be checked against learning records and reference answers. RegulaTutor improves historical-evidence accuracy, learner diagnosis, and judged teaching quality on average across evaluated model configurations. PairRoute results show an endpoint improvement in standalone-model historical-evidence accuracy over the refinement trajectory. Together, these results highlight RegulaTutor's potential to support lifelong personalized education through instructional policies that evolve with learner histories and targeted improvements to the tutoring system.

Tibetan, spoken by some seven million people on and around the Qinghai-Xizang Plateau, remains underrepresented in natural language processing (NLP). We present, to our knowledge, the first peer-reviewed systematic bibliometric survey of Tibetan NLP: 140 publications from 2004 to 2026 in the ACL Anthology and DBLP, annotated for venue type, research area, first-affiliation country, and resource release. Output has grown rapidly, with 84% of the publications appearing since 2018, and has expanded bottom-up along the NLP pipeline, from corpora and word segmentation through script, speech, and applications to large language models (LLMs). Researchers based in China first-authored 87% of the publications, which span all 25 represented research areas, whereas work from other countries concentrates on Classical Tibetan. We argue that this full-pipeline paradigm, realized under the compounded constraints of a complex script, dialectal diversity, and geographic remoteness, offers a reference model for other low-resource languages of the Global South. We also identify gaps in dialect coverage, responsible NLP, native and comparable evaluation, cross-community visibility, and resource release (only 29% of the publications release data, code, or models), and recommend open science, shared standards, and community participation for the LLM era.

We develop a model-free, noise-aware Bayesian clustering framework that constructs a martingale posterior over clusterings from a predictive Bayesian perspective. This avoids explicit likelihood-prior model specification, expensive MCMC sampling, and mode-seeking variational approximations, while accommodating irregular cluster shapes and enabling parallel generation of posterior draws. To interpret the posterior over high-dimensional, noise-aware clusterings, we introduce a Wasserstein-type summarization method. Using a novel noise-aware loss, this approach compresses the large collection of posterior draws into a small set of weighted representatives, capturing distinct cluster structures and noise assignments. Further, we marginalise the uncertainty in the density level, avoiding reliance on a specific choice. We establish theoretical properties of the proposed loss and guarantees for the level-marginal posterior approximation. Results demonstrate the framework's effectiveness in capturing clustering and noise uncertainty.

Coding agents are increasingly used to build programs from scratch as software engineers do rather than to resolve issues within a repository or to generate a repository from a given requirements document. However, scalable verifiable training environments that force agents to discover what is required and to decide how the system should be designed remain scarce. Moreover, existing agent workflows also cannot sustain long-horizon development in which the agent explores the reference program before implementation. To address these challenges, we introduce ProgramGym, a framework consisting of Program-Env and Program-Agent for scalable training environments of long-horizon program reproduction agents. Specifically, Program-Env constructs verifiable training environments from repositories through a feedback-driven loop that repairs runtime sandbox builds and the quality of black-box tests. Using this novel pipeline, we obtain 1,648 training environments spanning C, C++, Go, and Rust. Program-Agent combines a role-isolated workflow of analysis, implementation, and review with revision management for compilation checks and failure recovery. Experiments on ProgramBench show that both Program-Env and Program-Agent improve performance on program reproduction. Fine-tuning on trajectories with these constructed environments raises the performance of Qwen3.8-27B from 34.87% to 42.84%. Program-Agent further raises the base and SFT models to 55.71% and 56.25%, both exceeding some flagship models. We will release the ProgramGym to support further research on coding agents that build programs from scratch.

Geospatial machine learning often depends on task-specific feature engineering over large collections of sociodemographic, business, and environmental variables. Geospatial foundation models offer reusable representations of place, but leading approaches typically encode dynamic signals, such as mobility and search activity, or visual signals from satellite imagery. Few models represent the static structural composition of places (who lives and works there, which businesses operate there, and which risks they face) as transferable embeddings at fine spatial resolution. We introduce PlacePulse, a structural geospatial foundation model that compresses 572 attributes, expanded to 2,288 inputs through H3 neighborhood aggregates, into 256-dimensional unit-norm embeddings for 1,055,537 H3 resolution-7 cells in the United States. We also introduce a reusable two-axis benchmark that separates intrinsic representation quality from downstream predictive utility across multiple targets and spatial supports. Intrinsically, the embeddings reconstruct standardized inputs with mean $R^2 = 0.853$, use 251 of 256 approximately decorrelated dimensions, and show strong spatial organization, with Moran's $I = 0.93$ on the leading principal component and autocorrelation persisting beyond the input radius. Extrinsically, across 74 held-out targets from 12 external sources and four spatial supports, PlacePulse outperforms a dimensionality-matched feature-selection baseline on targets unrelated to the training inputs and is competitive with Google's Population Dynamics Foundation Model, leading on health, economic, election, economic-mobility and social-vulnerability outcomes, with relative strengths that vary by domain. These findings position structural composition as a distinct and complementary modality for geospatial foundation models, not simply a fallback where behavioral or imagery data are unavailable.

Stance detection remains challenging for large language models (LLMs) due to implicit opinions, sarcasm, and context-dependent expressions of stance. Although automatic prompt optimization offers an alternative to fine-tuning without updating model parameters, existing structured approaches are not specifically designed around stance-label semantics and class-specific ambiguities. We introduce MoPrO-SD, a stance-specific prompt optimization framework that decomposes prompts into functional semantic modules and refines them sequentially using LLM-based feedback. We study three refinement strategies: instruction-only optimization (I), instruction with reasoning (I+R), and instruction with per-example reasoning (I+R+E), using class-conditioned errors for stance-label refinement. We also introduce VaccineStance, a new human-annotated benchmark for vaccine stance detection. Experiments across four datasets and five open-weight LLMs show that MoPrO-SD improves upon the human-written Base prompt in many settings. On SemEval-2016, class-conditioned feedback achieves higher mean Macro-F1 than pooled feedback for four of five backbones. Both reasoning-based strategies, I+R and I+R+E, achieve strong performance, with their relative effectiveness varying across models and datasets, while remaining competitive with recent prompt optimization methods. The MoPrO-SD code and VaccineStance dataset are available on GitHub.

Kernel Ridge Regression (KRR) is a theoretically sound algorithm for nonparametric learning, which is also often used to investigate properties of deep learning. However, the exact KRR formulation incurs time and memory costs preventing applicability to large datasets, motivating the development of approximate solvers. These algorithms rely on (i) iterative methods, to avoid the computational burden of direct solutions; (ii) sketching or batching, to reduce the problem size; and (iii) preconditioning, to accelerate convergence. While crucial for fast convergence, preconditioning incurs memory costs that can be managed by lowering the number of free parameters of the model, potentially at the expense of expressivity and hence accuracy. To address this challenge, we propose BLOKS, a Conjugate Gradient-based solver for Nyström-sketched KRR built around a novel preconditioner that combines a block Jacobi approximation with sketching. The BLOKS preconditioner has memory cost independent of both the dataset size and the number of parameters, allowing scaling to large models. We provide a memory-efficient, GPU-accelerated implementation and show that BLOKS achieves excellent accuracy while retaining favorable time/memory complexity. Indeed, on standard large-scale datasets used to benchmark kernel solvers, BLOKS achieves comparable or better performances than state-of-the-art methods.

Rubrics have been widely used to evaluate and optimize language models on open-ended tasks. However, existing work primarily focuses on generating better rubrics, with limited understanding of what makes rubric design effective. In this work, we conduct a systematic mechanistic analysis of rubric design across criterion organization, criterion operationalization, and score aggregation. Our analysis shows that structural organization enables more comprehensive criterion construction, semantic specification alone does not ensure effective criterion-level judgments, and sophisticated score aggregation provides limited additional gains. Guided by these findings, we introduce Adaptive Rubric Tree Search with Hybrid refinement (ARTS-Hybrid), which combines adaptive tree construction with semantic-first, response-calibrated refinement. Experiments across benchmarks show that ARTS-Hybrid improves both end-to-end evaluation performance and intrinsic rubric quality. Further experiments suggest that, beyond evaluation, rubrics generated by ARTS-Hybrid provide effective signals for model optimization.

Mixture-of-Experts (MoE) models have become the standard architecture for frontier LLMs. The multiple parallel experts in each transformer block theoretically enable expert specialization according to categories such as domains, languages, or modalities during pretraining. However, when data is imbalanced across categories, minority categories can become concentrated in a small fraction of the expert pool or spread inefficiently across them, effectively limiting their available capacity. In this work, we demonstrate that category-conditioned expert-selection biases can preferentially route tokens from different categories toward specialized expert pools. By applying biases based on input category, we provide a simple way to induce soft category-aligned specialization. This intervention encourages categories toward distinct expert subsets while retaining cross-category sharing when useful. More importantly, it enables the over-allocation of expert capacity to categories that would otherwise be disadvantaged by data scarcity. We observe performance improvements across modalities, low-resource languages, and code in models up to 10.6B parameters. At our smallest and largest model scales, these gains are equivalent to allocating 18.5% and 2.7% additional baseline training tokens, respectively. Overall, our results show that expert biases can serve as a lightweight control for category-level expert allocation, making room for low-resource categories while retaining standard MoE inference.

Classifier-free guidance (CFG) is a standard technique for improving conditional generation in diffusion models, yet its activation schedule is usually selected heuristically. In practice, most existing CFG-based methods rely on a pre-defined activation policy throughout inference, while recent dynamic variants often depend on hand-designed schedules without a principled theoretical explanation of when conditional information is most useful. In order to understand how the informational relevance between conditioning inputs and the score function evolves over the course of generation, we explore when guidance should be activated from an information-theoretic perspective. In this work, we formulate the diffusion trajectory as a time-dependent information bottleneck and show that the effectiveness of conditional information is not uniform throughout sampling, but becomes significantly stronger after a critical stage characterized by a Fisher-information balance condition. Motivated by this insight, we propose a training-free dynamic CFG strategy for inference. Rather than prescribing a fixed schedule in advance, our method constructs a stochastic process from conditional and unconditional score statistics and activates guidance when the estimated guidance gain crosses a prescribed threshold. This yields an adaptive activation rule that aligns condition injection with the underlying generation dynamics, while also recovering interval-based guidance as a limiting special case. Extensive experiments across diverse architectures and tasks, demonstrate that our method is effective, robust, and broadly applicable. Beyond empirical improvements, our framework provides an interpretation of dynamic guidance methods and a principled foundation for inference-time control in conditional diffusion models.

As the production of software passes into the hands of software itself, coding agents are increasingly commissioned to build agents fitted to their users' own workflows. This delegation carries a quiet assumption, that the delivered agent will faithfully serve the requirement it was built for, rather than merely appear to. No existing benchmark tests that assumption. We introduce AgentSWE, a benchmark for agent software engineering spanning three stages of the agent development lifecycle: Creation of a complete agent from a natural-language requirement, Editing of production agent codebases, and Optimization toward a measurable target. Each of its 25 tasks is framed as a commission rather than an exam: the builder iterates against a handful of illustrative cases, while acceptance exercises the frozen submission on held-out cases designed to span the full requirement, grading behavior against requirement-derived rubrics under hard pass/fail gates or, for Optimization, against the source benchmark's own objective. Evaluating a range of coding agents across models and harnesses, we find that none comes close to delivering what a requirement asks for: the gap between what a builder demonstrates and what it delivers is the norm, not the corner case. An accompanying empirical analysis of the failure trajectories identifies recurring failure modes, among them overfitting to the visible cases, hollow runtimes behind code that looks complete, and unsupported claims of completion, each a signal by which coding agents can learn to build the agents they are asked for.

Coding agents are increasingly deployed for end-to-end software engineering tasks, where a readily available oracle evaluator that defines the success criteria is often absent. This motivates verifications for coding agents that can reliably judge and explain their behavior. However, verification is non-trivial as coding agents are increasingly capable of solving long-horizon software engineering tasks. Their execution trajectories can be extremely long, containing hundreds of iterative interactions with code repositories and external tools. Existing verifiers either rely on general-purpose LLMs without specialized training or train dedicated critics that predominantly focus on producing scalar rewards, highlighting the relatively limited exploration of training-based generalist verification. We introduce SWE-Critix, a post-training pipeline for training generalist verifiers of coding-agent trajectories. Given a task and its execution trajectory, the learned verifier predicts the completion label and generates a natural-language analysis that explains its judgment. Our pipeline combines supervised fine-tuning (SFT) with reinforcement learning (RL) and trains a series of 30B models on trajectories from CodeForge-Preview. On a held-out test set, SWE-Critix improves classification accuracy from 57% for the base Qwen3-30B-A3B reasoning model to 66% after SFT and 78% after RL. We release our models, datasets, and training code to facilitate further research.

The NIST Assembly Task Board #1 remains a challenging benchmark for learning-based methods in robotic manipulation. Over short horizons, it demands precise, contact-rich control of small, tight-fitting parts. Over long horizons, it requires recovery and coordination across extended action sequences. Demonstrations, task decomposition, and dense progress rewards have enabled research advances by providing guidance on how to assemble. But how much of the strategy must be specified? The progression from AlphaGo to AlphaGo Zero showed that RL at scale can discover superhuman strategies without prescribing how to win. AlphaNIST asks whether a sparse binary task reward and task diversity admitted by this task-agnostic reward can similarly let a robot discover how to assemble. In simulation, we train a single policy from scratch across all 20 assembly tasks on the NIST Assembly Task Board #1, reaching 98.2% overall training success without demonstrations, progress shaping, scripted primitives, or per-task tuning. The framework also trains policies to complete ten assembly pairs in a single episode without a prescribed order, with task success rewarded only when all ten pairs are simultaneously complete. The learned policies discover dexterous and planning-like behaviors: choosing grasps for later assembly, using surrounding geometry to reorient parts, and preparing one part, assembling another, then returning to finish the first without an explicit planner. For videos, see our website: https://alpha-nist.github.io/

Multimodal emotion preference learning uses pairwise comparisons to identify which open-ended emotion descriptions align more closely with human judgments. However, existing methods primarily focus on the final preference decision and provide little explicit supervision for localizing and correcting erroneous content. To address this limitation, we propose EmoGavel, a Grounding-Aware reinforcement learning framework with Verifiable rewards for multimodal Emotion preference Learning. We first introduce grounded supervised fine-tuning (Grounded SFT), which exploits controlled-error construction information to jointly learn preference prediction, error localization, repair, and missing-evidence recovery. We then optimize these verifiable outcomes with grounding-aware group relative policy optimization (GA-GRPO), combining preference-gated grounding rewards with dynamic zero-variance rescue to prioritize preference correctness and recover learning signals from all-wrong groups. Experiments on EmoPrefer and controlled-error evaluation sets show that EmoGavel improves both natural preference prediction and controlled-error discrimination while maintaining strong error-grounding performance. Our code and model checkpoints will be released once the paper is accepted.

Widely reused preprocessing pipelines for electronic health records can encode assumptions about when information became available, what clinical episode a label represents, and what missingness means. These assumptions are rarely audited systematically and may propagate across subsequent studies when preprocessing pipelines are reused or adapted. Using MIMIC-IV as a case study, we identify and quantify violations of three complementary forms of preprocessing alignment, i.e., temporal, semantic, and population alignment, which manifest as temporal leakage, feature-task misalignment, and biased imputation. We show how ambiguities in timestamp semantics, episode definitions, endpoint construction, modality selection, and missingness handling can alter the information available to a model and affect the interpretation of its predictions. Through analyses of MIMIC-IV data and empirical case studies, we demonstrate that such choices may remain largely invisible in aggregate predictive performance while changing the signals on which models rely. We derive practical recommendations for temporally valid, clinically aligned, and reproducible preprocessing of MIMIC-IV and related electronic health record datasets. The code for our experiments, including our proposed preprocessing, is available at: https://github.com/m2l-lab/mimic-iv-preprocess.

Large Language Models (LLMs) struggle to answer questions over long interaction histories, where relevant evidence may be scattered across distant sessions. Existing memory systems typically construct and maintain sophisticated memory representations offline, before user queries are observed, incurring substantial overhead and limiting their ability to adapt memory construction to query-specific information needs. In this paper, we propose T$^2$Mem, a Test-Time Memory construction framework that defers memory construction to inference and progressively builds memory along the reasoning process. Specifically, T$^2$Mem extracts lightweight episode-level metadata offline to capture key events and temporal associations as retrieval cues, while retaining links to the original interactions to facilitate the reconstruction of dependencies across episodes. At test time, the LLM adaptively retrieves and organizes metadata and raw interactions, progressively constructing query-specific memory as reasoning unfolds. Experiments on LoCoMo and LongMemEval show the effectiveness of T$^2$Mem by consistently outperforming strong baselines across different backbone LLMs while substantially reducing the overhead of memory construction. Further analysis demonstrates that deferring memory construction to inference enables the memory construction process to dynamically adapt to the information needs of individual queries. All code is available at https://github.com/OpenBMB/T2Mem.

Mamba-3 introduces input-dependent rotations into recurrent state updates, but their contribution beyond state retention remains unclear. We interpret these ro- tations as underdamped second-order dynamics, providing a framework for in- vestigating the role of oscillation in sequence modeling. Controlled comparisons show that underdamped dynamics outperform non-oscillatory damping regimes on state-tracking tasks, while language-model pretraining demonstrates benefits for long-context prediction and length generalization. Interventions in pretrained models further establish the importance of learned phase, and state analysis sug- gests that rotation regulates how retained information combines with incoming up- dates. Taken together, these findings identify underdamped dynamics as a useful mechanism for recurrent computation and suggest that oscillation supports long- context processing by regulating how information accumulates.

Attention sinks are tokens or positions that consistently receive disproportionate attention mass, often beyond what their semantic content would suggest. Despite extensive study in large language models, their behavior and functional relevance in text-to-image (T2I) Transformers remain largely unexplored. To fill this gap and better understand how text-to-image Transformers allocate and route attention, we explore this phenomenon of text-to-image transformers across bidirectional diffusion models and causal autoregressive models. Specifically, we explore whether attention sinks exist, where they emerge, and how they functionally influence generation. Through systematic attention analysis and targeted interventions, we identify that attention sinks emerge in both bidirectional diffusion and autoregressive models but exhibit distinct patterns. The autoregressive model concentrates attention on the first and global-scale tokens, whereas the bidirectional models develop sinks near modality boundaries, within text regions, and at spatial corners. Moreover, these patterns vary across generation stages and are localized to a sparse subset of layers and heads. To assess their functional relevance, we perturb sink-related heads and token positions. It turns out that perturbing sink-ranked heads produces larger average changes than non-sink comparisons, while some sink-position perturbations have limited effects on bidirectional model scores. We further introduce register tokens to redistribute attention sinks and find that several configurations preserve or modestly improve generation performance. Overall, attention sinks are architecture-, inference-stage-, and component-dependent phenomena, providing a basis for mechanistic analysis and principled attention design in text-to-image transformers.

Tool-using agents increasingly require a common substrate for building, training, evaluating, distilling, and maintaining agentic systems, yet existing tooling typically addresses these stages in isolation. We present AgentTune, a source-available Python library that provides a unified framework for developing and optimizing tool-using agents. At its core, AgentTune introduces a two-tier EventLog representation that preserves agent trajectories while supporting consistent projections for training, evaluation, and auditing. Building on this substrate, the library provides a shared agent runtime, tool-calling reinforcement learning, agentic distillation, retrieval-augmented generation, decision workflows, evaluation utilities, and a designed path toward self-healing agents. These components are exposed through composable interfaces so that agent strategies, training methods, evaluation procedures, and recovery mechanisms can be combined without requiring separate representations or pipelines. We demonstrate the framework across agentic tasks including legal contract search and domain-specific applications, illustrating its use for trajectory-based optimization, tool-use evaluation, and cross-task agent development. AgentTune is released as a source-available library at https://github.com/Lexsi-Labs/AgentTune

Motivated by the success of large language models (LLMs) with broad generalizability and robustness to noisy or unreliable pre-training data, we explore in-context prediction of PDE solution fields using approximate pre-training data. In addition, inspired by the posterior predictive mean-based inference mechanism of the in-context learning in prior-data fitted networks (PFNs), we propose PDE-PFN, a prior-data fitted neural solver that approximates the posterior predictive mean of PDE solutions via in-context learning. PDE-PFN builds on a PFN architecture with self- and cross-attention mechanisms of Transformer and is pre-trained on noisy approximate solutions generated by physics-informed neural networks, serving as diverse but not necessarily exact priors. Through experiments on a range of two-dimensional PDEs, we demonstrate that PDE-PFN achieves empirical generalization across heterogeneous equations, robustness under noisy priors, and zero-gradient-update in-context inference capability. PDE-PFN achieves competitive or improved predictive accuracy relative to the evaluated baselines while supporting flexible context and query representations.

Designing tool-integrated reasoning agent systems for complex chemical tasks remains a fundamental challenge. Although recent chemical agents have demonstrated promising performance through the integration of external tools, they are still built on the autoregressive paradigm, which is inherently constrained by causal attention and left-to-right generation. As a result, they often struggle with bidirectional molecular understanding, global coordinated tool planning and multi-source evidence integration. To address this challenge, we introduce ChemDiffAgent, the first masked diffusion language agent for tool-integrated chemical reasoning. It reformulates chemical reasoning as iterative denoising over interleaved reasoning and tool-use trajectories, enabling tool-use decisions to be jointly determined under full-trajectory context and thereby better aligning the agent's decision boundary with its knowledge boundary. We further analyze how this paradigm benefits agentic reasoning through harder infilling subproblems during training and flexible arbitrary-order decoding during inference. Then we develop a two-stage post-training framework, comprising agentic supervised fine-tuning and variance-reduced preference optimization, to enhance tool-use and reasoning capabilities. We also construct a comprehensive benchmark covering six chemistry tasks across both single-turn and multi-turn tool-use scenarios. Under matched budgets and evaluation, ChemDiffAgent consistently outperforms autoregressive agents, achieving an overall improvement of approximately 20%. Notably, our 8B model achieves performance comparable to, in some cases surpassing, significantly larger frontier models, such as GPT 5.5 and Claude Opus 4.7. Code is available at https://github.com/bilemond/ChemDiffAgent.

Serving a Mixture-of-Experts (MoE) model pays for expert precision twice: the resident footprint pays for every expert the router can reach, while each batch-one decode step pays only for the experts it selects. These are two different rate functionals over the same per-expert widths, yet published per-expert MoE allocators constrain only the first. We call the difference the store-read gap; in the released BitsMoE checkpoint, 2.54 bits per weight (bpw) of storage become 7.50 bpw of decode traffic. We propose DuoRate, which stores once and reads per workload. Offline, it stores one uniform tier of a prefix-readable weight stream and bakes a distortion tensor. Online, it profiles a workload's own prefill without labels and solves a Lagrangian rate-distortion allocation under the decode functional; routing frequency provably cancels from every per-expert decision, and the solve costs one dot product per expert and a bisection. A deployment therefore keeps a single artifact that prefill reads at the stored rate and decode reads at the workload's own rate. On Qwen3-30B-A3B, DuoRate reading 2.3 bpw exceeds every 2-bit-class baseline by 6.3 to 47.0 points of average accuracy, and at 2.9 bpw it matches 4-bit formats that read 1.4-1.6 bpw more. The advantage is largest under long decoding: on GSM8K, DuoRate at 2.9 bpw matches or exceeds every 4-bit format and trails BF16 by only 0.25 points, and at 2.6 bpw it still matches or exceeds the 4.5-bit NF4 and FP4. Against controls in its own format, the allocation alone is worth 0.07-0.11 bpw of read rate. Code will be released upon publication.

Agentic workloads resend long contexts that barely change between calls. Prefix caching now skips most of their prefill, yet every request must still be tokenized in full before the cache can even be consulted. Tokenization, a CPU stage long treated as negligible, becomes the largest part of time-to-first-token (TTFT), and a frontend that tokenizes for a whole GPU fleet limits how many GPUs it can feed. We present GraphTok, a GPU tokenizer that runs the entire encoding pipeline, from pre-tokenization through byte-pair encoding (BPE) merging to token-ID assembly, as one replayable CUDA graph. GraphTok reproduces the reference token IDs bit-for-bit on GPT-2, Llama 3, Qwen 2.5, DeepSeek-V3, and Gemma 3, hands short inputs to a CPU engine where the GPU cannot win, and plugs into unmodified vLLM and NVIDIA Dynamo. GraphTok encodes one million GPT-2 tokens in 1.7-3.2 ms including data transfers, and on Llama 3 it is 4.7-6.0x faster than gigatoken, the strongest CPU tokenizer. On TokTier's own single-request benchmark, it is 2.6-3.9x faster than gigatoken and up to 18.1x faster than TokTier. In vLLM with a cached prefix, it lowers mean TTFT by 2.1-2.4x. The largest gain is capacity. Within a 2 s P99 target, one Dynamo frontend with GraphTok sustains 518 requests/s of 121k-token traffic, the busiest-hour load of 56 GPUs, against 296 with gigatoken and 18.5 with the default tokenizer. Across 121k to one million tokens, this is 1.75-2.5x the load gigatoken sustains, with a P99 up to 64x lower where gigatoken's tail breaks.

Web crawling is a major source of large language model (LLM) pretraining data, yet data-selection pipelines discard a large share of crawled pages. We investigate how a crawler should allocate its budget across and within hosts to collect more high-quality pretraining data. Combining document-level quality classifier scores, web-graph centrality metrics, and Common Crawl's per-host crawl coverage, we find that roughly half of the variation in classifier scores occurs between hosts. In the crawls we analyze, a few pages from a host also provide a useful estimate of the quality of its remaining pages. Common Crawl uses harmonic centrality, a graph centrality metric, to guide its crawl allocation. In our simulations, using centrality to choose which hosts to crawl yields limited gains over random ordering and performs worse for some classifiers. For deciding how deeply to crawl each host, sampling a few pages and then fetching more from hosts with higher-scoring samples collects several times more high-quality pages per fetch than uniform depth or centrality-based allocation. Finally, Common Crawl's observed per-host coverage has little correlation with the evaluated quality signals. We estimate that tens of millions of high-quality pages remain among known but uncrawled URLs in each crawl examined, with most concentrated on a small share of hosts.

Partial differential equation (PDE) surrogate modeling often relies on either dense observations or access to governing equations/physics constraints. In practice, observations can be sparse and partial, and the underlying physics may be partially specified or unavailable. We introduce SBNO, an equation-free paired-data diffusion bridge matching framework that learns conditional operators for both forward and inverse PDE problems under partial observations. SBNO uses diffusion Schrödinger bridge matching to learn a pair of time-dependent drifts from a mask-aware endpoint coupling that preserves empirical input-solution pairing during training and supports both PDE-task directions within a single framework. At inference time, SBNO produces a deterministic point reconstruction by integrating the probability-flow ODE constructed from the learned drift pair, without inference-time observation or PDE guidance-weight tuning. On five PDE benchmarks with 5% observations, SBNO achieves superior or comparable accuracy to baselines while demonstrating lower test-instance error dispersion. Furthermore, SBNO is at least $9.7 \times$ faster and reduces GPU memory by at least $4.3 \times$ compared to guided-diffusion baselines.

This paper considers channel state information (CSI) prediction using neural networks for time-varying multi-antenna channels from noisy historical observations. The main challenge is to exploit temporal correlation while distinguishing channel evolution from observation noise. We propose JEPA-kMMVAE, which combines a joint-embedding predictive architecture with a k-memory Markov variational autoencoder to predict future CSI through temporally structured latent representations. The framework supports direct prediction and residual corrections relative to the latest latent state and noisy CSI observation. Simulation results show that the residual configuration outperforms the evaluated baselines at moderate and high input CSI signal-to-noise ratios, while sequence baselines remain more accurate under stronger noise.

The escalating environmental footprint of the Information and Communication Technol- ogy (ICT) sectorwith global Data Center electric- ity usage potentially reaching 3% by 2030 and the imperative for real-time responsiveness in smart buildings demand a foundational shift towards Edge Computing. This article proposes a holis- tic architectural and algorithmic approach for cre- ating autonomous, resilient, and privacy-preserving learning systems within smart buildings, advancing the core concept of the building itself as the data center. Our contributions, demonstrating originality compared to conventional centralized systems, are structured around three main pillars: 1) Validation of resilient data acquisition chains using the af- fordable Arduino ecosystem (Uno R4 WiFi/BME280) and secure persistence solutions including centralized archiving (Node-RED/MySQL) and secure remote transfer (Python/rsync/SSH). 2). The proposal of a lightweight, KOptim orchestration framework, inter- operable with the FIWARE standard, for the dynamic management of Service Level Agreements (SLA). 3) The adaptation and experimental valida- tion of the Generalized Hebbian Algorithm (GHA) for Principal Component Analysis (PCA) on low- power microcontrollers,We introduce and compare an iterative learning strategy using data buffers, featuring sorting and explicit preservation of extreme samples to maintain diversity. Experimental results on the high-dimensional Tour Perret sensor data confirm that this novel method, despite introducing a necessary temporal complexity of O(W log W ) due to sorting, achieves convergence and final result quality that is visually equivalent to processing the entire dataset in full memory, reliably identifying two distinct operational states.

Local online training on resource-constrained platforms requires low computational and training-state overhead. We study the numerical stability and computational efficiency of half-precision proximal policy optimization (PPO) on commodity GPUs. Directly converting PPO training to FP16 is unreliable because numerical errors feed back through policy updates and subsequent data collection, while PPO repeatedly compares a current policy with a fixed behavior policy. We propose PPO-FP16 with three structure-aware modules. Global scalar protection stabilizes statistics shared across a minibatch or parameter update. Policy-change residual protection extracts and preserves small differences between the current and behavior policies. Efficient persistent-update protection stabilizes optimizer state and parameter writeback while reducing state storage. These changes preserve the PPO objective, hyperparameters, and training budget. Across six continuous-control tasks and five seeds, PPO-FP16 achieves the lowest mean return-AUC rank (1.90) among the four methods and ranks first in 17 of 30 environment-seed comparisons. Eight of nine performance configurations yield PPO update speedups ranging from 1.16 to 2.03 times. Relative to Full FP32 Adam, optimizer-state memory is reduced by 71.7 to 74.9 percent, and peak allocated GPU memory is reduced by up to 44.5 percent. The results show that protecting a small set of structurally important quantities can deliver stable, efficient FP16 PPO without retuning.

We introduce MORAD, an online reinforcement learning framework for backbone-conditioned RNA diffusion models that trains a single shared policy across diverse target backbones. Structure-conditioned RNA inverse design seeks nucleotide sequences compatible with a target three-dimensional backbone, yet existing reinforcement learning approaches either optimize a separate policy for each target backbone or rely on offline preference optimization, leaving online post-training of a reusable backbone-conditioned model largely unexplored. MORAD combines backbone-conditioned group sampling, group-relative reward normalization, and a forward-process diffusion objective. Beyond existing RL rewards that primarily target tertiary structural quality, MORAD jointly optimizes six measurements spanning tertiary geometry, base pairing, ensemble behavior, and sequence composition. To aggregate these heterogeneous objectives despite differences in scale and optimization direction, MORAD maps each measurement to a bounded desirability score and combines them with a weighted geometric mean. Compared with a tertiary-structure-only reward, this multi-objective formulation improves 10 of the 11 evaluation metrics that have a preferred direction. Applied to pretrained RIDE, MORAD improves GDT-TS by 16.0% (from 0.3300 to 0.3827) and pairing MCC by 19.4% (from 0.6110 to 0.7295), while reducing normalized ensemble defect by 27.5% (from 0.3906 to 0.2830). On the test set, MORAD outperforms RiboDiffusion, gRNAde, and RDesign across tertiary structural accuracy, secondary-structure pairing, and thermodynamic quality, establishing a new state of the art in structure-conditioned RNA inverse design.

DAG structures can be learned from observations by incorporating acyclicity via smooth constraints, such as DAGMA's log-determinant function. Typically, observations and the underlying DAG are related through linear structural equation models, with Gaussianity assumption on the disturbances at each node. Extending DAG learning methods that account for scale and tail behavior of disturbances that vary across nodes remains a challenge. We propose a node-varying Student-$t$ DAGMA that jointly estimates the graph structure, node-specific disturbance scales, and the degrees-of-freedom parameter. A Gaussian scale-mixture representation yields an iteratively reweighted formulation that down-weights large standardized residuals. Building on existing results, we also establish identifiability of the graph and disturbance parameters. Synthetic experiments demonstrate improved structural recovery under heterogeneous heavy-tailed disturbances, and experiments on real finance data show that the proposed method yields lower test negative log-likelihood than the evaluated baselines.

LLM judges are increasingly deployed as panels, and disagreement among panel members is often read as a symptom of an unclear rubric. We test a label-free procedure that turns this disagreement into rubric repairs. The procedure decomposes each rubric into sub-criteria, ranks them by cross-judge disagreement on a development split, and has an editor model rewrite only the most contested ones. An additional stage that moves text-checkable sub-criteria to deterministic checkers transferred none in our evaluation. Human labels are never used to select or edit clauses; they are used only to evaluate the outcome. We pre-specified four hypotheses in a version-controlled protocol and evaluated six rubric conditions with a panel of five judges from three providers on SummEval, the HealthBench meta-evaluation, and USR Topical-Chat, for a total of 129,770 judge requests, excluding retries. On SummEval, the procedure raised inter-judge Krippendorff's $α$ by 0.069 (95% CI 0.046 to 0.092) over the original rubric and by 0.074 over a control that applied the same edits to an equal number of randomly chosen clauses that were not selected. Agreement with expert ratings showed no detectable change (Kendall $τ$ difference -0.007, 95% CI -0.037 to 0.021). On HealthBench and USR, neither inter-judge agreement nor human agreement met the preregistered effect criteria. In a preregistered secondary comparison, human agreement after editing the contested clauses was lower than after editing the control clauses on all three datasets in point estimate, although no interval excluded zero. Selecting clauses by disagreement makes judges agree with each other more than editing arbitrary clauses does, but in our setting the added consensus did not bring the judges closer to human annotators. We will release code, frozen rubrics, item identifiers, and judge scores.

Modeling human reactions to spatial audio requires understanding both what is heard and where it comes from. Prior work generates motion offline from complete acoustic and spatial input sequences. Humans, however, react to sounds without knowing what they will hear next. We present Sonic-AMG, to our knowledge the first framework for streaming human motion generation from spatial audio. The model encodes acoustic content and body-relative source geometry as separate tokens. Conditioned on these tokens, a causal transformer with a flow head generates continuous motion latents, from which a causal decoder reconstructs motion segments. Standard teacher-forced training derives motion histories and body-relative source locations from ground-truth motion, whereas streaming inference relies on generated motion for both, resulting in a training-inference mismatch. To reduce this history discrepancy, we introduce rehearsal training, which combines full-sequence teacher-forced supervision with supervised prediction along short autoregressive self-rollouts. To keep spatial conditioning consistent with generated motion, we propose spatial feedback, which recomputes body-relative source locations from generated root poses during both rehearsal training and inference. On SAM, Sonic-AMG achieves state-of-the-art FID and R-precision. It further supports real-time online generation of coherent reactions across sound events and source locations. Project page: https://jingchengni.com/sonic-amg/.

Pretrained vision transformers learn structured image representations spanning spatial patch tokens and non-spatial tokens. Recent analyses suggest that non-spatial register tokens encode global scene properties such as illumination, blur, and style. Yet patch-based image decoders discard these outputs, leaving the decoder to recover image appearance from patches alone. We investigate whether retaining register features improves decoding when the complete spatial patch representation is already available. We study this question using representation autoencoders (RAEs) with frozen register-equipped encoders. Our decoder processes four register tokens together with all patch tokens through joint self-attention. Appearance probes, paired reconstruction-error analysis, and fixed-decoder register substitutions examine which global image properties are accessible from registers and how their values affect reconstructed pixels. For generation, we train a multi-head attention pooling (MHAP) predictor to recover register features from sampled patches, keeping the pretrained diffusion model and its samples unchanged. On ImageNet-1K, our ViT-B decoders achieve mean reconstruction FIDs of 0.472 with DINOv2 and 0.442 with DINOv3, reductions of 9.5% and 5.2% relative to their respective patch-only baselines. The corresponding PSNRs are 19.73 and 20.54 dB. For generation, our DINOv2 models achieve a mean FID of 1.641, 9.8% below the patch-only baseline on identical sampled patches.

High-precision retrieval over external knowledge bases is essential for LLM agents to operate effectively. However, existing Information Retrieval (IR) research primarily optimizes query-side retrievers (e.g., ColBERT, BGE) and generators (e.g., Self-RAG, GraphRAG), while reducing corpus-side representation to chunking. This leaves index-side quality an underexplored lever. We propose Pre-Retrieval Index Modeling & Enhancement (PRIME), an index-side optimization framework built on two complementary tracks: an LLM-driven Semantic Restructuring (SR) track that transforms long documents into structured, information-dense sections, and a deterministic Metadata Synthesis (MS) and Hierarchical Retrieval (HR) track that extracts semantic anchors and heading structure for coarse-to-fine section retrieval. Our evaluations show that deterministic metadata matches structure-aware chunking's retrieval quality while cutting embedding computation by 45.8 $\times$ and index vector count by 22.4 $\times$. On top of these gains, Semantic Restructuring (SR) further improves Answer-F1 by 13.7% in full deployment, with characterized quality-fidelity-cost trade-offs. Across QASPER, BRIGHT, and NFCorpus, we further characterize three qualitative regimes that determine whether index-side optimization helps or degrades retrieval, while showing that these index-side effects are neutralized under a strong reranker.

Lagrangian Neural Networks (LNNs) learn mechanical systems from trajectory data by parameterizing a scalar Lagrangian and deriving the dynamics through the Euler-Lagrange equations, biasing the model toward motion that respects the variational structure of mechanics. The interactions studied so far have been pairwise-additive, leaving open whether LNNs can recover an irreducible many-body interaction from motion alone. We address this using the Axilrod-Teller triple-dipole interaction, the leading three-body dispersion term, whose dependence on the collective geometry of three particles cannot be represented by any sum of pairwise terms. We prescribe the kinetic energy exactly and learn only the potential, as a function of interparticle distances and angles. Combined with a signal-aware sampling strategy that concentrates training data where the three-body signal is strongest while preserving broad coverage of configuration space, this recovers the interaction from accelerations alone. The learned potential reproduces the radial scaling, angular dependence, and characteristic sign reversal of the Axilrod-Teller interaction, and generates stable trajectories. Embedding it within a pairwise Lennard-Jones potential lets us examine its observability: the three-body term is recovered once its dynamical signature exceeds the residual modeling error, which it does cleanly at amplified coupling but not at the coupling of a real noble gas, bracketing the transition between the two. Finally, symbolic regression on the trained potential, with no assumed functional form, returns the Axilrod-Teller expression with angular numerator, radial law, and coupling constant intact. An LNN with domain-specific structure can thus recover an irreducible three-body interaction from trajectories as a closed-form law, pointing toward learning interaction laws whose structure is not known a priori.

Prompt optimizers rewrite a prompt to raise one target metric, but prompts also state requirements the metric does not score, such as an output format, citations, or abstention. In a preregistered study, we run GEPA, MIPROv2, and TextGrad on HotpotQA, SQuAD 2.0, and IFEval and count per-item violations of unscored requirements with deterministic checks. By the preregistered criterion, all three optimizers produce regressions, substantially on HotpotQA, where net regression in citation validity reaches 10.7 to 41.6 points while exact match rises by 21 to 34 points. A post-hoc decomposition changes this reading. The optimized prompts elicit reasoning before the required JSON object, and under the 200-token output cap used throughout, 13.9% to 47.0% of responses end before any answer appears. Among parseable outputs, net regression is within 3 points of zero. With a 1,024-token cap, extraction failures disappear, content regression stays below 1 point, and exact match rises further. Two simple gates fail the preregistered criterion. The side effect is narrower than hypothesized: optimized prompts are more accurate and break an explicit output contract, and a lenient target metric does not show the cost.

Distilling long-chain reasoning trajectories from large teacher models provides an effective way to improve the reasoning ability of small language models. However, existing methods typically select a fixed training set before distillation, overlooking that the usefulness of a reasoning example changes as the student evolves. This mismatch becomes especially costly when only a small number of teacher trajectories can be used for training. Besides, examples already mastered provide little new supervision, while examples far beyond the capability of student may be difficult to learn from. Training data should therefore adapt as model capability evolves, yet reliable data selection is difficult when the student is initially weak. To address these challenges, our key idea is to first establish the basic reasoning capability and then adapt the training data to the evolving capability boundary of the student. We propose BoundaryEvo, a two-stage data--model co-evolution framework. In Capability Priming, we directly distill a small set of teacher-generated long reasoning trajectories without using a separate model to judge and filter them, allowing the student to acquire an initial reasoning capability before capability-aware data selection. In Capability-Boundary Co-Evolution, we repeatedly select questions that the current student has not mastered but the expert can reliably solve, and update the training data as the student evolves. With only 1,000 training examples, BoundaryEvo consistently outperforms static and dynamic data selection baselines across 11 reasoning benchmarks, raising the original Qwen3.5-4B accuracy from 62.9 to 75.3 on IMO-AnswerBench and from 26.8 to 46.0 on AMO-Bench. The code and data will be made publicly available.

Vision-language models (VLMs) exhibit complementary capabilities, yet selecting the most suitable expert for a given image-question query remains challenging due to heterogeneous multimodal inputs and varying expert strengths. Existing routing approaches rely on text-based cues or static model profiles, failing to capture fine-grained visual evidence and expert-specific failure patterns. To address this, we propose MHR-Router, a multimodal heterogeneous retrieval-enhanced adaptive routing framework for VLM inference. It accurately dispatches each multimodal query to the most suitable VLM expert via local retrieval evidence. MHR-Router consists of two stages: (1) Multimodal Routing Representation, which fuses frozen multimodal features into a comprehensive query representation; and (2) Heterogeneous Retrieval-Enhanced Adaptive Routing, which retrieves expert-specific success and failure memories as local evidence and employs a shared scorer to adaptively select the optimal expert. Experiments show that MHR-Router effectively achieves heterogeneous modality routing and consistently improves accuracy over traditional baselines.

Memory-augmented assistants retrieve evidence from stored conversations and documents to answer queries. Most embedding-based retrievers compare single-vector representations of queries and memories, even when relevance depends on a small part of the content. Such local evidence can be matched to queries using the contextual token representations already computed by the same embedding models. Our method uses sparse token matching to improve retrieval without additional training. Each query token retrieves a small set of matching tokens across candidate memories. The matching scores are aggregated for each memory and combined with whole-memory similarity to rank memories. An optional low-rank query adapter can be trained with supervision while keeping the encoder and stored memory representations fixed. Our method improves retrieval across eight benchmarks and multiple model families without additional training. Query adaptation further increases the gains to 13.3 and 14.3 percentage points in average Recall@10 and NDCG@10 on Qwen3-Embedding-4B. These retrieval gains translate into higher question-answering accuracy on long-term memory benchmarks.

Set-based language model routing aims to increase the coverage of the event where at least one model answers a query correctly while controlling the number of inference calls made. Existing risk-controlled routing policies bound the all-miss risk (the probability that every model in the selected set fails) by calibrating a single conformal threshold and applying it to a nested family of model sets. Because each model is scored in isolation, these policies ignore the dependence structure of model failures. Models that share training data or architecture tend to fail on the same queries, so two individually strong models can be jointly redundant while a weaker but complementary model covers the queries that the others miss. To account for this dependence, which existing risk-controlled policies ignore, we propose CopRoute, which orders models by how many new queries each one answers correctly beyond those already selected, decides per query how far down that order to go, and then picks the smallest cutoff that still meets the target risk. We show that CopRoute's ordering attains, on the design data, the best guarantee any efficient fixed-ordering algorithm can give for the average number of calls to the first correct model, and that a separate calibration step certifies the all-miss risk of the deployed sets. On RouterBench and RouterEval, CopRoute attains the fewest certified calls on nine of ten datasets at a target risk of 0.05 and on eight of ten at 0.10, making up to 79 percent fewer calls (median 24 percent) than the state-of-the-art risk-controlled policy at the same certified all-miss risk.

API specifications serve as critical documentation for modern software development, directly impacting integration efficiency and client code generation. While recent advancements have leveraged Large Language Models (LLMs) to generate initial specifications from source code, existing approaches treat this process as a one-time event - overlooking the needs for continuous validation, long-term maintenance, and market-driven evolution of these artifacts. To address this issue, we present Agent4Spec, an autonomous multi-agent framework featuring two distinct pipelines: (1) a validation pipeline that detects and self-corrects OpenAPI syntax errors, and (2) a market analysis pipeline that conducts competitive research, identifies functional gaps against industry leaders, and automatically generates missing API-spec functionalities. We evaluate Agent4Spec on two curated datasets of API specifications spanning both general-purpose and industrial scenarios. Among 9 SOTA LLMs, the top-performer achieves 100% syntax error detection rates and over 95% implementation success for proposed corrections, while securing over 50% adoption rates for suggested endpoint improvements, representing a practical tool for API development.

Large language model (LLM) agents are highly vulnerable to memory-poisoning attacks. Existing defenses primarily rely on external modules for either static memory isolation or continuous online auditing, resulting in low memory efficiency and high computational overhead. To address these limitations, inspired by neuroscientific mechanisms that weak reactivation can destabilize memories and induce decay in the absence of reinforcement, we propose an agent-intrinsic training-free memory repair framework, DREAM (Dynamic Reactivation for Engram Attenuation in Memory), that enables selective functional forgetting of poisoned memories. Specifically, DREAM implements a three-stage pipeline: perturbation-induced implicit reactivation to reveal the structural fragility of malicious memories, adaptive anomaly diagnosis based on multi-dimensional activation patterns to detect topological anomalies, and a dynamic memory repair module that selectively suppresses harmful memories while reinforcing benign ones. Extensive experiments across diverse attackers and real-world tasks demonstrate that DREAM reduces attack success rates by over 95\% against backdoor poisoning while preserving strong benign utility. DREAM also achieves competitive robustness against injection attacks. In terms of runtime and token consumption, DREAM achieves up to a 2.13$\times$ speedup and a 37.7\% reduction in token consumption compared with A-MemGuard. Furthermore, DREAM also achieves a task success rate of 92.96\% on poisoned multi-agent systems.

On-policy distillation (OPD) reveals versatility and convenience during VLM post-training, providing dense supervision for student from teacher. However, we observed an anomaly that the distilled VLM could still answer the questions accurately even when the related visual regions were masked or replaced with noise. Building on this observation, our quantitative and qualitative experiments demonstrate that OPD, which uses the answer as a hint, has failed to teach students to genuinely seek out and rely on visual evidence. Our theoretical analysis suggests that this failure arises from a shortcut enabled by the lack of visual constraints during distillation. To address the issue, we introduce Contrastive Vision On-policy Distillation (CV-OPD), which compares teacher distributions induced by positive and negative crops while conditioning on the same student prefix and augments optimization with a margin-based contrastive constraint, which is modulated with discrepancy gating across token and sample level. Specifically, CV-OPD first introduces a negative teacher unrelated to the answer and designs a Contrastive Divergence Loss (CDL) that constrains the feature distribution of student by contrasting the positive and negative teachers, correcting how the student exploits visual regions. The CV-OPD introduces a Discrepancy Aware Gate (DAG) that uses token-level and sample-level discrepancies to select visually relevant tokens for optimization, directing the student's attention toward visual content. Extensive experiments across a broad range of fine-grained visual benchmarks demonstrate that CV-OPD substantially alleviates shortcut and consistently improves the visual reasoning capabilities of existing VLMs.

In this paper, we propose Data-Adaptive Mahalanobis Cross-Head Attention (DAMCHA), a simple and effective attention mechanism that generalizes Multi-Head Attention (MHA) through the lens of metric learning. Standard MHA implicitly learns a static, head-wise Euclidean similarity corresponding to the diagonal blocks of a unified cross-head metric matrix, discarding all off-diagonal interactions across heads. DAMCHA addresses this limitation via two key innovations. First, it learns the full cross-head metric matrix, inducing a Mahalanobis-like similarity that explicitly captures inter-head interactions beyond the restricted block-diagonal structure of MHA. Second, it parameterizes the metric matrix as an input-dependent function, enabling the attention geometry to dynamically adapt to the intrinsic manifold structure of the data rather than remaining fixed after training. We further introduce principled regularization strategies, most notably stack-wise parameter sharing, to ensure computational efficiency and stable optimization. DAMCHA serves as a drop-in replacement for MHA and its variants in various Transformer architectures, delivering superior expressive power and learning performance with comparable or reduced computational overhead. Extensive experiments demonstrate that DAMCHA-based Transformers consistently outperform competing baselines across a range of benchmark tasks. Code is available at https://github.com/erwinmsmith/DAMCHA.

Neural surrogates provide computationally efficient auto-regressive modeling of wave propagation, enabling e.g. fast inverse solutions. Despite recent progress, the current surrogates are not yet sufficient for the needs of ultrasonic sensing, characterized by high frequency and heterogeneous propagation media. We show how methods building on Fourier Neural Operators (FNO) fail to capture essential signal features, such as reflections, whereas U-Net -based methods learn the wave fields well but lack the ability to transfer information across different spatial scales. We then propose a new surrogate that captures the best of the two worlds, by integrating FNOs and local-integral resampling into a U-Net architecture. We show it can be trained on limited data, and that the predictions are robust to input noise and work across spatial scales.

Cross-tokenizer on-policy distillation (OPD) requires aligning teacher and student sequences before their scores can be compared. Decoded-text alignment provides such correspondence at the chunk level, but leaves a distinct problem unresolved: credit alignment. A chunk-level signal specifies how much a span should change, but not how that signal should be distributed across student tokens, while tokens without teacher correspondence remain unconstrained. We propose CAST (Chunk-aware credit Assignment with Selective Token Anchoring), a shift-or-stay framework that addresses both issues. For aligned chunks, CAST redistributes each chunk's advantage across student tokens using predictive uncertainty and local teacher-signal smoothness, emphasizing learning opportunities supported by locally consistent teacher feedback while preserving the chunk's total advantage. The resulting weights exactly solve a KL-regularized credit-allocation problem. For unaligned positions, CAST selectively anchors the student to its initial policy through a KL penalty, limiting drift where teacher supervision is unavailable. Because credit assignment is separated from chunk construction, CAST also applies to punctuation-defined or fixed-length chunks under shared tokenization. Across six mathematics, science, and code benchmarks, CAST improves the average over the strongest evaluated baseline by $2.45$ and $1.32$ points for cross-tokenizer transfer from Qwen3-8B to Mistral-7B and Llama-3.1-8B, respectively, and by $2.10$ points for shared-tokenizer transfer to Qwen3-1.7B.

In-context-learning Tabular Foundation Models (TFMs) are uniquely suited to scenarios with limited resources or privacy constraints, as they solve new tasks without additional training. However, their architecture consists of multiple transformer blocks, resulting in high computational and memory footprints. Meanwhile, traditional alternatives like Gradient-Boosted Decision Trees (GBDTs) remain highly competitive and run efficiently on CPUs, suggesting that tabular tasks may not inherently require deep architectures. This contrast raises a fundamental question: can TFMs be simplified by removing architectural redundancies without sacrificing performance? To answer this, we introduce a closed-form, label-free affine translator that replaces a contiguous window of up to 96% of the blocks, leaving the rest unchanged. Evaluating eight frontier TFMs on 171 classification datasets, we reveal that tabular transformers are vastly overparameterized: replacing half of a model's blocks costs at most $0.002$ median AUROC across all architectures, with the half-depth TabFM still outperforming every other full-depth TFMs and tuned GBDTs. Furthermore, different model families are overparameterized in different ways: TabPFN-3 relies on its encoders, maintaining performance regardless of which single block is retained. Other single-axis models spread computation thinly, needing only their first block, while dual-axis models concentrate it in one critical block. Ultimately, removing depth saves up to 1.5B parameters, speeds up GPU inference by up to $6\times$ at a median cost of $0.03$ AUROC, while a single-block TabPFN-3 serves cached queries $2.4\times$ faster using $6\times$ less memory at negligible cost.