The Inference Report

Reasoning models often generate very long reasoning traces, making inference computationally expensive. Existing approaches typically improve efficiency either through inference-time early-stopping mechanisms or by explicitly encouraging shorter reasoning during training, for example through reinforcement learning with length penalties. We show that substantial efficiency gains can instead emerge from a different kind of supervision: \textit{confidence}. Using a self-supervised procedure, we fine-tune reasoning models to predict their confidence in the answer at intermediate points along their own reasoning trajectories using only 600 training problems. Confidence is used only as a training target: the loss contains no objective for reasoning length, efficiency, or stopping. At inference, the fine-tuned models use the standard generation procedure, with no confidence elicitation or early-stopping mechanism. Despite this, self-supervised confidence fine-tuning makes reasoning more efficient, reducing generated tokens by up to 25\% at matched accuracy across Gemma, Qwen, Nemotron, and GPT-OSS models on mathematical, scientific, and coding reasoning benchmarks, with efficiency gains comparable to methods that explicitly optimize for shorter reasoning. Analysis of reasoning episodes further shows that confidence supervision largely preserves the base models' high-level reasoning composition rather than selectively suppressing particular behaviors. Our results suggest that efficient reasoning may emerge as a downstream consequence of learning metacognitive signals, without being directly optimized.

We give a gap-free differentially private algorithm for the principal component analysis (PCA) problem with Gaussian data.

Recent literature has shown a strong connection between optimization and sampling. We develop the corresponding first-order theory for diffusion models. First, the SDE-based reverse-time flows of overdamped and underdamped Langevin diffusions contract relative Fisher divergences at explicit exponential rates whenever the stationary potential of the forward process is strongly convex---a condition on the noising process one chooses, not on the data. This is a unique advantage of SDE-based reverse diffusion, absent in the reverse process based on ODEs. Second, we incorporate discretization and establish averaged first-order stationarity bounds---the sampling analog of averaged gradient-norm guarantees in nonconvex optimization---for samplers of both overdamped and underdamped diffusion models. As in nonconvex optimization, the convexity-free certificate is local: it guarantees score consistency, not global mode weights.

Generative AI systems are increasingly used, but aligning their outputs with user requirements poses a continuing challenge. Here, we aim to ensure that the distribution of an attribute of an AI-generated output aligns with a user-specified target. This is motivated by examples such as fairness, where we want to ensure that a protected attribute (e.g., gender, race, or age categories) follows a desired distribution, and synthetic data generation, where we want the generated data to be representative of a target distribution. We study the practically important black-box access setting, where a user can repeatedly query a generative AI model. The goal is to return $m\ge 1$ outputs whose joint attribute distribution is as close as possible to this target. For both exact and approximate alignment, we develop algorithms that minimize the expected number of queries to the generator, and we further demonstrate their optimality as the number of requested outputs $m \rightarrow \infty$. Experiments on text-to-image generation and geocoded persona generation tasks show that our post-processing algorithms improve statistical attribute alignment, complementing prompting-based interventions.

Large language models (LLMs) implicitly infer attributes of their users and adapt their behavior accordingly, yet these beliefs remain difficult to inspect and causally manipulate. We introduce Belief Self-Distillation (BSD), a unified read-write framework that bridges linear and causal probing by learning a compact user representation that can be both decoded and written back into the model. The frozen LLM acts as its own teacher, distilling beliefs from natural conversations without external annotations. Unlike conventional probing, BSD isolates not only information present in activations, but a state whose causal role can be directly tested. Across multiple model families, BSD faithfully recovers user beliefs and enables substantially stronger interventions than matched hidden-state steering. Crucially, we find that refusal depends not only on the request, but on the model's inferred user intent: changing this belief alters refusal while holding the request fixed. We further uncover a striking cross-model regularity: independently trained LLMs converge on a shared geometry for representing their users. Together, these results reveal implicit user models as readable and causally writable internal states with direct implications for AI safety, shaping how models condition safety decisions on whom they believe they are interacting with.

Low-rank adapters (LoRA) make it cheap to fine-tune a large language model once per task, but combining several independently trained adapters into one model remains difficult: merging the updates in weight space causes interference, retraining on all task data is expensive, and routing between separate adapters gives up the goal of a single combined model. We trace the difficulty to two choices that every composition method makes implicitly. A LoRA update admits infinitely many equivalent factorizations; the choice among them is invisible while an adapter serves alone, but it determines what a learned interaction between adapters can see. A coupling between an old skill and a new one can likewise point in either direction, and the direction decides whether the old skills keep computing what they computed before. We introduce READ (Read-only Expansion of Adapter Deltas), which fixes both choices: each adapter is rewritten into a balanced canonical form that preserves its update exactly, and the coupling grows in one direction only, so a new skill can read the input subspaces of old skills but cannot write into their output subspaces. The only trainable object at each append is the new skill's row of the coupling matrix, and the composed update folds into the base weights with no inference cost, routing, or task-specific rules. We evaluate READ across four benchmark suites and two model families, adding skills one at a time. Across several families, READ improves every suite average over the strongest published baselines built from the same adapters---by more than twenty points on SuperGLUE and more than seven points on the domain suite---and nearly all complete addition sequences end above every direct baseline. Factor coordinates and coupling direction, which a lone adapter never exposes, are what decide whether composed skills survive.

The Manufacturer Usage Description (MUD) standard allows IoT manufacturers to define expected network behaviors in a MUD file. This file can be translated into enforceable access-control policies, restricting compromised devices to operate solely through manufacturer-defined communication patterns. However, practical adoption of MUD depends on profiles that are accurate, complete, and maintainable. Existing approaches use traffic-based automation but require device deployment and prolonged monitoring, capturing only behavior exercised during observation. Rare, failure-triggered, or configuration-dependent communications may remain absent, producing incomplete policies that disrupt legitimate operation and offer limited insight into the software components responsible for each rule. We present AutoMUD, a source-code-driven tool that generates traceable MUD profiles for IoT devices from their firmware and software source code. AutoMUD combines static and syntactic extraction, retrieval-grounded language-model reasoning, and deterministic validation and compilation to recover the communication behavior characterizing an IoT device and translate eligible endpoints into policy rules. By analyzing code-level evidence, AutoMUD exposes rarely exercised and conditional communication paths, links every generated rule to its source-level provenance, and preserves excluded findings with explicit reasons for review. Our evaluation on a Linux-based repository demonstrates that AutoMUD recovers complete communication behavior, consolidates validated behavior into semantic endpoint groups, and generates structurally valid MUD profiles. Through a controlled semantic fault-injection campaign, we demonstrate that AutoMUD enables analysts to detect, localize, explain, and correct propagated errors, recovering policies semantically identical to their clean counterparts.

Direct feedback alignment (DFA) trains hidden layers through fixed random projections of output error. With tanh hidden units and independent sigmoid outputs, plain stochastic gradient descent can stall near the loss of a constant predictor of class frequencies. We trace this stall to the error's common mode, the component shared across inputs. An exact mean-covariance decomposition separates a rank-one update formed by the mean teaching signal and mean presynaptic activity. Its leading component drives tanh units toward saturation. At initialization, random feedback provides no systematic correction of the shared error on average; readout learning limits its duration. A reduced model initialized from the network, without fitted parameters, predicts the concentration of activation sensitivity across 48 settings. On MNIST, class decodability largely survives collapse, but readout learning remains slow at a fixed learning rate. Adam learns faster despite deeper collapse. Calibrating the baseline readout to the class prior suppresses collapse and speeds learning; weaker feedback trades less collapse for slower learning. Replacing errors by their signs sustains collapse; subtracting the signal's batch mean prevents sustained collapse and improves learning in the tested setting. Related effects occur in deeper and convolutional networks and on CIFAR-10, with severity and cost depending on the readout, optimizer and input statistics.

We investigate whether natural-language documentation helps coding agents resolve software issues, and we build the tools to construct and evaluate it. We introduce a roundtrip benchmark that scores code descriptions by whether code regenerated from them passes the original tests, and show that completeness, not length, drives a description's fidelity. Using the benchmark as an optimization signal, we discover a description-writing prompt that reaches full fidelity and generalizes to unseen files. We then test the hypothesis that motivated the work: that better documentation helps an agent resolve real repository issues. Across two model families and ten repositories, and against a positive control confirming that our evaluation can detect a genuine improvement, we find that it does not. When the source is present, neither static compact documentation nor retrieved context beats the issue alone. We report this negative result together with the benchmark and the optimizer, and we characterize the boundary at which documentation helps.

Decision Transformer performance degrades on long rollouts because the conditioning context drifts out of the training distribution. We show that this drift is visible through the model's own next state prediction error, which rises during rollout and stays elevated, giving a direct signal of when context has become unreliable. We introduce Trust Guided Decision Transformer (TGDT), which selects context before applying value guidance. At each step, TGDT evaluates several recent context suffixes using rolling next state prediction error, calibrated against held out offline data via split conformal prediction. It keeps only suffixes whose error stays within the calibrated threshold, then uses a frozen critic to choose the highest value action among the trusted suffixes. This reverses the order used by value only elastic selection, where the critic may choose an action generated from a context the model itself has flagged as unreliable. Experiments on D4RL navigation and locomotion tasks show that state prediction, critic guidance, and hard context reset each solve only part of the problem. TGDT reduces persistent high error runs and improves return over vanilla Decision Transformer, reset based context control, and value only context selection.

Uneven rotation and dropped frames make equal-angle assumptions unreliable for turntable reconstruction. We present OC-GS, an object-centric Gaussian splatting that refines each image's angle while maintaining a shared camera, rotation axis, and pivot. This orbit-consistent refinement jointly optimizes image-derived geometry and angles to reconstruct objects from sparse, irregular captures. On rendered objects with 12, 8, and 6 irregularly spaced views, OC-GS achieves mean foreground PSNR scores of 21.26, 19.36, and 15.83dB, respectively, exceeding all four evaluated pose-free Gaussian splatting baselines in each condition. Under a shared trainer, refining image-estimated angles improves mean foreground PSNR by 7.88dB over keeping those estimates fixed. An ablation study shows that both image-derived angle initialization and the shared motion model contribute to the improvement. On real captures, OC-GS's refinement increases mean foreground PSNR by 0.70dB. Results show that refining uncertain angles within a shared motion model improves reconstruction from sparse, irregular turntable captures.

Many LLM training and inference methods, including RL and test-time scaling, depend on repeated sampling, but benefit only when the responses meaningfully differ. Self-training faces the same challenge: training data is typically constructed by sampling IID responses and filtering primarily for correctness, thereby overrepresenting strategies a model already favours. We investigate strategic diversity, or substantive variation among approaches to a problem, as an alternative principle for constructing self-training data. We generate strategically diverse data with two sampling methods: GROOT, a new method which constructs a hierarchical tree of approaches and samples distinct paths, and Verbalized Sampling (VS), adapted to produce an unstructured set of approaches. Across competitive programming and Next-Chapter Prediction domains, models trained on strategically sampled data outperform IID-trained counterparts on difficult tasks and provide strong initializations for RL and test-time scaling. Most strikingly, self-training on strategically diverse but incorrect traces from Qwen3-4B outperforms IID distillation from a 235B teacher. These results challenge prevailing assumptions about what makes useful self-training data and show that diversity of approaches can matter more than correctness or teacher scale.

Deep learning has substantially accelerated the calibration of complex stochastic-volatility models, but neural point calibration alone does not capture the uncertainty remaining after an implied-volatility (IV) surface has been observed. We develop a simulation-based inference framework for rough Heston (rHeston) calibration that learns the posterior distribution of the model parameters conditional on an IV surface. Using neural ratio estimation, we obtain calibrated posterior samples that can be propagated through heteroscedastic neural surrogate pricers for path-dependent exotic options. The resulting posterior-predictive distributions combine residual parameter uncertainty with conditional surrogate uncertainty and yield uncertainty-aware price intervals. We further introduce Hellinger-SHAP, an information-theoretic explainability method for posterior inference. Rather than attributing a single parameter point estimate, it applies local-background Kernel SHAP to a posterior-information functional measuring contraction from the prior to the posterior. This identifies maturity--moneyness regions associated with posterior information gain for individual rHeston parameters. In a simulation study, posterior-predictive intervals provide calibrated or conservative coverage across forward-start, barrier, and realized-variance claims, while point plug-in prices can be materially unreliable for selected contract regimes. Together, the UQ and XAI analyses provide a transparent framework for uncertainty-aware neural calibration and downstream exotic pricing under the specified prior-predictive model.

Conversational AI tools are entering children's everyday experiences, and schools are interested in adopting them. However, successful classroom integration depends not only on the technology but also on the work teachers do to make it usable and appropriate for their students and classroom context. There is little known about how elementary teachers work as they implement conversational AI tools in real classrooms. In this study, we examine three teachers' experiences implementing an AI literacy and English Language Arts (ELA) curriculum built around ToyTalk, a conversational AI toy development platform, over 13 instructional days, a three-week summer camp. Drawing on daily individual reflections, group reflections, and post-camp interviews, we find that teachers' adaptive practices of repair, differentiation, translation, and balancing sit at the intersection of three tensions (technology, learner, and instruction). Teachers' understanding of AI and their role evolved over the camp experiences. From these findings, we contribute design implications and considerations for deploying conversational AI within elementary classrooms.

AI tutoring could markedly improve learning outcomes for students in developing regions such as Vietnam, yet the two obvious paths both fall short. Cloud assistants such as ChatGPT route sensitive student data to foreign servers---violating data-sovereignty laws such as Vietnam's Decree 53---and, pre-trained on Western-centric corpora, are not organized around the national textbook curriculum, so their knowledge of local content is unsystematic and frequently hallucinated. Self-hosting an open model keeps data on-premise but hits a two-fold wall: post-training quantization (AWQ, GPTQ) tames the static weight footprint, yet the dynamic KV cache and prefill latency of long tutoring contexts still cause out-of-memory failures and slow responses on consumer GPUs, while the model keeps hallucinating on region-specific material. We present DeepEdu-v1, an AI-tutoring system for Vietnamese education built on SCALE (Self-improving Context-Aware Learning Engine), a framework with two innovations. First, a long-context inference engine amortizes token selection from per-sub-chunk to per-cluster granularity; on long-context retrieval it issues x7.7 fewer retrieval calls than a state-of-the-art selective-attention baseline, cutting prefill latency (TTFT) by roughly 35% while matching or improving task accuracy. Second, a self-improving agentic layer continuously curates a verified playbook from past interactions instead of fine-tuning, a design intended to progressively reduce reliance on dominant-language priors as trustworthy local knowledge accumulates. In its deployed configuration, DeepEdu achieves a nearly x2 TTFT speedup over standard vLLM serving and lifts agentic accuracy from 70.0% to 79.5% on complex tasks, with the strongest per-track gains across financial-reasoning and interactive-agent benchmarks.

We show that neural network weights can be explicilty fintuned to admit a smaller grammar. Weight Pair Encoding (WeightPE) does so by placing a lossy Re-Pair compressor inside a straight-through estimator. The int8 weights of the network are flattened into one string, and near-matching Re-Pair patterns are made exactly equal within a global L2 budget. The network computes with the rewritten weights and trains through them with a straight-through estimator. Unlike a flat codebook of fixed-size entries, a grammar offers variable-length patterns and reuses them hierarchically inside larger ones. On the MLP weights of ViT-B/16 and ViT-L/16 finetuned on CIFAR-10, WeightPE produces a Re-Pair grammar 0.43x and 0.38x the size of the one produced by an equivalent int8 QAT run, at a cost of 1.9 and 1.1 accuracy points. The trend extends to different grammar compressors (LZ78, SEQUITUR), over which the networks has not be finetuned against. To our knowledge, this is the first time grammar size has been used as an explicit training objective for network weights.

Multi-agent LLM systems are often expected to improve as team size increases, yet the scaling behavior may depend on task structure. Our central contribution is to introduce Steiner's taxonomy of group tasks as a framework for analyzing multi-agent LLM scaling and focusing the analysis on disjunctive and compensatory tasks. We model independently sampled agents as conditionally independent given the item, which yields their large-team limits: plurality voting converges to the model's modal answer, and averaging converges to the model's item-level bias. Across selected representative benchmarks, 13 open-weight models, and teams of up to 30 agents, we find qualitatively different scaling behavior. On disjunctive tasks, the probability that at least one agent is correct grows by 5-20 points with team size, but plurality voting over agents that answer directly realises almost none of this potential, as the model predicts to within 0.5 points on average. Multi-round revision raises accuracy considerably, yet the gain is nearly the same with one peer as with 29. In contrast, scaling provides little benefit on Fermi estimation, despite its natural suitability for aggregation: item-level biases shared across the samples of a model account for about 87% of the squared error, so averaging reduces error by only about 6%. Combining model families helps on Fermi estimation but does not surpass the strongest member on disjunctive tasks. These results show that task structure, together with the mechanism combining member outputs, is a fundamental determinant of team scaling.

Large neural networks can be compressed by rounding or "quantizing" their weights to numbers that admit representations with fewer bits. One algorithm for quantization, OPTQ, progressively quantizes the weights of a neural network so that the squared quantization error on a specified calibration dataset is as small as possible. We study the performance of OPTQ and a variant algorithm, stochastic OPTQ, in a generalization setting and derive bounds for the expected squared error accrued by the algorithm when a test point is drawn from a fixed distribution. We prove two results. One result relates the generalization error to the error on a calibration dataset comprising independent samples from the same distribution as the test distribution. The other result bounds the generalization error of stochastic OPTQ for all sufficiently nice distributions, regardless of the calibration dataset. In both of these results, the regularization term $λ$ plays an important role. We use insights from these results to make a new recommendation for the choice of $λ$ and see that this choice of $λ$ preforms favorably in experiments when compared to prior recommendations in the literature.

Online learning in non-stationary environments requires models to adapt rapidly from streaming data under strict computational constraints. A principled approach casts online learning as Bayesian state tracking, where model parameters are updated sequentially via Bayesian filtering. However, applying Bayesian filters directly to modern deep models is computationally prohibitive due to the high dimensionality of parameter space, forcing existing methods to rely on restrictive approximations or manually designed low-dimensional subspaces. In this work, we identify the absence of a suitable low-dimensional dynamical representation as the core bottleneck in Bayesian filtering-based online learning. Accordingly, we propose Adaptive Update through Representation Adaptation (AURA), a meta-learning framework that learns offline a low-dimensional latent state-space model governing the evolution of optimal model parameters under distribution shift. Online adaptation is then performed via extended Kalman filtering in this learned latent space followed by reconstruction of the full model parameters through a learned lifting map, enabling efficient single-step online adaptation while preserving model expressiveness. Evaluated on online adaptation of neural wireless receivers under time-varying channels and on non-stationary image classification, AURA shows substantial improvements in adaptation speed, accuracy, and computational efficiency over existing online learning and Bayesian filtering baselines, demonstrating that an adaptation-aware latent geometry is beneficial for effective Bayesian online learning in high-dimensional models.

Ensuring online safety through content monitoring had raised Hate Speech Detection as a crucial task to be addressed. By essence the task demands the capture of contextual cues, which are essential for a precise understanding of the content's intent. Although automated detection approaches for the task have advanced significantly, the scarcity of non-English resources persists, limiting the ability of models to adapt to the subtle, context-dependent, and culturally related nature of multimodal content. In this paper, we introduce MexHat, a video dataset designed to capture the linguistic and cultural cues for the hate-speech detection task in a Mexican Spanish context. Our dataset comprises around 1k video clips annotated across two tasks: a three-way class evaluation (no negative content, offensive content and hate-speech content), and a fine-grained class evaluation including three hate-speech sub-categories. The dataset statistics and the baseline results highlight the inherent challenges associated with the task. Disclaimer: This paper contains sensitive content that may be disturbing to some readers.

Disaggregating the two stages, Prefill and Decode, onto separate GPU pools is now a standard optimization for (text-only) LLM serving. However, multimodal LLMs (MLLMs), which add a third phase, Encode, pose new challenges for resource allocation. Encode turns images, video, or audio into embeddings that the language model can consume, yielding a three-stage Encode-Prefill-Decode (EPD) pipeline. Existing frameworks offer only partial answers: text-only PD systems lack Encode, while EPD frameworks expose it as a separate service without regulating downstream request flow. The pipeline also carries a structural resource imbalance: every request enters through Encode before downstream work can begin, yet per-request execution leaves the encode GPU severely underutilized even at high loads, starving the downstream Prefill and Decode workers. Addressing this, we reposition Encode as the control point of the EPD pipeline, exposing three tightly coupled dimensions: when work enters downstream, where prefill executes, and how the GPU is shared. We instantiate this in EAServe across two co-designed layers. Its runtime manages load-adaptive micro-batching, rate-controlled partial offload to a co-resident prefill worker, and dynamic SM partitioning for predictable co-location. The configuration layer, Hybrid Auto Selection (HAS), navigates the joint space of GPU allocation, encode batch size, and offload ratio by pruning unbalanced allocations with per-stage capacity profiling and refining the remainder through TPE-based Bayesian optimization. Evaluated on three MLLM architectures spanning image, video, and audio, EAServe delivers up to 4.3x and 1.7x higher goodput than NVIDIA Dynamo and vLLM, respectively, under identical SLO constraints, sustains more balanced and higher GPU utilization across the EPD pipeline, and reaches near-optimal configurations faster than baseline search methods.

We study false-alert control when screening for text generated by artificial intelligence (AI). The screening procedure selects document prefixes and detectors from observed evidence and may stop before exhausting its inspection budget. We give two finite-sample constructions under document-level exchangeability between human calibration documents and a new null document, with no restriction on dependence among tokens within a document. Construction A registers a finite family of prefix-detector scores and allocates a false-alert budget across their conformal ranks. A union bound protects any executed subset of that family. Construction B calibrates the complete-path maximum of a development-fixed adaptive policy. Each partial-path maximum is bounded by the complete maximum, so a terminal conformal rank protects early stopping without splitting the error budget. We prove marginal control of any false alert across the permitted inspection path and derive necessary calibration counts for rejection. We also state oracle testing, distribution-shift, and independent-audit bounds with their additional assumptions. Both constructions protect stopping within their specified scope; neither proof constructs an e-process or justifies multiplying conformal ranks. Detection power and computational savings remain questions for empirical evaluation.

Electrocardiogram (ECG) foundation models typically tokenize the signal into fixed-length patches that ignore cardiac structure, so a patch may split a heartbeat and the number of beats in each patch shifts with heart rate. This matters most for infants, whose heart rates are higher and whose ECG differs from the adult, clinic-recorded 12-lead data these models are built on. A model for infant ECG should therefore reason about heartbeats directly rather than recover them from arbitrary patches. We propose BeatGraph, which makes the heartbeat its unit of representation, modeling each 30-second window as a graph of beats. A shared beat encoder embeds each heartbeat from its waveform and inter-beat intervals, a Transformer with positional encoding orders the beats in time, and residual graph attention layers relate every beat to every other before attention pooling yields a window embedding. We pretrain BeatGraph on our new corpus of unlabeled infant recordings by predicting masked-beat embeddings, then fine-tune it for each task. One backbone supports sleep-wake detection, infant-state classification, activity-source identification (infant- or caregiver-initiated movement), and affect recognition, improving macro-F1 over the strongest baseline on each task by 0.076 to 0.158. It also transfers across age groups, reaching 0.892 AUROC on the ZZU-pECG pediatric benchmark (ages 0 to 14), within 0.001 of the best published self-supervised ECG model, and matching that model under linear evaluation on the adult PTB-XL benchmark despite infant-only pretraining. Finally, to our knowledge, we release the first public infant ECG corpus collected in homes, classrooms, and laboratory settings with state and affect labels. It contains 3,408 hours of single-channel ECG from 143 infants aged 3 to 11 months, with unlabeled pretraining data, benchmark tasks, and subject-level splits.

Neural representational dissimilarity quantifies differences between neural response distributions, and is essential for comparing neural codes across stimuli, brain areas, tasks, and models. Commonly used distance metrics involve different assumptions and are estimated with separate methods. Here, we show that a variety of distance metrics can be unified under a flow matching framework developed in deep generative models. That is, these distances arise as Jeffreys divergences under different velocity constraints. We find that flow matching has advantages for estimating distances involving complicated distributions and continuous variables. Furthermore, this framework enables the design of new distance metrics in a principled way. Together, flow matching provides a unified approach for understanding, estimating, and designing neural representational dissimilarity metrics.

Sending every network-automation input to a third-party frontier LLM exports sensitive artifacts such as production configurations, topologies, and logs. Querying small language models (SLMs) locally avoids this egress, but SLM outputs can be error-prone for direct use. This work introduces checkability as a criterion for determining which tasks are suitable for local inference. A task is checkable when it exposes a cheap, deterministic test - an intrinsic check - that rejects outputs violating a necessary correctness condition. We instantiate this idea in Touchstone, a local-first pipeline that uses seven off-the-shelf SLMs (1-8B parameters) to generate candidates, uses task-specific intrinsic checks to reject responses, and escalates unresolved inputs to a frontier LLM. On conflict detection and intent translation tasks, Touchstone reaches 98.6% and 93.8% end-to-end accuracy while escalating only 16% and 17% of inputs, respectively. On TeleQnA, a knowledge-only control that has no task-specific intrinsic checks, Touchstone is unable to match the accuracy of the frontier baseline. Our results support a simple deployment rule: keep inference local when task semantics support precise, low-cost checks; escalate the rest.

Physics-Informed Neural Networks (PINNs) build neural representations of time-dependent PDE solutions, naturally incorporating physics knowledge and observational data, which makes them well suited to both forward and inverse PDE problems. PINNs, however, are known to suffer from spectral bias and lack of causality. Neuro-Spectral Architectures (NeuSA), a recently proposed alternative to PINNs, mitigate both issues, but their numerical integration becomes unstable for stiff differential equations arising in many relevant physical problems. This study proposes Neuro-Spectral Exponential Time Differencing Architectures (NEXT), which combines the spectral representation of the PDE solution in NeuSA with high-order exponential integrators. Within this approach, the linear stiff part of the vector field induced by the PDE is integrated exactly through matrix exponentials, while the possibly nonlinear remainder is modeled by a neural network. The effectiveness of NEXT is verified through benchmark experiments on a set of stiff PDEs, in which NEXT is stable and accurate while NeuSA diverges numerically. It is also shown that NEXT can be applied to inverse problems, where the model has to learn unknown parameters or boundary conditions from sparse data. All code used in this work is publicly available at: https://github.com/marcioh2m/next.git .

Cooperative multi-agent reinforcement learning under partial observability and shared rewards requires assigning team outcomes to individual agents and high-order coalitions. A MAPPO-style critic compresses joint behavior into one global value, while critics that dynamically reconstruct the grouping topology change the mapping from agents and coalitions to value components as interactions or active agents evolve. We refer to this inconsistency as structural target drift. We introduce HySTAR, a MAPPO-based framework that separates adaptive representation learning from a temporally consistent high-order value-decomposition basis. HySTAR anchors an overlapping sparse hypergraph as a uniformly covered decomposition scaffold, uses a spatiotemporal encoder to represent physical and task-dependent interactions, and combines temporal and structural relevance to construct agent-specific advantages. Experiments on SMAC, GRF, Traffic Junction, and MPE demonstrate consistent improvements over MAPPO-style, value-factorization, and dynamic-grouping baselines. On the hardest SMAC settings, HySTAR achieves relative gains of 16.7\% over MAPPO and 15.6\% over HYGMA, ranks first on all six GRF scenarios, reduces Traffic Junction convergence epochs by up to 40.2\% relative to MAGIC, and obtains the highest MPE episode rewards. Controlled topology, agent-death, neighborhood, and parameter analyses support the benefit of anchoring the decomposition scaffold while adapting the propagated representations.

Sintering is widely used to manufacture ceramics, but coupled densification and grain growth, material-dependent kinetics, and sparse measurements complicate predictive modeling and process design. We present Sinter-PiNDiff, a retrainable physics-integrated neural differentiable framework for predicting density and grain-size evolution. Two neural networks learn densification and grain-growth coefficients within coupled rate equations, while a smooth saturation factor attenuates densification near theoretical density. The same governing structure, network architecture, and training procedure were fitted independently to published data for MgO, Al-doped ZnO, and CaO-doped ThO2. Tests at held-out temperatures and compositions yielded the lowest mean error in all twelve material-metric comparisons against multilayer perceptron and residual network baselines. For MgO, Al-doped ZnO, and CaO-doped ThO2, respectively, density normalized root-mean-square errors were 14.6%, 10.8%, and 14.4%, and grain-size errors using the same metric were 8.6%, 12.1%, and 19.3%. Removing evolving density from both neural-network inputs increased density and grain-size trajectory errors in all three systems and ten of twelve aggregate errors, supporting density-dependent kinetic feedback. Deep ensembles estimated model disagreement, but empirical coverage showed that the uncertainty bands were not calibrated and did not capture all model-data discrepancies. These results establish Sinter-PiNDiff as a retrainable framework for sparse-data prediction and uncertainty-informed selection of sintering conditions.

We develop a statistically explicit sentiment index for Google Play user reviews and establish the mathematical results supporting its construction. Normalized star ratings and text-sentiment scores are treated as noisy measures of latent review valence and fused by covariance-aware inverse-variance weighting. Review-level estimates are aggregated with bounded helpfulness and recency weights, then shrunk toward a population mean using estimated precision rather than an arbitrary review-count threshold. App-level rating histograms provide a distributional diagnostic for samples returned under different API sort orders; because star ratings are discrete, classical continuous Kolmogorov-Smirnov critical values are not used. A local-level state-space model and the Kalman filter provide a denoised temporal trend. Full proofs cover the BLUE and Gaussian maximum-likelihood result, Gaussian-conjugate shrinkage, the Glivenko-Cantelli and Donsker theorems, count transformations via the delta method, and exact Gaussian Kalman filtering. A worked three-review example shows how textual complaints can materially reduce an apparently perfect star-only score.

We present Muslim, a production Arabic voice AI platform serving grounded, sourced Islamic knowledge to real users. Beyond a real-time voice pipeline (NeMo Arabic ASR, an OpenAI-compatible LLM endpoint, self-hosted TTS) and a deterministic multi-source retrieval layer routed across six Model Context Protocol servers, we report three things a research prototype typically lacks. First, a released family of fine-tuned Arabic Islamic model artifacts: an efficient tool-routing LLM (Muslim-6B-PRO, 5.94B parameters) and a Modern Standard Arabic TTS model (Fasih-TTS-V1) that ranks 5th of 17 overall and 2nd of 11 open-weight systems on the community-voted Arabic TTS Arena for MSA. Second, an account and metering layer - a free per-account turn allowance, capacity-aware refusal, and email verification deferred to the point it actually matters - that turns an open demo into an operable, abuse-resistant product. Third, a three-layer observability stack (liveness, error reporting, product analytics) built specifically around the system's characteristic failure mode: a GPU-bound agent host going silent while the web tier keeps serving normally. We report real, measured latency and accuracy figures (98.4% recitation-validation accuracy on 124 cases; end-to-end voice latency of 0.9-1.7s) and discuss the concrete engineering trade-offs and limitations of running an Islamic-knowledge voice product in production.

We present ClearGS for 3D Gaussian Splatting (3DGS) from handheld videos with uneven viewpoint coverage and mixed frame quality. Rather than selecting frames with binary decisions, ClearGS uses Reliability-aware View Allocation (RVA) to assign graded raw-supervision weights based on appearance reliability, degradation risk, and geometric utility, while weakly reactivating useful suppressed frames to maintain trajectory coverage. Since weighting cannot restore details lost to blur or distortion, ClearGS further introduces Render-Guided In-Video Restoration (RIVR). The current 3DGS render provides a pose-aligned structural candidate, a frozen no-reference restoration expert restores the corresponding raw video observation without any clean reference image, and no-reference perceptual scores select among the render, restored observation, and high-frequency fused candidate. ClearGS then applies Full-Trajectory Repair Consolidation to revisit accepted repairs and preserve details introduced early. On GS2E and GSOTM, ClearGS achieves state-of-the-art overall performance, with consistent CLIP-IQA and MUSIQ gains and LPIPS reductions in most degradation settings, without paired sharp supervision or matched clean references.

Large language models (LLMs) exhibit substantial performance disparities between high- and low-resource languages. Beyond lower task performance, they often fail to capture the cultural norms and values of underrepresented communities. In this work, we present the first systematic evaluation of cultural awareness in LLMs for Haitian Creole, a language spoken by millions but severely underrepresented in digital resources. We assess cultural awareness along four complementary dimensions---specificity, bias, diversity, and variation---using a benchmark of culturally salient prompts curated by native speakers in a text infilling setting. Our results reveal a clear gap between cultural awareness in Haitian Creole and higher-resource French, with Haitian performance being more uneven across domains and more affected by French linguistic interference. Story generation further reveals recurring portrayals of Haitian characters through hardship and resilience, showing that even positive characterizations can encode stereotypical narratives. Our code, benchmark, and evaluation framework are publicly available.

Despite the growing capabilities of large language models (LLMs), prompt design remains largely heuristic and ad hoc. This project will explore $\textit{prompt minimization}$, the process of reducing prompts to their smallest, most information-dense form while preserving output fidelity. Practically, shorter prompts reduce computational overhead and inference latency, especially when large contexts, such as entire documents or codebases, are included unnecessarily. Further, longer prompts can damage LLM reasoning and accuracy. Theoretically, the existence of multiple prompts yielding equivalent outputs suggests a high degree of redundancy in the input space, raising fundamental questions about what information is essential to elicit specific model behaviors. We propose three variant frameworks to identify and evaluate minimal prompts and demonstrate that minimal prompts often produce outputs comparable to those of their longer counterparts. These findings suggest new directions for efficient prompt engineering and deepen our understanding of input compression in LLMs.

Search is one of the most important features in e-commerce, directly driving customer engagement and business growth. A good product search system must show both relevant and desirable results. However, retail search presents unique challenges. User intent can range from exact matches to open-ended discovery. Search systems must also balance multiple goals, such as relevance, revenue, and profit, while keeping response times low. Traditional keyword-based methods often fall short in handling natural language or semantic queries. Vector search helps alleviate these issues, but it can miss key intent signals or return low-precision results. In this paper, we present the design of a hybrid search system at Target that combines lexical and vector search. We describe our approach to data processing, embedding training, precision control for the final result set, multi-channel result fusion (where we compared fusion strategies and adopted weighted interleaving), and the performance optimizations used to maintain low latency for production deployment. Our method improves offline evaluation metrics, and in online A/B testing it raised click-through rate by 0.97%, order conversion by 0.98%, and demand per visitor by 1.10% over lexical-only search, while roughly halving zero-result searches. The resulting system is deployed at scale and serves millions of guests daily.

Forecasting short-horizon mid-price movements from limit order book (LOB) data is central to algorithmic trading, yet most deep LOB forecasters are point predictors: they output a direction or a displacement, but never indicate which of their forecasts can be trusted. We introduce UQ-LOB, a lightweight, encoder-agnostic uncertainty quantification module that attaches to any pretrained LOB encoder and, in the spirit of attentive neural processes, conditions each forecast on a context set of recently completed windows whose outcomes are already realised. The UQ-regression variant outputs a calibrated Gaussian over the future tick displacement, while the UQ-classification variant outputs a categorical distribution over down/up/stationary. Both expose a scalar confidence (predicted signal-to-noise ratio or class probability) that supports selective prediction. On 5.2 billion LOB events across seven cryptocurrency assets and horizons of 5, 10 and 15 seconds, UQ-regression attains near-nominal 68% interval coverage, and restricting to the most confident 10% of predictions raises directional macro F1 by 0.11-0.15 for UQ-regression and 0.05-0.11 for UQ-classification, at every horizon. On large, economically meaningful moves, the tightest confidence tier reaches a directional F1 of 0.88 (down) and 0.83 (up) at the 5-second horizon.

Density functional theory (DFT) strikes a practical balance between accuracy and computational cost in many problems of computational chemistry and materials science. However, many DFT calculations are limited by fixed atom-centered basis sets, which dictate how accuracy and cost scale with system size. We propose Gaussian Splatting for Density Functional Theory (GS-DFT), which represents molecular orbitals as a cloud of Gaussians whose positions, shapes, and mixing coefficients are optimized jointly by gradient descent to minimize the energy without training data. Conceptually, GS-DFT is 3D Gaussian splatting with the renderer replaced by quantum mechanics. We introduce two key solver components: adaptive density fitting with screening for efficient evaluation of two-electron integrals, and a regularized differentiable orthogonalization of the molecular orbitals. Empirically, the optimized basis reaches the accuracy of the largest conventional basis sets with a fraction of the parameters, converging systematically in energy, density, and nuclear forces. At equal parameter count, it captures the stretched-bond and anion physics that fixed bases only recover with specialized basis augmentation. The resulting solver exhibits quadratic peak memory scaling in the cloud size, allowing us to simulate systems of up to 2,742 atoms (10,406 electrons) without any modifications at triple-zeta scale using a single four-GPU node.

Generative AI is reshaping the cultural infrastructures through which knowledge is found, synthesized, and held accountable. To make sense of this shift, scholars and policymakers reach for historical analogies of technologies such as the printing press, steam power or electricity. But these comparisons are typically imprecise about which property of the technology carries the comparison, and imprecise analogies produce imprecise governance by designing interventions against the wrong property of the system. This paper offers a diagnostic framework for analyzing how generative AI can transform epistemic and cultural practice. This paper offers a diagnostic framework for analyzing how generative AI can transform epistemic and cultural practice. We decompose each intervention into three coordinates: the epistemic site at which a technology acts, the governing logic by which it organizes its object, and the technical mechanism through which the logic is instantiated. This framework allows us to distinguish between structural cultural consequences, which follow from the mechanism itself, from contingent ones, which remain open to design and institutional choice. Applying the framework to information discovery and knowledge synthesis, we show how the shift from indexicality to inference and from editorial authority to statistical consensus produces specific, traceable cultural effects and reveals governance levers that gestalt analogy obscures.

We introduce Kaggle Game Arena, an open and ever-expanding platform to evaluate large language models (LLMs) through competitive games. Different from static benchmarks, game arena enables models to play head-to-head matchups in structured environments where the gameplay strength naturally increases as models evolve, preventing performance saturation. This technical report details the infrastructure behind Game Arena and describes the three pilot game environments: Chess, Poker, and Werewolf. These environments span perfect information, imperfect information, and multiplayer game settings, enabling a systematic study of models' strategic planning, adaptation, and robustness under uncertainty. For each game, we provide a detailed description of the environment, evaluation metrics, and results from running full competitions across models. Through robust infrastructure and large-scale ground-truth based evaluation, Game Arena ensures reproducibility, transparency and generalizability to new games and variants over time.

Outliers are essential for evaluating and improving the robustness of machine learning systems, especially when future distributions may differ significantly from historical training data. In high-stakes applications, robustness often depends on rare cases that finite datasets fail to capture, making simple resampling or perturbation insufficient for stress scenario generation. Existing outlier synthesis methods typically rely on sparse neighborhoods, low support latent regions, or classifier boundary crossings, which can be heuristic, unstable, and tied to specific modalities or architectures. We therefore propose Sinkhorn Boundary Outlier Generation (SBOG), a structured framework for latent-space outlier generation that couples Sinkhorn optimal transport geometry with distributionally robust boundary modeling. The resulting Sinkhorn-induced support cost guides the sampler toward weakly supported boundary regions, while semantic constraints prevent uncontrolled drift from the intended context, yielding controlled deviations from the in-distribution reference measure rather than arbitrary sparse-region samples. Experiments on time series anomaly generation and image outlier synthesis show that our framework produces informative, semantically controlled outliers and improves downstream robustness evaluation across modalities, providing a foundation for stress scenario generation beyond empirical support.

Shapley values, a solution concept from cooperative game theory, have recently become a standard tool for feature credit allocation in machine learning. They provide an axiomatically justified method to fairly distribute a model's prediction among the data features. Shapley values have not yet been applied to explain machine learning models trained on features from topological data analysis. We develop what we believe is the first such approach, focusing on the persistence landscape featurization of persistence diagrams. Because each landscape coordinate is a rank statistic, crediting a model's prediction back to individual persistent homology classes (persistence diagram points) is nontrivial. We introduce LandscapeSHAP, a method for fair credit allocation to persistence diagram points based on a model's prediction. For linear models on persistence landscapes, LandscapeSHAP has a closed form expression that gives the exact Shapley value of every persistence diagram point. In particular, there is no coalition sampling required. We further prove that the four Shapley "fairness" axioms uniquely characterize this credit allocation for any model, not only linear ones. For a general nonlinear model, this unique value can only be calculated exactly from its defining coalition averaging formula, which requires considering all $2^N$ many coalitions, where $N$ is the number of points in the persistence diagram. This is computationally intractable for persistence diagrams of realistic size. We complement the exact linear model result with an efficient Monte Carlo sampling of persistence diagram coalitions. We give convergence rates in terms of number of samples needed to approximate to a desired degree of accuracy. We also prove stability results for the LandscapeSHAP credit allocation, for any model.

LLMs increasingly act as purchasing agents, which makes the LLM, not the user, the one choosing among the options that satisfy a request; its preferences quietly fix what gets bought and what it costs. Hotel booking is a clean instance: a high-volume choice settled on a few comparable attributes, where the pick reveals those preferences. We introduce PriceBench, a diagnostic benchmark that recovers an LLM's price, quality, and brand preferences from its booking choices with a logit choice model, applied to 28 LLMs from 8 providers on 3,600 hotel tasks from 179 real New York City properties. We find that capability is associated with how consistently an LLM chooses, not with what it chooses: more capable LLMs hold stronger, more consistent preferences, while weaker ones either lock onto one position, exploitable by whoever controls listing order, or choose almost indifferently. What those preferences favor varies sharply across providers and even within one family: price sensitivity spans more than an order of magnitude, and the price/quality trade-off moves mean booked nightly price from \$247 to \$393 on identical tasks. What an agent buys must therefore be measured per LLM, not inferred, and we release the tasks, code, and all 28 response sets.

Graph neural networks (GNNs) rely on message passing over graph edges, making their computational and memory costs strongly dependent on graph density. Graph sparsification offers a natural way to reduce these costs, but removing edges indiscriminately can distort important communication structure and degrade predictive performance. We introduce Scaffold, a topology-based, unsupervised graph sparsification framework derived from support graph theory preconditioners. Scaffold explicitly controls two complementary structural quantities: dilation, which measures the length of rerouting paths induced by removed edges, and congestion, which measures how strongly these rerouted paths concentrate on the retained support. By jointly controlling dilation and congestion, Scaffold preserves short communication paths while avoiding structural bottlenecks. To our knowledge, Scaffold is the first scalable GNN sparsification framework to use a joint supporting-path dilation-congestion criterion. Across 19 homophilic and heterophilic benchmarks spanning small to large graphs, Scaffold achieves the best aggregate rank among the evaluated sparsification and related methods. Using only 10%-50% of the original edges per sparse support, Scaffold recovers or closely approaches full-graph GNN performance while using less than half the memory of full-graph training and reducing end-to-end training time, including sparsification overhead. We provide an open-source software package at https://github.com/siddhartha047/Scaffold.

Infant movement analysis provides valuable biomarkers for the early identification of neurodevelopmental disorders. Recent advances in deep learning have enabled automated analysis of infant movements from video-derived skeletal representations, achieving performance comparable to expert assessment for tasks such as General Movement Assessment (GMA). However, most existing approaches rely on centralized training, requiring data from multiple institutions to be collected and stored at a single site. Such assumptions are often impractical in clinical settings due to privacy, governance, and data-sharing constraints. To address these challenges, we present, to the best of our knowledge, the first federated learning framework for automated infant movement analysis and General Movement Assessment using skeletal motion data. As a clinically relevant use case, the proposed framework is evaluated on fidgety movement classification. To quantify model confidence, Monte Carlo (MC) Dropout is employed to estimate predictive uncertainty during inference. Building upon this, we propose an Uncertainty-Aware Federated Averaging (UA-FedAvg) strategy that incorporates predictive entropy derived from MC-Dropout into the federated aggregation process, enabling client contributions to be adjusted according to their predictive uncertainty. Experiments were conducted using a cross-subject evaluation protocol under a three-client federated learning setting. Results demonstrate that federated learning substantially improves classification performance compared with independently trained local models while achieving performance approaching that of centralized training. Furthermore, UA-FedAvg and its variant incorporating validation loss generally outperform conventional FedAvg across the evaluated data-split configurations.

Topic modeling is an effective technique for discovering hidden themes within documents and is widely used in text mining and data analysis across a variety of industry sectors. Recently, large language model (LLM)-based topic models have been emerged that prompt LLMs to generate topics then assign the topics to documents, producing more natural and human-readable topics than conventional topic modeling algorithms. However, the nature of topic assignment process causes certain drawbacks, such as the incapability to produce topic distributions over a document, too broad or narrow topics, and high resource consumption, which increases with the number and length of of documents being assigned topics. These issues are particularly critical for industrial applications, which require high-quality, in-depth analysis and the processing of large volumes of documents. In this context, this paper introduces a framework called SeLATM, which addresses these concerns by employing segment-level topic generation and topic refinement through agentic feedback loops. Experimental results on various datasets demonstrate that SeLATM significantly reduces the LLM resources compared to methods based on topic assignment process, while maintaining superior performance.

Transformers have become a central architecture for in-context learning (ICL), particularly through their state-of-the-art performance in large language models. This success motivates understanding how transformers exploit task-relevant structure in geometrically heterogeneous data. However, existing nonparametric ICL theory has largely focused on Euclidean domains or single-manifold models. To address this gap, we study the prediction problem under unknown local geometry, modeled by sample size-dependent mixtures of manifolds with heterogeneous dimensions, smoothness, and sampling masses. Under local separation and small-perturbation conditions, we establish a minimax lower bound capturing the aggregate difficulty of the components and construct an oracle tangent local-polynomial estimator with a matching upper bound. This estimator is connected to a structure-informed, two-stage softmax transformer with a geometric preconditioner and chartwise reduced local-polynomial solvers. The transformer achieves negligible approximation error relative to the minimax rate with logarithmic depth and polynomial size. Finally, we derive an in-context generalization bound for near empirical risk minimizers over this class. Together, these results identify conditions under which the resulting predictor exploits local geometry and attains the aggregate minimax rate.

Federated learning (FL) data corruption can affect either inputs or labels, but it remains unclear whether input-conditional uncertainty and prediction-label loss expose these corruption modes equally. This paper compares two corruption-detection signals in FL: input-conditional uncertainty and prediction-label loss. The uncertainty signal is characterised using a learned aleatoric variance estimate together with Monte Carlo (MC) dropout variance and entropy measures, while the loss is computed against the supplied label. We test these signals against additive image noise and persistent random label flips. On ResNet-20 with CIFAR-10 and SVHN under Dirichlet partitions with data that are not independent and identically distributed (non-IID), the two corruption types behave differently. For persistent random label flips, the within-client per-sample area under the receiver operating characteristic curve (AUC) is 0.85 on CIFAR-10 and 0.95 on SVHN for prediction-label loss, while every uncertainty estimator stays at chance (0.49--0.50). This pattern is consistent with the model remaining confident in the underlying image despite the supplied label being wrong. For image noise, expected-entropy uncertainty rises above chance (0.67 on CIFAR-10 and 0.66 on SVHN), while loss responds comparably (0.64 on both). Each signal is therefore the stronger detector for a different corruption: the prediction-label loss for persistent label flips, and expected-entropy uncertainty for image noise, with its advantage becoming apparent as federation-wide corruption prevalence increases. Robust FL data-quality assessment should match the signal to the corruption rather than rely on uncertainty alone across corruption types.

Medical vision-language model (VLM) post-training is commonly evaluated through answer accuracy. We examine how changes in accuracy and training objectives relate to image-conditioned decisions in a controlled Qwen2.5-VL-3B study on PMC-VQA. We compare supervised fine-tuning (SFT) with low-rank adaptation (LoRA) restricted to the language model, expanded multimodal adaptation scopes, standard answer-only Group Relative Policy Optimization (GRPO), and a counterfactual evidence objective. On 2,000 clean-test questions, language model LoRA SFT changes correct-image accuracy by +1.10 percentage points (95% paired bootstrap CI:-0.85 to +3.05), while visual-benefit events decrease by 2.40 points and image sensitivity decreases by 5.60 points. Paired records reveal 155 acquired and 203 lost visual-benefit events. Broader adaptation yields lower correct-image accuracy than language-model LoRA SFT. Standard GRPO produces mixed-reward groups and parameter updates, with an uncertain clean test accuracy change. A generation audit reveals that canonical option scores can follow a different token path from generated answers. With scores taken along the greedy generation path, the evidence target improves on the training set; its gains over standard GRPO remain inconsistent on validation data at matched training doses. Sample-level analyses trace how evidence scores, decision margins, and generated answers change during post-training. This empirical and measurement audit identifies gaps between optimization activity, target acquisition, and useful held-out visual behavior.

Clinical prediction models estimate risk from patient measurements, while large language models support medical text understanding and question answering. Yet their language capabilities do not ensure accurate prediction from structured, high-dimensional clinical time series. Improving this ability would connect risk estimation with flexible questions about a patient's evolving condition. We introduce ViSTA, a compact adapter that incorporates irregular numerical measurements into a pretrained vision-language model's chart representations. It learns corrections to visual tokens while leaving all pretrained parameters unchanged. On MIMIC-IV, ViSTA has the highest mean scores among the compared adaptations on all four metrics for acute kidney injury and mortality prediction across models with 2-9 billion parameters. With 0.516 million trainable parameters, the 2-billion-parameter model reaches an area under the ROC curve of 0.7376 for acute kidney injury, compared with GPT-5.6 Sol's 0.7380 with text input and high reasoning effort. Training for temporal question answering yields 69.27% accuracy at 4 billion parameters with over 90% fewer trainable parameters than low-rank adaptation using charts or numerical text, at a 2.82-4.88 percentage-point accuracy gap. ViSTA extends pretrained language models to numerical prediction and temporal questions.

With the advent of snapshot cameras, hyperspectral video is becoming more readily available. In recent years, new applications have emerged which have led to increasingly larger datasets. However, hyperspectral video compression remains in the early stages. In this study, we explore the use of implicit neural representation as a candidate solution. We propose a novel extension of an existing RGB video compression model, achieving Bjøntegaard Delta PSNR gains of +4.99 dB and Bjøntegaard Delta rate of -88.88% compared to traditional hyperspectral image compression methods applied frame-by-frame. In addition to reconstruction quality, the effects on downstream task performance are measured in the form of object tracking success. Compared to video compressed with methods based on principal component analysis and JPEG2000 in low data regimes, our proposed method improves tracking area under the curve by up to 23.42% and distance precision by up to 35.56% on examples from the HOT2026 dataset.

Tool observations dominate the context of software-engineering agents, making long interaction histories costly to maintain. Existing context compression methods can discard information needed by later actions, while adapting agents to soft-token representations can compromise their original behavior. To reduce context while preserving action-critical information and agent behavior, we combine Latent Observations, Hard Actions (LOHA), a context layout that separates compressed history from text needed for exact reference, with Anchored Context Distillation (ACD), a training method that enables latent reading while constraining behavioral drift. LOHA compresses older tool observations into soft tokens while retaining the agent's own turns and the last K observations in text, providing compact access to historical information and exact access to recent content. To enable the agent to use this representation, ACD distills the base model's full-text predictions into the latent view while anchoring its behavior on plain-text inputs to the same base model. On SWE-bench Verified, K=3 reduces context per call by 43% for Qwen3-4B and 57% for SWE-Master-4B-RL, with resolve rates of 12.1% and 21.8% versus 14.5% and 27.5% for their uncompressed bases. A single-run recency sweep reaches 14.4% and 23.0% at K=8, with larger windows generally favoring task performance over compression. Under a 32K-token limit, Qwen3 with K=3 resolves 21.1% of a 199-instance subset versus 11.1% for the same adapted agent using full text. In concurrent single-GPU serving, it achieves 1.9 times that full-text agent's instance throughput.

Large language model-based multi-agent debate (MAD) systems are being increasingly used as complex decision pipelines in distributed processes. However, their final synthesis phase still remains inadequately controlled. Even with detailed debate logs, summarizing models are prone to fabricating smoothly written debate consensus that is not grounded in the debate's history. To address this safety gap, this paper presents empirical research and studies if the introduction of active post-debate verification can mitigate the production of such factually unsupported summaries, while still providing valuable information. Furthermore, it is examined whether explicitly signalling divergence is preferable in the absence of a reliable compromise. The Active Provenance Gate (APG) is introduced as a post-debate verification layer that treats the source as a hard constraint, analysing the debate logs, auditing each claim, and applying self-correction. In crisis simulations, the self-healing mechanism more than doubles the average data Provenance Fidelity in difficult condition scenarios, before the strict gate blocks unsupported claims and generates divergence reports. In the human study, a vast majority of the users (over 75%) preferred a report explicitly stating failure in critical scenarios, despite most of them perceiving fabricated consensus from the baseline system as more fluent. Our main contribution is the transition of data origin tracing from passive logging to active conditional blocking before publication.

Position-independent KV cache reuse aims to reduce latency in retrieval-augmented generation by reusing chunk-level KV caches across prompts. We show that current evaluations of KV cache reuse techniques rely on measurements that fail to faithfully capture the loss of accuracy attributable to reuse, often artificially inflating the reported effectiveness. We also show that existing datasets do not exhibit the reuse dynamics needed to thoroughly evaluate such techniques. To address these issues, we propose an evaluation methodology that measures this accuracy loss without ambiguity and we introduce Boxoffice, a tool that programmatically generates evaluation datasets that exercise challenging KV cache reuse patterns.

Third-party libraries (TPLs) are widely used in Android apps, but their reuse can introduce security risks and interfere with downstream program analyses. Existing Android TPL detection approaches face two key limitations: their hand-crafted features are fragile under aggressive code transformations, and their whole-library matching strategies are ineffective when apps retain only part of a TPL. In this paper, we propose LibFan, a learning-based Android TPL detection approach based on context-aware functional modeling. It realizes this modeling through two complementary components: context-aware contrastive learning at the method level and functional partitioning at the library level. At the method level, it learns semantic representations through contrastive training while incorporating outgoing call relationships and class-level context, improving robustness to obfuscation, shrinking, and optimization. At the library level, it partitions each TPL into functionally coherent units and determines library presence using the best-matching partition, thereby accommodating partial library reuse. To evaluate LibFan, we construct a new benchmark comprising 200 apps and 46 vulnerable TPLs, with each app compiled under four transformation configurations. Under the most challenging R8 full mode, LibFan achieves F1 scores of 81.3% at the library level and 47.6% at the version level, representing relative improvements of 64.9% and 35.6% over the state of the art, respectively.

Vision-language models (VLMs), despite their success in optical character recognition (OCR) tasks, are vulnerable to typographic attacks and have a fragile structure for images with multiple text layers. In this study, the DecoyBench dataset was created using the Decoy Font method. The dataset consists of 300 images, each containing text with sharp contour lines superimposed on another text with soft shading. Six recent closed-source models from three different model families were evaluated using this dataset under two different prompting conditions (naive and guided) and at two different resolutions ($512\times512$ and $64\times64$). A validation study showed that human participants could read both text layers with high accuracy. In contrast, the models, with most variants and both prompting methods, read the contour text with near-human accuracy at high resolution, but almost never fully extracted the shading text. At low resolution, the contour text could not be read by either the models or humans, while the shading text could be extracted with high accuracy. The findings indicate that the evaluated VLMs exhibit a consistent behavioral limitation when processing typographic structures containing multiple spatial frequency layers.

Auditory Attention Detection (AAD) utilizes electroencephalographic (EEG) signals to identify a target speaker in a multi-speaker environment. Despite considerable progress, existing deep learning architectures often lack explicit mechanisms for handling noisy EEG data. To address this limitation, we propose Auditory Focus Attention Networks (AFA-Net), a machine learning framework that replaces vanilla attention with a simple yet flexible differential attention mechanism to help focus on task-relevant neural activity. AFA-Net achieves an upward accuracy of 96.8% at the 2s decision window, while using substantially fewer parameters than most existing methods. To the best of our knowledge, AFA-Net is among the first frameworks to explicitly try to combat EEG noise to improve AAD.

Prior work established that a probe can decode an in-context binding on model errors and that probe-guided steering can repair some of them. We follow probe accuracy, model output, and steering response across public pretraining and post-training checkpoints. Probe accuracy rises during Pythia pretraining, while probe-guided steering moves from negligible all-trial benefit to a larger benefit at two model sizes. Saved scores distinguish probe-correct errors with low and above-uniform model probability for the correct candidate. Oracle-target steering already repairs many early errors, but saved aggregates cannot separate target quality from intervention sensitivity. A held-out comparison of decoders trained on the final state or candidate logits finds no detected final-state advantage on late-checkpoint model errors. An information-theoretic counterexample explains why decodability on errors alone cannot establish discarded output information. The connection to downstream omissions remains open.

Multimodal emotion recognition (MER) increasingly pairs EEG with speech, treating internal neural signals and external vocal expression as informative views of affect. In practice, naive fusion underperforms the stronger single modality, because EEG artifacts inject noise that corrupts the shared representation. We introduce EmoSpeechBrain, a multimodal framework built on the insight that noise suppression is a precondition for effective fusion. Its EEG encoder uses differential attention, taking the difference between two attention maps to cancel shared noise and isolate discriminative neural activity. An attention-based gating adapter aligns both modalities in a shared space and weights each one's contribution to the prediction. On two datasets - PME4 and EAV, EmoSpeechBrain improves MER accuracy by up to 12.9% over other state-of-the-art (SOTA) EEG encoders, and surpasses unimodal speech and EEG baselines by up to 13.1% and 23.1%. These results show that once EEG noise is suppressed, fusion delivers gains that naive combination cannot.

Automated and highly usable Quality-of-Service (QoS) enforcement requires translating high-level service intents into deployable traffic-management policies. Although intent-based networking (IBN) has simplified policy specification, bridging the gap between business-level intents and executable network configurations remains complex, error-prone, and difficult to automate. This paper presents Intent2Tc, a closed-loop language-model-driven framework that translates business-level traffic-shaping intents into declarative sub-intents and subsequently into validated, executable Linux traffic control (tc) configurations. The framework integrates an Active Queue Management (AQM)-based digital twin (DT) semantic model, automated metadata extraction, critique-driven refinement, and Retrieval-Augmented Generation (RAG)-based knowledge reuse to improve semantic consistency and configuration reliability. We evaluate multiple open-source large language models (LLMs) and small language models (SLMs), together with Claude Sonnet-4.6, on 100 Request for Comments (RFC) 9315-compliant traffic-shaping intents. Across both translation stages, Intent2Tc achieves high semantic fidelity, configuration accuracy, and deployment readiness, with Claude Sonnet-4.6 reaching 0.98 semantic similarity, 1.0 semantic unit coverage, and 0.045 normalized edit distance. Furthermore, RAG reduces token consumption and inference latency while enabling compact models such as Phi-4-mini to approach the performance of substantially larger models. Linux tc serves as the target configuration platform, demonstrating the practical applicability of the proposed framework.

Agentic LLM inference accumulates long KV caches across iterative observation-reasoning-action loops, imposing substantial memory overhead and limiting serving throughput. Existing compression methods emphasize overall output quality, overlooking the asymmetric importance of actions in driving task progress. Our key idea is to establish a compression criterion that values KV entries by their contribution to action generation and prioritizes action quality. However, iterative execution, dynamic memory demands, and scattered action-critical entries pose challenges to eviction policies, budget allocation, and paged memory integration. To this end, we propose ActKV, the first KV cache compression framework tailored for agentic LLM inference. (i) Action-oriented KV cache eviction exploits stable action access patterns to retain entries critical to future actions, supporting reliable task progress under compression. (ii) Confidence-driven adaptive budget allocation uses LLM's intrinsic confidence to adapt the budget to evolving action-critical memory demands. (iii) Page-aware compression management standardizes compression into three primitives with customized kernels, realizing practical throughput gains. On long-trace tasks, ActKV retains an average of 98.53% of FullKV's accuracy with only 25.98% of its peak KV cache memory. It also achieves 3.97 times and 3.58 times FullKV's token and task throughput, delivering state-of-the-art performance.

End-to-end driving policies are commonly trained through open-loop behavior cloning, yet they must ultimately operate in closed-loop when deployed on a vehicle, creating a fundamental mismatch between training and execution. Beyond the commonly studied effects of covariate shift and causal confusion, we identify a complementary factor for this open-loop/closed-loop gap: waypoint-based supervision and displacement metrics do not ensure that the intermediate trajectory is physically coherent or easy for the controller to track. We observe that these inconsistencies concentrate primarily at intermediate waypoints, while the predicted endpoint remains comparatively reliable. Based on this observation, we introduce Endpoint-Constrained Optimization (ECO), a lightweight postprocessing layer that anchors the trajectory to the vehicle's executed history, preserves the policy's predicted endpoint, and reshapes the intermediate waypoints to improve feasibility. ECO requires no map, privileged simulator state, or additional training, and can be inserted between a broad range of waypoint-emitting policies and their controllers. Across two closed-loop simulators, it improves the aggregate closed-loop score of all six evaluated generative and regression-based driving policies, and the gains tend to increase with how often the base plans violate motion limits. On HUGSIM, ECO improves VaVAM from 18.1 to 31.0 HD-Score (+71%), achieving 1st place on the HUGSIM Closed-Loop Driving Challenge. Similarly, on AlpaSim, ECO increases the scene scores of VaVAM and DiffusionDrive by 123% and 22%, respectively. These results show that for a broad collection of end-to-end driving models, repairing the intermediate geometry of predicted trajectories without changing the policy's predicted endpoint can substantially improve closed-loop performance.

Long-context understanding requires large language models (LLMs) to reason over lengthy documents, conversations, and code, yet task-relevant evidence is often sparse and scattered amid substantial irrelevant and redundant content. We propose Highlight-Then-Summarize (H2S), a compress-then-reason paradigm that first identifies source-grounded, question-relevant evidence and then integrates it into a compact, question-conditioned summary before producing the final answer. To train this behavior, we construct H2S-Dataset, comprising 6,647 examples from 11 benchmark families with an average context length of 43.9K tokens, and introduce H2S-RL, which provides process-level rewards for evidence selection and summary construction in addition to final-answer correctness. We evaluate on H2S-Bench, a seven-task long-context suite. Under a shared 128K input and 4K output budget, H2S-14B achieves an average score of 32.60, outperforming Qwen3.8-27B by 10.17 points and obtaining the strongest overall result among the evaluated open-source models. H2S-14B also achieves the highest Evidence-Summary Quality score and retains 97.1% of its 16K-budget performance with only a 4K output budget. These results show that explicitly selecting and integrating evidence improves long-context reasoning while enabling more compact generation.

Context projection replaces older tool observations with compact, addressable excerpts, reducing repeated input while potentially adding evidence-retrieval turns. We study this trade-off in ReVerPi, a Pi extension with archived observations and matched full/projected continuations. In an 86-run source-reading campaign with 641 model requests, the 15 completed pairs show identical success: 12/15 per arm. Twelve further boundary runs stop, with the runner suppressing the companion whenever the first arm fails to complete. Restoring all 27 boundary runs bounds projected-minus-full success between $-$9 and +1 tasks. One omitted, selector-chosen projected continuation successfully retrieves archive text yet exhausts twelve requests; its full counterpart answers in three. The eleven jointly correct pairs form a fully observed success stratum within this recorded frame: projection reduces aggregate logical tokens by 25%, while increasing the median pair's tokens by 29% and total suffix requests from 35 to 55. Separating fitting from evaluation changes the selector's apparent tie: outside its four fitting pairs, it incurs one extra failure and 8.6% more logical tokens over thirteen comparable runs. This methodological case study connects stopping rules, known bounded failures, unexecuted companions, and resource aggregation. Its findings concern the recorded campaign, rather than population noninferiority or superiority over unrestricted Pi. Evaluations should retain every intervention boundary, execute both allocated arms independently of the first arm's completion, and report completion alongside interaction and token expenditure.

Large language models (LLMs) have demonstrated promising performance in log anomaly detection, yet how their adaptation strategies, architectures, and deployment configurations affect detection effectiveness remains insufficiently understood. To investigate these factors, we conduct a systematic empirical analysis across three public log datasets, examining different adaptation strategies, model architectures, parameter scales, and quantization settings. Our results reveal substantial performance differences across adaptation strategies, while model scaling yields varying detection gains across datasets. We further observe that models with comparable detection accuracy can exhibit markedly different computational costs, and that low-bit quantization largely preserves detection performance in the evaluated configurations. Finally, we examine detection robustness under structural, semantic, and label noise at different perturbation levels. These findings provide empirical insights into the performance, efficiency, and robustness of LLM-based log anomaly detection, highlighting practical considerations beyond conventional accuracy-oriented evaluation.

Tree-Adjoining Grammars (TAGs) have recently been introduced to Nonlinear System Identification (NLSI) as a means of encoding an entire model class as a finite set of grammatical rules, from which candidate models are assembled as trees. Existing TAG-based identifiers rely on evolutionary optimisation and return point estimates of the model structure. This paper instead proposes the TAG framework within a Bayesian setting. A generative prior is defined over tree structures and their parameters, and a Reversible-Jump MCMC sampler with structure-preserving tree moves is used to infer the joint posterior over model structure, parameters and predictions. Two training objectives are considered; that is, a one-step-ahead objective with conjugate parameter proposals, and a simulation-based objective handled by likelihood-free inference. The approach is validated on a simulated polynomial NARX system, the Silverbox benchmark, and wave-loading data from the Christchurch Bay Tower, where embedding Morison's equation as a fixed initial tree yields a grey-box model that outperforms the physics-driven baseline. The results demonstrate that Bayesian TAGs are well suited to quantifying uncertainty in equation discovery for dynamical systems and to fitting physics-informed models.

Distributionally robust optimization (DRO) provides a principled framework for learning under distribution shift, but its practical use is hindered by the difficulty of evaluating worst-case risks for nonconvex loss functions. We study a penalized DRO formulation in which the adversary may choose any distribution but incurs a Wasserstein penalty for deviating from the empirical distribution. We show that the adversary's problem can be reformulated as an optimization problem over transport maps that push empirical samples to adversarial ones, and we prove that optimal maps are cyclically monotone. We also show that standard adversarial training---based on per-sample local optimization---violates cyclical monotonicity and wastes transport costs unless the adversary is severely restricted. We propose two remedies. First, we introduce multi-start particle ascent, which alternates parallel gradient ascent with reassignment to enforce cyclical monotonicity across samples. Second, we parameterize adversarial maps as gradients of input-convex neural networks, which guarantees cyclical monotonicity by construction. Experiments on robust regression, image classification, and robust control show that our methods consistently outperform standard adversarial training and state-of-the-art baselines, achieving improved robustness and better generalization under distribution shift.

Inference in LLMs is conventionally a fixed-depth, fixed-order forward pass through every layer, regardless of how difficult the input is. The human brain does not work this way: using the thalamus as a central hub, it routes information flexibly to all regions of the cortex according to demand. Li et al. (2026) recently showed, with a system they call program-of-layers (PoLar), that transformers can be given an analogous flexibility if their layers are treated as a library of functions rather than a fixed sequence. Performance improves over the standard forward pass when each input is dynamically routed through an adaptive sequence of skipped or repeated contiguous layer blocks. We reconstructed PoLar's diagnostic MCTS in more detail than the original paper and applied it across 5 models. We reproduced several of PoLar's findings: skipping outperformed the standard pass, repeating outperformed skipping, and combining both outperformed either alone. Shorter programs sufficed for easier questions, while harder questions required more layer repeats. However, we failed to replicate the main claim regarding their learned router for single-shot inference: its top-ranked prediction consistently collapsed back to the standard pass, even though its top-k predicted programs, taken together, did show a real accuracy gain. Beyond reproduction, we find that a small number of generic programs are enough to solve most of the questions. We also provide a much deeper analysis of these programs' structure and robustness: for example, we found that programs that correct errors are highly brittle: undoing even a single edit inside a program typically breaks the correction. Connecting this to the brain's routing mechanisms, PoLar mirrors principles of thalamo-cortical coordination between cortical-area-like transformer layers. We publicly release the code at https://datexis.github.io/RE-PoLar/

The Model Context Protocol (MCP) provides a common interface through which AI applications discover and use external resources and tools. It allows language-model agents to ground their reasoning in current system state and interact with heterogeneous services. In medical environments, however, exposing device state and action affordances requires deterministic constraints on possible effects. We present an IEEE 11073 Service-Oriented Device Connectivity (SDC)-to-MCP gateway that exposes metrics, alarms, context references, and semantic metadata as read-only resources, while representing selected action affordances as policy-validated dry-run tools. The term safety-bounded denotes a narrow no-execution property: agent-facing requests dispatch no SDC device operation. A Python prototype supports simulated fault and lifecycle experiments, a software-reference protocol path spanning independent Java and Python implementations, deterministic baselines, representation ablations, and multi-model agent evaluation. The results show semantically explicit resource exposure, visible rejection of invalid or outdated state, and preservation of the no-execution boundary across resource, proposal, and authorization paths. Explicit semantic metadata improved conformity to required metric identifiers in structured alarm outputs relative to a generic representation, while retained structured-output failures reveal a distinction between plausible narrative answers and task-compliant machine-readable results.

Vision-Language Models (VLMs) exhibit remarkable zero-shot generalization, yet they often encode unwanted or hazardous stylistic domains such as idealized textbook diagrams in medical AI or cartoon vehicles in autonomous driving. Approximate Domain Unlearning (ADU) aims to selectively erase a model's recognition of a target visual domain while preserving accuracy on the remaining domains. However, existing ADU methods operate under a flawed closed-vocabulary assumption: they evaluate unlearning solely on the specific object classes seen during the unlearning fine-tuning phase. Consequently, these methods do not unlearn the domain itself; they merely overfit to seen class-domain pairs, leaving the domain easily recognizable for unseen classes and providing a false sense of removal. We argue that true domain erasure must be class-agnostic. To address this, we formalize Open-Vocabulary Domain Unlearning (OVDU), a rigorous protocol that mandates domain forgetting must transfer to held-out classes. To solve the OVDU challenge, we propose a surgical parameter-editing framework. First, a Fisher Information mask isolates domain-sensitive weights, mathematically protecting foundational zero-shot generalization. Second, our Targeted Manifold Scattering (TMS) objective uses preference-based mining to locally scatter the forget domain's stylistic geometry. Evaluated across PACS, OfficeHome, and DomainNet, our method vastly improves open-vocabulary generalization over existing baselines. Crucially, it delivers exceptional sample efficiency, outperforming peak 8-shot baseline results with only 4 shots.

Contemporary large language model (LLM) chat systems treat conversation history as an immutable sequence of turns that defines the model's working context. However, user intent in real interactions is not static: it evolves through correction, refinement, and shifting constraints. This mismatch between dynamic intent and static transcripts can result in context pollution, where outdated or irrelevant information persists and continues to influence subsequent responses. We introduce mutable transcripts, a new interaction paradigm that enables users to revise prior turns through natural language edit requests, allowing the conversation history itself to be updated rather than appended. This reframes the transcript from a passive record into an editable representation of conversational state. We present a working prototype that integrates transcript-level revision into a standard chat interface and evaluate its feasibility through a controlled user study (n=17) and an illustrative transcript analysis of representative interaction scenarios. Participants significantly preferred mutable transcripts over standard chat across measures of clarity, confidence, and ease of use, with reduced intent to restart conversations. Transcript analysis of representative user study conversations shows that mutable transcripts can reduce conversation length and eliminate obsolete retained context. These findings provide initial evidence that user-driven revision of conversational history can improve interaction quality and help maintain a more current representation of user intent. The source code and prototype can be accessed at https://github.com/QxLabIreland/ReChat

Time-series carry structure simultaneously at multiple scales (fine-grained variation, mid-range motifs, and global properties) and downstream tasks operate at correspondingly different scales. Most existing self-supervised learning approaches supervise representations globally via instance-level contrastive losses and limited temporal neighborhood supervision, but do not explicitly exploit the structural hierarchy. We propose a learning framework that explicitly enforces a structural hierarchy across three scales independently: a local objective for token continuity, a mid-range objective for window-level motifs, and a global objective for sequence-level agreement. Realizing this framework requires the backbone to expose a representation at each scale; we introduce \textbf{Progressive Memory Transformer} (PMT), which augments a transformer with writable, window-aligned memory that exposes the mid-range scale alongside the token and sequence-level representations conventional transformers already provide. Across seven UCR/UEA/UCI classification benchmarks, a cue-retention probe, and forecasting benchmarks, PMT learns representations that probe well at the global, mid-range, and local scales---strong low-label classification (1--5\% labels), competitive forecasting performance across multiple horizons, and quantitative and qualitative evidence that memory states capture mid-range motifs.

Large video diffusion models offer expressive priors for embodied prediction and learning, yet their many-step sampling remains costly for interactive downstream use. Distribution Matching Distillation (DMD) enables few-step video generation, but can suppress robot--object motion while preserving visual quality. Examining DMD's teacher and fake-score signals, we find that weak re-noising keeps the teacher posterior concentrated near motion-deficient rollouts, limiting motion-restoring guidance. Meanwhile, stronger-motion rollouts tend to incur larger fake-score fitting errors, which can hinder the generator's learning of interaction dynamics. We propose DyMD, a DMD framework that adapts both teacher supervision and critic fitting to the evolving student. Temporal affinity--conditioned re-noise sampling adapts the timestep distribution to each rollout's current interaction fidelity by mixing the base schedule with a teacher prior motivated by local posterior variation, thereby balancing motion recovery and appearance refinement. To better track stronger-motion rollouts, dynamics-guided fake-score tracking uses a noise-conditioned predictor to estimate noise-relative fitting difficulty from latent temporal dynamics, then upweights predicted-hard rollouts in the critic loss. Using DyMD, we distill a 14B teacher into a four-step 1.3B student with no auxiliary modules at inference. On embodied-video benchmarks, the student improves R-Bench task adherence by $9.6$ percentage points and PAI-Bench-G Domain score by $5.1$ points over Base DMD while maintaining comparable visual quality. As a backbone for downstream action planning, our student achieves 34% mean success across two WorldArena tasks, compared with 16% for Base DMD.

Retrieval-augmented generation (RAG) is often used to address outdated knowledge by providing external evidence. But retrieval helps only when that evidence is still valid. We identify a temporal alignment failure, stale-document poisoning, in which outdated evidence makes a model wrong despite answering correctly without retrieval. We construct a benchmark of 317 verified knowledge reversals across medicine, law, software, and platform policy, grounded in dated official sources. Across 12 models, recent medical reversals are harder than long-established ones. More importantly, outdated retrieval flips 30% of Llama and 37% of Qwen answers even without instructions to trust the document; explicit follow instructions raise these rates to 66% and 75%. Across four open models and four domains, poisoning ranges from 17-91%, while matched up-to-date evidence is followed in 97-100% of trials. To isolate temporal applicability, we keep the historical evidence unchanged across 50 reversals and vary only the evaluation date. A clear pattern emerges: dates alone produce only modest adaptation, but when models are explicitly told when the old evidence stops applying, the larger models switch to the appropriate answer almost perfectly. Causal interventions confirm that this validity information directly shapes the final decision. The same internal components also support broader comparison tasks, suggesting that temporal applicability can recruit a general reasoning mechanism used for other comparisons. Finally, a fixed recency-aware hybrid re-ranker reduces poisoning by 4.6-10.0 points when dates are accurate, with gains that depend on reliable temporal metadata. Reliable RAG therefore requires selective trust: models must determine not only what retrieved evidence says, but whether it still applies.

Whether an information extraction pipeline should process page images or parsed text depends on the document, and the answer flips across the layout spectrum. We study this trade-off under a constraint that rules out (closed) cloud services: privacy-sensitive documents processed on-premise by small ($\le 8\mathrm{B}$ parameter) text-only and vision--language models, evaluated on both accuracy and energy over a design space spanning input representation, model family, and inference configuration. Benchmarking on the near-plain-text Kleister-NDA contracts and the layout-rich VRDU forms, we find that batching is the dominant energy lever, cutting energy per page by 38-85% at no cost in accuracy, while FP8 quantization saves 27-32% when requests are served one at a time but less than 1mWh per page (9-19%) once batching is applied. Preprocessing dominates what remains: neural OCR costs $17\times$ more energy per page than classical OCR and never reaches the Pareto frontier. Which representation wins flips with the type of document: vision--language models on layout-rich documents and small text-only models with a cheap parser on near-plain text, where they are both more accurate and cheaper than any vision--language configuration. Our work yields concrete guidelines for energy-efficient, privacy-compliant local information extraction.

Morphology--control co-design jointly optimizes an agent's body structure and control policy as an integrated embodied system. However, existing methods typically model morphology design and control with separate networks coupled only indirectly through a shared task objective, limiting explicit high-level coordination. Inspired by natural genes that coordinate biological development, we introduce \textbf{Morphogene}, a compact latent blueprint that bridges an agent's body and brain. Through AdaConcat, Morphogene jointly conditions morphology and control generation at the limb level, allowing its variations to induce coordinated changes in both components. Building on this representation, we propose \textbf{GeCode}, which formulates co-design as exploration in the compact Morphogene space. Each Morphogene anchors a local design region in which nearby body--brain designs are explored, while performance-guided updates move these anchors toward promising regions for more efficient exploration of the broader design space. This process combines local refinement with global exploration while preserving body--brain compatibility. Extensive experiments across diverse 2D and 3D co-design tasks demonstrate that GeCode consistently outperforms existing state-of-the-art methods, achieving substantially faster convergence and higher final performance.

Ordinal acne severity grading requires distinguishing visually similar neighboring grades while jointly weighing holistic facial appearance and localized lesion burden - evidence that most existing approaches collapse into a single opaque representation. We introduce CG-HAF, a global-local fusion framework that instead keeps this evidence explicit: averaged holistic severity probabilities from independently trained classifiers are combined with structured lesion-burden descriptors from an object detector (lesion count, detection confidence, lesion area) into a compact representation, from which a lightweight, interpretable classifier produces the final grade. On a widely used benchmark, this fusion yields a clear, statistically supported improvement over global-evidence-only baselines, with the largest gains on the most severe cases. Testing on an independent dataset with a different grading standard shows that strong within-dataset performance does not transfer automatically, and a follow-up diagnostic attributes much of this gap to mismatched grading criteria rather than detection failure alone. These findings support interpretable global-local fusion as an effective strategy for ordinal acne grading while highlighting criterion alignment as key to cross-dataset portability, with a further illustration of how the resulting severity signal can support transparent, non-diagnostic decision-making in skincare applications.

Continual spatio-temporal forecasting supports traffic management and environmental monitoring under evolving dynamics and expanding sensor networks. However, conventional graph-based continual learning methods tie forecasting representations to the current sensor layout, so sensor expansion can alter the representation of learned spatial relationships. Our key insight is that sensor expansion changes the evidence available about a process without necessarily changing the dynamics to be learned. We propose STFO (Spatio-Temporal Field Operator), which parameterizes forecasting knowledge as a shared field-evolution operator and handles changing sensor layouts through observation and query interfaces. Normalized coordinate-based aggregation lifts irregular sensor histories onto a fixed latent grid, enabling reuse of learned spatial maps across observation sets without sensor-specific parameters. To accommodate process drift, a spectral descriptor summarizes variation across spatial scales and conditions Fourier propagation and attention to adapt operator responses to the current spatial regime. Coordinate-based decoding queries the evolved field at sensor locations and combines spatial corrections with local-history predictions. Experiments on PEMS-Stream, CA-Stream, and AIR-Stream demonstrate state-of-the-art average forecasting performance. STFO-Large reduces average MAE over DOL by 8.4% on PEMS-Stream and 4.7% on CA-Stream. Our code is available at https://github.com/Xielewei/Spatio-Temporal-Field-Operator.

AI agents combine language models with external data and tools that can modify files, call APIs, or execute code. Security failures can arise when adversarial content changes an agent's tool use or when the surrounding software contains vulnerabilities such as path traversal or command injection. We study authorized white-box pre-deployment auditing, where the auditor has access to the target repository and a controlled runtime, but successful attacks must still act through the task-defined attacker interface and be confirmed by an external verifier. We present AgentXploit, a two-role auditing system that separates repository-level attack-path discovery from runtime exploitation. The Analyzer Agent traces attacker-controlled inputs to sensitive operations and records code-supported candidate attack paths; the Exploiter Agent turns these paths into concrete attacks and revises them using runtime feedback. We also introduce AgentXploit-Bench, containing 72 reproducible vulnerabilities across 12 open-source AI-agent systems and frameworks. Across three runs, AgentXploit reaches 59.3% end-to-end success, compared with 38.4% for Codex. Under a token-budget-matched comparison, Codex reaches 46.3%. On AgentDojo, where injection points are provided, the Exploiter Agent reaches 79.2% attack success versus 52.7% for AgentVigil. These results highlight repository discovery and runtime exploitation as distinct challenges in end-to-end agent security auditing.

Unobserved common causes are pervasive in real-world time series and can induce spurious associations that causal discovery methods mistake for direct edges. We propose LUCID (Learning Under Confounding for Inference and Discovery, a regime-adaptive deconfounding layer that first estimates the confounding regime from data using a Marčenko--Pastur spectral router, then applies a deconfounding strategy matched to that regime. When the spectrum indicates pervasive factor confounding, LUCID attenuates factor-dominated variation and recovers contemporaneous (lag-$0$) structure from the resulting innovations, with edge selection calibrated against a data-driven edge-free null. Rather than being tied to a particular discovery algorithm, it can wrap existing discovery engines; we demonstrate consistent improvements across three such methods. On a diverse synthetic out-of-distribution benchmark spanning changes in confounder strength and sparsity, loading density, lag structure, volatility dynamics, edge heterogeneity, persistence, intermittency, and tail behavior, LUCID achieves the best family-weighted directed, lag-resolved graph $F_1$ ($0.60$), improving over the strongest baseline by $0.19$ absolute ($\approx\!46\%$ relative). Its advantage widens relative to looser lag-collapsed scoring, and remains robust under intermittent and heavy-tailed confounding. Code reproducing the method, the benchmark generators, and every reported experiment is available at https://github.com/bloomberg/causal-ts.

Vision-Language-Action models (VLAs), while showing strong potential for robot control, require massive amounts of high-quality imitation learning data. Moreover, the absence of an explicit world model casts further doubt on their control capabilities. In this concept paper, we propose a novel architecture that addresses sample efficiency in VLAs by training a predictive world model on the embedding space of the VLA's vision encoder. We hypothesize that these embeddings are action-relevant and usable for future prediction. To this end, we propose using the suggested architecture to investigate how well these embeddings predict the future based on actions, as the inability to do so would mark a key limitation of VLA architectures: the lack of a non-lossy implicit world model to simulate real-world dynamics. The proposed architecture differs from the standard world model dynamics as the loss comes from the embedding space rather than the pixel space, similar to joint embedding predictive architectures. Furthermore, the trained world model can be utilized for short-term planning tasks by sampling VLA actions given goal images. We intend to examine the richness of vision embeddings in VLAs and reduce their high data requirements through a world model that can also generate plans during inference.

Tabular in-context learners such as TabPFN, Mitra, or ConTextTab rely on alternating row and column attention over 2D sequences of latent embeddings. These attention patterns differ markedly from the one-dimensional case in language models: row attention involves longer sequences while column attention operates on much shorter ones, and the strided memory layout of tabular data makes producing contiguous tensors costly. Moreover, the hidden dimensions used in current models are small compared to recent language models. Yet efficient attention has been studied mostly for one-dimensional sequences, leaving the two-dimensional tabular setting unexplored. To this end, we create a reproducible benchmarking setup and study the unique characteristics of tabular attention across several backends -- Torch SDPA (efficient and cuDNN), FlashAttention-2/3/4, and the inference-only backends vLLM and SageAttention -- measuring forward and backward throughput across realistic tabular shapes on three GPU generations (A100, H100, B200). We find that the optimal backend choice differs between column and row attention and varies across hardware as well as model specifics: While the FlashAttention implementations tailored for each GPU generation perform overall best, they are at times outperformed by CuDNN in the case of column attention at longer sequences with cross-over points depending on the head dimension. Among inference-only backends, SageAttention performs well for row attention and large sequences beyond 16\,k rows. Our reproducible benchmark lays the foundation for future improvements to table-native attention. The self-contained benchmarking and evaluation code is openly available at: https://github.com/SAP-samples/tabular-attention-benchmark

We study geometric moment contraction (GMC) of the constant-parameter stochastic Nesterov recursion \[ Y_k=Θ_k+β(Θ_k-Θ_{k-1}),\qquad Θ_{k+1}=Y_k-γG(Y_k,X_{k+1}). \] Under mean strong monotonicity and stochastic $L^p$ Lipschitz continuity, an explicit Perron comparison proves synchronous $L^p$ contraction when $βγL_p<(1-β)(1-q_{γ,p})$. This direct criterion includes infinite-variance gradients for $1<p<2$, but its small-step regime requires $β<μ/(μ+L_p)$. A complementary power-Lyapunov argument establishes a positive, generally much smaller, step-size interval for every fixed $β<1$ and every $p>1$, using only a finite $p$th gradient moment. At $p=2$, a simpler explicit certificate gives \[ 0<γ<\frac{2μ(1-β)^2}{L_2^2(1-β+2β^2)}. \] Its quadratic high-momentum scaling is a limitation of the chosen metric, not a sharp stability boundary. We quantify this loss, provide a general mean-only quadratic $S$-procedure, and exploit endpoint Lyapunov inequalities under stronger samplewise sector information. Verified endpoint certificates can be orders of magnitude less conservative than the explicit metric.

Large language model agents increasingly act on software systems, no longer merely generating text but also changing databases and online services. However, an approved database update may succeed yet leave an unapproved notification because execution can produce persistent effects beyond the requested change. Current safeguards can approve an action or record its aftermath, but without checking the persistent result before continuation, an unapproved outcome can be accepted as success and propagated to later steps. We present EffectMatch, a runtime that collects persistent changes within a controlled execution boundary and compares them with what the application approved for the current state and execution. The comparison governs commit and dependent execution. In comparative evaluation on 206 public business tasks, EffectMatch preserved all clean executions and prevented all tested incorrect commits. Six 20-run ablations exposed the failure caused by each removed mechanism, while 80 task-topology cases preserved truthful handoffs and blocked invalid continuation. Together, these results show that EffectMatch blocks the silent acceptance and downstream propagation of persistent outcomes inconsistent with application approval.

Autism Spectrum Disorder (ASD) is a complex neurodevelopmental disorder for which early and accurate diagnosis is critical to improving long-term developmental outcomes. However, existing ASD recognition methods are often constrained by the scarcity of diagnostic text data, forcing them to rely mainly on visual analysis and limiting their ability to model clinically meaningful semantic reasoning. To address this challenge, we propose UniAR, a unified framework enhanced by multi-granularity prompt learning for robust ASD recognition under heterogeneous data variations. Specifically, UniAR leverages a large multimodal model to generate hierarchical diagnostic descriptions at the word, phrase, and sentence levels, compensating for the lack of paired clinical reports. To align the generated semantics with visual evidence, we further design a Mixture-of-Experts-based Multi-Scale Alignment Module, which dynamically matches vector-quantized visual prototypes with semantic representations at corresponding granularities. Extensive experiments on four benchmarks covering brain MRI and facial expression scenarios show that UniAR consistently outperforms existing state-of-the-art methods, achieving average accuracies of 75.9\% on MRI benchmarks and 91.6\% on facial benchmarks, while improving average Accuracy on MRI benchmarks by 1.5 percentage points and average Accuracy on facial benchmarks by 1.2 percentage points over baselines. These results demonstrate that UniAR offers a robust and interpretable framework for ASD screening under semantic scarcity.

Speech-based screening is a promising, non-invasive approach for detecting Alzheimer's disease and related cognitive risks. However, models trained on a single domain often generalize poorly to unseen languages, tasks, or recording protocols. This paper investigates this deployment gap using a leave-one-corpus-out evaluation across four distinct datasets. Among 70 interpretable speech and language features, 59 exhibit direction conflicts between healthy control and cognitive risk groups across corpora, with pause, silence, and speech rate showing high protocol sensitivity. Furthermore, while the XLM-R text baseline achieves strong average performance, its Area Under the ROC Curve (AUC) drops to 0.520 on the weakest held-out domain. A standard GroupDRO baseline reaches a 0.766 mean speaker AUC and a 0.504 worst-domain AUC under the same protocol. To address this, we propose a fusion method that integrates XLM-R text baseline scores with evidence anchors selected during training. Balanced fusion achieves a 0.785 mean speaker AUC, while anchor-heavy fusion raises the worst-case speaker AUC to 0.615. This work highlights the need to audit feature transferability and report worst-case domain robustness in cognitive speech screening.

Large vocabularies make output heads a substantial inference cost in small language models. We propose softmax reparameterization, a post-training method that selects a functionally equivalent output head before quantization. The method subtracts a scalar multiple of the vocabulary-row mean from every output row and selects the coefficient by validation KL separately for RTN, activation-weighted MSE, and full-Hessian GPTQ. This one-dimensional search includes the original head and fixed mean-centering, preserves the full-precision softmax distribution, and leaves the trained decoder unchanged; a rank-one correction handles nonlinear logit paths such as soft-capping. Across seven heads, W4 gains concentrate where baseline quantization substantially distorts predictions: on Phi-4-mini, AW-MSE KL falls from 0.936 to 0.256. The gains survive stronger GPTQ calibration and remain complementary to exact per-channel scaling and affine quantization. Across four heads and three W4 quantizers, frozen WikiText-selected coefficients also transfer to C4 and OpenWebMath, outperforming mean-centering in all 18 comparisons where the frozen coefficient differs from $1$ and matching it in the remaining six. At W2, used as a compression stress test, benefits broaden across nearly the full model--quantizer matrix. Matched residual analysis shows that improved fidelity can accompany greater logit reconstruction error while reducing the residual's Fisher-weighted cost. For shift-compatible heads, reparameterization adds no inference operation and preserves packed W4 execution: with the decoder held in BF16, quantizing the Phi output head reduces batch-one generation latency by 10.8% relative to the BF16-head baseline.

Applications built on commercial large language model (LLM) APIs depend on model versions that providers retire on their own schedule, with notice periods ranging from one year to two weeks. We ask what actually happens to applications when a model is retired. We mine GitHub for commits that migrate away from officially deprecated models and endpoints of OpenAI, Anthropic, and Google, matching each commit to the provider's published announcement and shutdown dates. From 22,555 commits in 17,703 non-fork repositories (2024-2026), 5,139 are matched to an official event; two independent coders validated a stratified sample of 300 (kappa = 0.89-0.95), and we reweight all estimates by their labels. We find that an estimated 82% (95% CI 79-84) of migrations away from retired models were committed after the shutdown date - after the application had started failing - regardless of repository popularity, prior retirement experience, or the presence of a provider-abstraction layer. The share tracks the provider's notice policy: 89% for Anthropic's 60-114-day notices versus 13% for OpenAI's one-year Assistants API notice, and each e-fold increase in notice length reduces the odds of post-shutdown migration by about three quarters. Model identifiers are hard-coded in 94% of migrating applications, migration effort scales from a median of 6 added lines for prompt-only applications to nearly 700 for fine-tuned ones, and only 8% of migrations switch provider. We release the dataset and pipeline and discuss implications for deprecation policy, dependency-risk assessment of LLM products, and tooling.

Offline multi-agent reinforcement learning (MARL) learns cooperative policies from fixed datasets without further environment interaction and a learned policy is frozen at deployment. Such a frozen policy typically proposes a single joint action and executes it directly at deployment time. However, this one-shot deployment often commits to a suboptimal proposal, even when better nearby alternatives remain consistent with the behavior data. To address this issue, we propose Gradient Guided Multi Agent Flow (G2MAF), a refinement framework for optimizing joint policies at test-time. G2MAF applies one globally normalized, projected critic gradient to guide and coordinate all agents' corrections while keeping the action both feasible and close to the frozen policy proposal. Across 24 MPE and SMAC settings, its canonical variant improves 20 frozen settings, with mean relative gains of 9.2% on MPE and 8.9% on SMAC, with model inference latency increased by about 6% only.

This paper proposes a blockchain-backed agentic security framework designed to safeguard the complete software development lifecycle (SDLC) while also securing the agentic AI components responsible for monitoring it. The framework coordinates a set of specialised security agents, covering source integrity, dependency and SBOM analysis, CI configura tion auditing, artifact verification, and runtime policy evaluation, each supported by a large language model (LLM) that interprets artefacts, reasons over tool outputs, and produces structured security reports. To ensure agent trustworthiness, every agent generates a cryptographically signed attestation that is recorded in a permissioned blockchain via smart contracts, including an agent registry, an immutable attestation log, and an enforceable release-policy module. Communication among agents and with blockchain nodes is secured using a consortium-operated certificate authority, ensuring authenticated and tamper-resistant interactions. A detailed use-case and sequence flow demonstrate how a source code security agent performs analysis, anchors its attestation on-chain, and triggers a verifiable allow/block deployment decision. The proposed framework of fers decentralised integrity transparent provenance, uninterrupted security assurance and a generalisable architecture to incorporate the agentic AI into the modern software supply chain security.

Multi-agent cooperative tasks require different agents to execute a joint action simultaneously, and each agent's action affects both the observations and responses of the other agents. Hence, a world model is needed to predict the team return resulting from the joint actions of all agents. A naive extension directly applies a single-agent world model to each agent's action when predicting the team return step by step. However, such an extension fails to capture the dependencies among the simultaneous actions of multiple agents. We propose Multi-Agent World-Action Model (MA-WAM), a test-time planning framework that enables a frozen multi-agent flow policy to evaluate futures of candidate joint actions. To our knowledge, MA-WAM is the first test-time world-model planner for multi-agent flow policies. MA-WAM predicts the consequences of each joint action according to cross-agent dependencies and enables efficient candidate scoring. Across 30 offline multi-agent reinforcement learning (MARL) settings on MAMuJoCo, SMAC, and MPE, MA-WAM achieves mean relative gains of 22.0% over direct execution and 25.6% over uniform action selection. Under the standard evaluation protocol on an A100 GPU, MA-WAM adds 12.1 ms, accounting for 2.5% of the measured generation-and-scoring time.

Expert-driven Bayesian networks can support recurring software decisions when historical data are limited, but quantifying their conditional probability tables (CPTs) requires many probability judgments. Semi-automatic methods reduce direct elicitation, yet practitioners have little comparative evidence from operational decision processes. We report an embedded case study comparing the Weighted Sum Algorithm (WSA) and Ranked Nodes Method (RNM) as complete elicitation-and-quantification pipelines in two contexts within a single software research and development (R&D) organization: feature selection for Internet of Things projects and user interface design selection. Within each context, the pipelines shared the graph, root priors, and decision evidence. We evaluated them through 15 expert-defined model-walkthrough scenarios and retrospective reconstructions of alternatives recorded in decision meetings. WSA matched 7/7 and 6/8 walkthrough expectations, whereas RNM matched 4/7 and 4/8. Both pipelines placed the selected alternatives near the top of the retrospective rankings. Their complete distributions nevertheless differed: the median total variation distance was 0.304 across four distinct feature patterns and 0.340 across seven distinct design patterns, rising to 0.890 for one design pattern. These results concern two ordinal value-estimation models in one organization. They show that similar shortlist behavior does not imply equivalent representations of uncertainty. For comparable software decision-support models, method selection and validation should consider node semantics, the judgments experts can provide, calibration requirements, complete output distributions, and the intended downstream use of the probabilities.

AI is becoming increasingly integrated into daily workflows, especially in computing. We are gradually shifting towards an AI-rich future, an impending yet unknown one. One important emerging concern is whether we are accordingly preparing our future computing workforce. Further, we need to know what the important cognitive skills are to remain relevant in the computing workforce and if there are changes in cognitive skill importance. To investigate this direction, we conducted a mixed-methods study, collecting perceptions from computing students and computing experts regarding the importance of cognitive skills in the past, present, and future. We report that the perceived importance of most cognitive skills will decrease in the future, with an AI-rich environment, but critical thinking skills remain important. Further, we report reasons collected through interviews on why the importance of cognitive skills will change and how future computing students can prepare for it.

With the growing importance of science-related discourse on the Web and the erosion of the classical knowledge order, it is important to identify different user groups, such as scientists, automatically. This work proposes an approach for identifying scientists and non- scientists on X/Twitter based on their user biographies and tweets. We show that we are able to classify accounts as scientists and non- scientists on two different datasets, reaching an F1 score of up to 0.88 using Random Forests with linguistic features and up to 0.96 using a contrastively fine-tuned DeBERTa model in an ensemble setup. Furthermore, we provide two datasets with X users labeled as scientists or non scientists and their respective tweets and user biographies.

The quadratic complexity of dense self-attention remains a central bottleneck for long-context language modeling. Many efficient alternatives address this cost by deciding in advance where attention should be sparse or local. We argue that attention approximation should instead be approached as a geometric problem, with the relevant interaction geometry learned from data: natural-language dependencies are input-dependent and difficult to prescribe in advance, so the model should learn where positional relevance can decay and where broader interactions must be preserved. We introduce Mixture of Semantic Attention Regimes (MoSAR), which learns such an adaptive, controlled-decay geometry over query--key interactions. Input-conditioned query and key routers, applied after positional encoding, select mixtures over short, medium, and global regimes, inducing a continuous distance-dependent attention field rather than a fixed sparsity pattern. This geometry is learned during training and can subsequently be discretized through top-1 routing. In controlled pre-training experiments with matched 500M-parameter models, MoSAR learns a substantially lower-reach attention geometry without degrading language-modeling quality, improving perplexity over dense RoPE at the training context length. Under length extrapolation, MoSAR achieves the best perplexity among all evaluated variants, including strong baselines such as ALiBi. Moreover, the learned geometry remains stable under deterministic top-1 discretization, suggesting that it is not only adaptive, but also amenable to low-cost approximation at inference time.

We investigate the systemic macroscopic dynamics emerging from Limit Order Books (LOBs) populated exclusively by autonomous reinforcement-learning agentic traders. By formalizing agent interactions within a microscopic order-matching engine, we examine two fundamental quantitative phenomena: equilibrium phase transitions in order flow regime shifts, and the structural dynamics of market impact. We show that agentic LOBs exhibit distinct phase boundaries separating orderly price discovery from hyper-volatile cascade states, governed by critical thresholds in the number of agents and observable market depth. Furthermore, we demonstrate that market impact under agentic liquidity provision deviates from classical square-root dynamics, exhibiting distinct dissipative, balanced, and non-dissipative regimes under non-linear feedback loops.

General-purpose agent memory summarizes conversations: it extracts salient snippets, embeds them, and retrieves the top-k into the prompt. A health agent cannot run on summaries: a dose becomes a sentence, "since last week" is resolved at the model's discretion, and a three-month glucose trend cannot be answered by text similarity. We present PIA, a personal intelligence agent deployed alongside a consumer health agent. PIA receives the agent's natural-language requests, decides for itself whether and how to write or read, and turns conversations into typed clinical records and records into a synthesized understanding of the user. Its memory harness consists of four controls -- extraction, memory, retrieval, and understanding -- each a domain-agnostic mechanism with a pluggable health module: schema, medical alias dictionary, knowledge graph, and temporal rules. We show how the same query receives a different answer as the memory injected into the response context deepens from one-dimensional recall, to a two-dimensional health snapshot, to a three-dimensional trajectory with causality, and report lessons from operation: self-reported health data are missing not at random, question phrasing governs the quality of synthesized understanding, and nearly a third of candidate causal links are structural noise that rules alone remove.

We report fine-grained, deterministic instability in Dynamic Tensor Rematerialization (DTR), an online eviction policy for memory-constrained DNN training, measured on the reference DTR simulator (simrd) using public execution traces. On an LSTM trace, memory budgets differing by 0.10% of unconstrained peak memory select fast and slow execution regimes whose overheads differ by as much as 7.3x; the slow regime is driven by broadly repeated re-eviction of the same storages (evictions per storage rise from 1.33 to 8.27 while the set of distinct evicted storages is essentially unchanged: 5,233 vs 5,236, with the two sets overlapping at Jaccard 0.999). On a ResNet-32 trace, a fine budget sweep reveals a deterministic feasibility inversion: the run is feasible at ratio 0.101, infeasible (OOM) across 0.102-0.106, and feasible again from 0.107. We trace the immediate cause of the OOM to a fully pinned recursive rematerialization frontier that exceeds the budget after every evictable tensor has been evicted. Ablations using the DTR authors' own variants implicate the joint size-staleness scoring term in the observed LSTM instability. We argue these are at least two distinct budget-sensitive pathologies rather than one mechanism, and we separate what is demonstrated from what remains hypothesised. All results concern the reference simulator; reproduction in a production runtime is future work. Code, instrumentation, and raw results accompany this preprint.

Novel view synthesis (NVS) models can produce realistic new views of the same scene from different viewpoints. However, these generated views are not always geometrically consistent with one another. Multi-view (MV) consistency has shown promise as a tool for evaluating these NVS models. Its potential for multimedia forensics, however, remains largely unexplored, particularly for localizing geometric inconsistencies across wide-baseline image pairs. To enable research in this direction, we introduce DeformView, a wide-baseline MV dataset with pixel-level annotations of geometric inconsistencies. Using DeformView, we evaluate state-of-the-art MV consistency-scoring methods and show that approaches developed for NVS evaluation transfer poorly to the forensic task of geometric inconsistency localization. To address this limitation, we propose DEFECt3R, a lightweight learning-based classifier that uses cross-view feature relationships to localize geometric inconsistencies at the pixel level. By learning from explicit supervision, including hard negatives from geometrically consistent yet deformed views, DEFECt3R improves localization performance and substantially reduces false positives compared to existing consistency-scoring methods. Ablation experiments further show that both feature representations and correspondence quality contribute to localization performance. Overall, our findings demonstrate that MV geometric consistency is a promising yet underexplored signal for multimedia forensics and establish a benchmark and baseline for geometric inconsistency localization in wide-baseline MV image pairs. Code and dataset are available at https://github.com/IDLabMedia/DeformView-DEFECt3R

Individuals turn to large language models (LLMs) for guidance across a wide range of economic tasks, from comparing loan options and planning savings to deciding what raise to ask for or how much to charge for their services. LLMs are known to reproduce social biases, and biased economic guidance may influence what users believe they are worth, what they ask for, and what they ultimately accept. This risk is especially salient in India, where economic outcomes are shaped by demographic categories such as caste and urban-rural location. Existing LLM bias benchmarks, however, are largely designed around Western demographic categories and therefore miss key axes of economic disparity in the Indian context. We introduce RupeeBias, a benchmark for auditing demographic bias in LLM-generated economic guidance across Indian economic settings. RupeeBias consists of 39,150 prompts spanning four use cases: salary estimation, salary increment estimation, counter-offer recommendation, and service pricing recommendation. The benchmark follows a single-attribute counterfactual design, holding the description of the user's qualifications, experience, or service offering fixed while varying one demographic identifier at a time. RupeeBias covers 87 India-specific demographic identifiers across six axes: caste, religion, regional identity, gender, disability, and urban-rural location, with all prompts constructed in both English and Hinglish. We evaluate nine LLMs on RupeeBias and find systematic demographic disparities across all six axes. For otherwise identical prompts that differ only in demographic identifier, LLM-generated economic outputs differ by 20.2% on average. We publicly release RupeeBias to support future research on demographic bias in LLM-generated economic guidance across India-specific demographic and economic contexts.

Artificial neural networks (ANNs) usually represent neural transmission with fixed trainable weights during a training batch, which omits short-term changes in synaptic efficacy. In addition, the discrete time-step simulation requires additional temporal processing that many conventional ANN architectures do not use. To overcome these challenges, we propose Purin, a biology-inspired and ANN-compatible mechanism, that introduces synaptic efficacy modulation into conventional convolutional neural networks. Purin uses a time-interval-based abstraction for neural activities, which allows Purin to introduce short- and long-term synaptic efficacy changes without using discrete time-steps. Purin introduces a bounded factor to represent temporary synaptic efficacy changes, together with two weight matrices that represent input-side and output-side efficacy. The weight matrices are updated by backpropagation and interpreted as the long-term synaptic efficacy changes. Experimental results show that after removing the confounding factors in the AlexNet, VGG11, and GoogLeNet architectures, Purin improves the classification accuracies in all three models across the evaluated datasets.

Build pipelines are integral to modern software development, yet their energy footprint remains largely invisible to practitioners. Existing CI energy tools either rely on model-based estimation (due to hardware access restrictions in cloud runners) or report only total pipeline energy without decomposing it into meaningful phases. Joule-Profiler is an open-source command-line tool for Linux that measures hardware energy consumption via Intel RAPL (CPU), NVML (NVIDIA GPU), and attributes it to user-defined program phases by monitoring standard output. In this tool paper, we demonstrate Joule-Profiler in the context of Maven build pipelines using Google Gson as a case study. By applying a token pattern-matching Maven plugin invocations, we analyze builds across the 5 most recent Gson releases into per-phase energy profiles, comparing cold builds (empty local repository) and warm builds (cached dependencies).

Full-duplex speech language models continuously accumulate acoustic key-value (KV) states, making long-running interactions memory-intensive. During listening, the model can finish processing an audio unit before the next arrives; we term the remaining interval listening-time slack. We propose acoustic-to-text KV compression, which introduces a transcription side channel to convert incoming speech into compact textual memory within this interval. When the cache exceeds a target budget during inference, older acoustic states are evicted while transcripts and recent acoustic context remain. We train the side channel with LoRA using cross-entropy on transcription segments. To preserve listening and speaking behavior, we apply knowledge distillation to the original model's token-level output distributions at native prediction positions. On ten-minute LongSpeech sessions, our MiniCPM-o 4.5 implementation reduces peak streaming KV-cache size by 64.6% compared with the same model without eviction. The proposed method also improves transcription, temporal question answering, and summarization over the baseline. Full-Duplex-Bench evaluations further show comparable pause-handling, turn-taking, and interruption performance.

Pocket-conditioned molecular diffusion updates ambient atom coordinates, but many lead-optimization objectives are expressed on quotient features such as distances, contacts, and anchored substructures. We introduce budgeted quotient-residual guidance (QRG), an inference-time correction that makes these quotient objectives active without retraining the molecular generator. QRG lifts quotient covectors to metric-horizontal ambient directions and delivers them through a trust budget set by the frozen sampler's own step norm: quotient geometry chooses the direction, while sampler motion bounds the scale. We derive the horizontal lift, closed-form sampler-budget update, KL/kinetic interpretation around a frozen reverse step, equivariance conditions, and a product-budget split for budget-capped section and residual controls. Controlled quotient tasks confirm that sampler-relative delivery activates signals that raw local quotient gradients leave dormant. On frozen TargetDiff backbones, official seed-0 CBGBench ligand-generation/editing sweeps show practical quality-runtime gains: Local-QRG improves validity from 0.815 to 0.864 on fragment growing, 0.664 to 0.707 on scaffold hopping, and 0.681 to 0.712 on linker design, while PredNext-QRG improves fragment/scaffold and remains near-neutral on linker. Novelty remains 1.000 and diversity is preserved in the matched multi-seed molecular slice, giving task-dependent improvements without sampler retraining or backbone modification. Overall, QRG provides a lightweight route to quotient-aware inference for frozen molecular samplers with explicit runtime accounting.

Large language models (LLMs) as a judge enable scalable evaluation, but their judgments can be sensitive to response order and, even after removing such position effects, can still diverge systematically from human preferences.We introduce DIAL, a unified framework that combines abundant LLM comparisons with limited human comparisons to separate judge-specific position effects, learn shared structure in position-debiased LLM preferences, and adaptively calibrate that structure toward the human preference target. Theoretically, we study three aspects of DIAL: (i) identification of latent LLM preferences, position effects, and human calibration; (ii) adaptive estimation that balances LLM anchoring against limited human evidence; and (iii) fixed-weight uncertainty quantification for the calibrated human preference. Empirically, we evaluate position debiasing and human alignment separately in controlled simulations and on three human-preference benchmarks, showing that DIAL remains robust to unbalanced response order, achieves strong human-aligned rankings with limited labels, and adapts toward human evidence when LLM information is imperfect. Our real-data study collects over 410K judgments from 21 LLM judges in both display orders, providing a resource for future studies of LLM-judge bias, heterogeneity, and human alignment.

Estimating the influence of training examples on model behavior is essential for data debugging, valuation, and attribution. Existing influence estimators often produce incompatible rankings, which are commonly ascribed to approximation error. We argue that a more fundamental source of disagreement is specification mismatch: influence depends on the behavior being attributed, the intervention applied to each training example, and the counterfactual training process that maps the intervention to a model response. These choices are especially important when the target behavior requires a tractable surrogate, such as query loss, a logit, or a margin. We formalize influence as a counterfactual estimand, distinguish specification mismatch across estimands from approximation error in estimating a fixed estimand, and organize representative estimators by their implied specifications. We further derive a local decomposition that exposes how behavior signals, training signals, and counterfactual parameter responses interact. Controlled experiments show that exact estimands under different specifications can induce different rankings, whereas approximation error grows as perturbations move farther from their linearization points. Experiments on noisy label detection and LLM attribution show that specification choices significantly affect attribution quality, especially for the choice of behavior surrogate. Behavior-aligned specifications can identify target-specific training examples obscured by default loss-based or similarity-based specifications. These results establish specification analysis as a necessary first step for interpreting and comparing data influence estimators.

Auroral spectrographs such as the Auroral Spectrograph In Skibotn (ASIS) record hundreds of thousands of emission spectra, but only a few hundred can be labelled by an expert. To exploit the rest, we pretrain a 1D Vision Transformer with a masked autoencoder on 223,000 unlabelled spectra. Without labels, its representation recovers the emission-line intensity ratios that physicists use to diagnose the precipitating particles (R^2 0.91 vs. 0.77 for an untrained control) and, under one linear probe, classifies as well as 13 features designed by experts. Fine-tuned, the model outperforms the previous supervised auroral classifier on its own benchmark (macro-AP 88.5 vs. 77.8), reaches 0.870 mAP, and exceeds the same architecture trained from scratch by +0.159 with 10% of the labels; attribution shows that it uses both N2+ bands. Could an existing pretrained model replace it? Two astronomical spectral foundation models and a time-series model transfer according to their spectral window: SpectraFM, trained in the infrared, falls below the untrained control, whereas SpecFormer, trained in the optical, approaches in-domain pretraining without reaching it.

Scaling studies typically represent training data by a single count of samples. For hierarchically and spatially structured data, however, the same number of samples can be drawn from few or many sources and distributed differently across the underlying domain. We therefore study data scaling as an allocation problem, separating unique sample count, source diversity, and spatial coverage. We study this decomposition in microscopic whole-brain histology, where a source is an individual brain, and a sample is an image patch at a specific spatial location. Across 93 controlled pretraining runs of a contrastive model that uses spatial proximity for supervision, we vary data allocation, compute, and model capacity over 11.6 million spatially anchored image patches from 21 human brains. Performance improves with more unique samples, broader spatial coverage, additional compute, and larger model capacity. At fixed sample count, distributing samples across one to 18 subjects produces no detectable improvement, even though representations generalize substantially better to subjects encountered during pretraining. Inter-subject variation therefore strongly affects generalization, but additional subjects provide no benefit when a fixed sample budget is distributed across more sources. These results establish sample count, source diversity, and spatial coverage as distinct axes of data scaling in spatially structured representation learning.

Machine learning for scientific discovery is almost systematically data bound. Producing relevant high quality data, under budget constraints, is amongst the most promising ways to advance the field. Active learning (AL) offers promise wherever labelling requires expensive experiment, measurement, or simulation. Most existing tools cover only part of the data acquisition loop, and typically focus on either offline benchmarking or online deployment, but not both. We present ALF, a modular AL Framework that runs the full data acquisition loop via five modular components. One clear API for both settings: offline, against an existing dataset for controlled and reproducible experimentation; and online, against an oracle for acquiring new candidates in real-world deployments. ALF is open-source and available at https://github.com/instadeepai/alf.

Data quality research has usually treated data as an input that is stored, processed, and validated. In AI-driven software-intensive systems, data also shapes model behavior, evaluation, and lawful use. Empirical evidence remains limited on how practitioners define, assess, and manage quality under these conditions. We interviewed 16 practitioners from nine organizations and analyzed the transcripts using reflexive thematic analysis and developed six themes from participants' accounts. In AI systems, traceability shifted from modular debugging to attributing model behavior, while using models as quality assessors introduced circularity. Agent context and memory became data objects, and synthetic and pseudo-labeled data made authenticity a quality concern. In foundation-model development, lawfulness became a gate for training data, while representativeness was judged through coverage of situations in which the system must behave safely. Prior ML research examines many of these problems separately. Our study provides a practitioner-grounded account of how they are encountered together as an engineering and organizational concern. We also interpret five recurring conditions as helping explain how the themes relate to reduced trust in data and AI outcomes. We synthesize these findings through lifecycle assurance: a conceptual framing focused on producing evidence that data can support a specific AI claim when its influence may be embedded in model behavior, model-based judgments, or agent actions.

Artificial intelligence is advancing rapidly, with increasingly capable systems taking larger roles in reasoning, decision-making, scientific discovery, and autonomous development. As AI begins to participate in its own improvement, from model training and experience accumulation to agent evolution and automated AI development, the prospect of recursive self-improvement (RSI) is becoming increasingly relevant. This transition raises a fundamental safety question: how can safety be maintained when the system, its accumulated experience, and even the process producing its successors continue to change? We introduce Evolutionary Safety as a perspective for studying safety under persistent and recursive self-improvement. It concerns not only whether an AI system is safe at a particular moment, but how safety properties change, persist, accumulate, and propagate throughout evolution. We characterize recurring manifestations, including intent drift, error accumulation, experience contamination, safety-property erosion, evaluator drift, and risk propagation. We then develop a taxonomy spanning persistent agent state, model state, evaluation and environmental feedback, computational substrate, and meta-level update mechanisms. Building on this taxonomy, we examine how evolutionary risks can be discovered and evaluated across states, updates, trajectories, and lineages, and derive governance principles for modification, selection, authorization, provenance, and recovery. Finally, we outline open problems toward maintaining safety guarantees as AI systems become increasingly persistent, adaptive, and recursively self-improving. Project resources and proposed evaluation systems are available at https://chaunceykung.github.io/evolutionary-safety-rsi.

LLM-as-a-judge has become the de facto standard for scalable, subjective evaluation, yet current leaderboards compensate for systematic measurement bias by running ever more comparisons, an approach that is both statistically unsound and computationally wasteful. The root cause is an incomplete measurement model, treating LLM judges as neutral, interchangeable instruments ignores documented biases like position bias, verbosity bias, judge severity, and self-enhancement, that no volume of additional data can eliminate. We propose a unified latent variable framework that jointly models pairwise and ordinal data while explicitly correcting for these confounders, recovering reliable rankings from substantially fewer comparisons. Because fitting this model costs negligible compute relative to a single round of LLM inference, bias correction is not only more statistically rigorous but also a more sustainable approach to trustworthy evaluation.

Black-box model identification works by scoring a model's response to natural-language prompts. One line of work feeds models a degenerate input -- their own token, repeated -- to find a failure mode rather than an identity. We take that input and ask where the model goes when it does not. For each token t, read argmax p(. | t, t) in one forward pass; the result is a map on the whole vocabulary, with two halves. The first -- which tokens are fixed points -- is partially anticipated, and we report it as a failed estimand: the natural distance on it is 83% cardinality, separates a corpus manipulation by two bits in 3471 against a precision floor of zero, and attributes families at 0.5833. The second half, where the map sends tokens that are not fixed points, is unrecorded; the one paper holding those tokens logged them as a zero. Pairing on the source token removes the cardinality confound by construction (r from 0.9128 to -0.0932) and attributes families at 0.8333 -- twelve models scored against a pool of nineteen -- with chance 0.1389, across seven tokenizer groups and several corpora. Two nulls clear it: frequency-matched destinations agree at 0.1429, independent marginals at 0.0798. Family predicts agreement better than tokenizer (0.2031 against 0.1205), and recurrent architectures cluster at balanced accuracy 1.0 against a 0.7895 majority rate, or 0.90 once each model's dominant destination is excluded -- the figure we stand behind. We measure the robustness envelope: 8-bit weight rounding moves the map less than deduplicating the training corpus does (0.9004 against 0.6353, on one support), 4-bit destroys it (0.0098; 0.1812 at deployment granularity, so not a coarseness artefact), and the precision floor varies by model from 0.201 to 0.9778. All estimands and kill conditions were registered before the data, and the failed one is reported as fully as the surviving one.

Decoding and interpreting naturalistic speech from the brain increasingly relies on alignment to pretrained speech and language representation spaces. However, current CLIP-style brain-speech alignment ground neural activity to a single anchor modality-audio or text-despite the brain's inherently multimodal speech processing. This induces a trade-off: audio anchoring preserves temporal structure but weakens linguistic separability, while text anchoring captures semantics yet discards acoustic detail. We propose BAT-CLIP, the first CLIP-style trimodal alignment framework for iEEG that jointly aligns neural embeddings to both pretrained audio and text anchors in a shared, frozen audio-text manifold. On the naturalistic Podcast benchmark, BAT-CLIP yields more robust representations than bimodal CLIP baselines. We also highlight the importance of using self-supervised foundation models for CLIP training.

Large neighborhood search (LNS) relies critically on destroy and repair operators, whose effectiveness depends on both adaptation to the evolving LNS state and interaction between the two roles. We introduce Stackelberg Program Optimization (SPO), an LLM-based framework for discovering adaptive executable destroy-repair programs. SPO conditions operator decisions on a compact LNS state, allowing state-dependent behavior to emerge through program discovery, and organizes destroy-repair discovery as a Stackelberg interaction over program space that reflects their asymmetric dependency. Role-specific credits evaluate destroy programs as leaders and repair programs as conditional follower responses, guiding a coupled optimization process that combines LLM generator learning with population-based evolutionary search over programs. Experiments on the traveling salesperson problem and capacitated vehicle routing problem show that SPO outperforms strong baselines across a broad range of settings and generalizes beyond the discovery scale to larger instances and benchmark sets. Behavioral analyses further demonstrate state-dependent operator behavior and coupled destroy-repair improvement during discovery.

Repository-level issue localization aims to identify and rank the files and functions relevant to resolving a reported issue. LLM agents approach this task iteratively: they identify a set of potentially relevant locations, inspect the corresponding code, and revise their judgments about these candidates as new evidence is acquired. Existing environments, however, provide limited support for this loop: agents must search for unresolved relation targets, reconstruct entity semantics from raw source code, and revise candidates without evidential basis. To address these limitations, we present SemNav, a framework that leverages deterministic retrieval to seed a broad candidate set and an LLM agent to continually refine that set, thereby combining initial coverage with evidence-guided revision. SemNav supports this process through three key components. A Semantic Navigation Graph resolves program relations on demand through a language server, enabling direct navigation to related entities across files. Issue-conditioned Semantic Cards provide compact, source-grounded interpretations of each entity's role and relevance to the issue. A persistent Candidate Workspace records each candidate together with its evidential basis, enabling grounded verification, revision, and ranking. Across SWE-bench Lite and PLocBench, SemNav outperforms existing baselines, improving File Hit@10 from 68.33\% to 82.67\% with Gemma 4B. Component ablations and trajectory analysis support the complementary roles of all three components, while Semantic Cards reduce working-context load by 48.2\% relative to full-source reading. SemNav further ranks first on all seven evidence-quality metrics on SWE-Explore and improves downstream issue resolution from 44.00\% to 52.33\%.

Multimodal Machine Translation aims to incorporate additional signal from non-textual modalities to improve translations by resolving ambiguities. While models, through multimodal fusion, are able to accept images related to the source text, they can ignore this information. Therefore, increasing their visual sensitivity remains an active research area. In this work, we introduce a training method, Metric-based Loss Weighting, that improves visual grounding of translations by increasing the loss function for tokens that benefit from the accompanying image. We identify these tokens using the Point-wise Cross-mutual Information (PCXMI) metric, which compares the model's output probabilities with and without visual context. We introduce a Congruency-based PCXMI metric and experimentally show that both metrics working in combination yield the best results. We evaluate our method by fine-tuning three pretrained Multimodal Large Language Models on the task of Image-guided Machine Translation for three language directions. Metric-based Loss Weighting outperforms other tested methods on the CoMMuTE contrastive dataset, improving accuracy by up to more than 7 percentage points compared to standard fine-tuning, while maintaining strong general translation performance.

My doctoral work aims to explore Audio Emotion Recognition (AER) in the context of atypical listening. This research focuses on auditory hypersensitivity in people with autism, a phenomenon that is often difficult to evaluate and unique to each individual. Our core idea is to leverage our understanding of affect from acoustic traits, relying on the possibility of generalizing affective responses from a small amount of annotated data. As a first step, we fine-tune a large foundation model, Contrastive Language-Audio Pretraining (CLAP) using low-rank adaptation (LoRA), trained on a valence and arousal dataset of neurotypical listeners.

EEG foundation models increasingly use masked prediction to learn from unlabeled recordings, but optimizing this objective does not ensure transferable neural representations. A central challenge is that stable positional cues and local correlations can make masked regions predictable without integrating distributed neural context. To reduce this reliance on low-information prediction paths, we introduce Neural State Prediction (NSP), a latent-predictive framework that constrains both the prediction target and the available context. NSP uses a Target Encoder updated by an exponential moving average (EMA) to define latent supervision. Identity residualization removes additive effects associated with channel identity and relative time from the targets, while topology-separated context excludes their immediate spatial and temporal neighborhood from the visible input. We pretrain NSP on 2.2 million EEG segments from TUEG and evaluate it across 30 downstream datasets spanning clinical diagnosis, sleep staging, emotion recognition, motor imagery, event-related potentials, cognitive-state decoding, and language retrieval. Under full-parameter multi-task fine-tuning on EEG-FM-Bench, NSP achieves 63.94 macro balanced accuracy across 14 datasets, exceeding the strongest evaluated baseline by 2.35 percentage points. Controlled component ablations assess the contribution of each mechanism, while matched context controls and held-out interventions characterize the role of context geometry, signal content, and positional information. Jointly designing latent targets and their context offers a promising direction for EEG foundation models that learn from distributed signal structure.

Recommender systems have traditionally been developed for platforms. However, this has given rise to many phenomena that may be advantageous for platform lock-in but are a nuisance to users, such as clickbait, filter bubbles, and the spread of fake news. Recently, user-side recommender systems have been proposed as a new paradigm for solving this problem. If users deploy their own recommender systems, they are no longer at the mercy of the platform's interests. However, building a user-side recommender system is not trivial; in particular, customizing one for oneself requires additional data. We propose AgentRecommender, a method that leverages the investigation capability and internal knowledge of LLM agents to flexibly build user-side recommender systems without additional data. AgentRecommender allows users to easily create recommender systems tailored to their own preferences.

Speech-breath segmentation is a fundamental preprocessing step in respiratory audio analysis, enabling applications such as respiratory acoustic biomarker extraction, lung function prediction and disease monitoring. Existing approaches, including threshold methods, Fourier Transform-based techniques, and unsupervised and pretrained voice activity detection (VAD) models, primarily focus on speech detection and often classify breathing events as non-speech or silence, limiting their applicability for precise breath detection. To address this limitation, we propose BreathGRU, a semi-supervised Bidirectional Gated Recurrent Unit (BiGRU) framework specifically designed for speech-breath segmentation. The proposed framework combines frame-level acoustic feature extraction with bidirectional recurrent modelling, pseudo-label refinement and duration-constrained Segmental Viterbi decoding to produce speech and breath segmentation. BreathGRU was evaluated against the existing approaches, using manually annotated recordings. Performance was assessed using event-based, time-based, overlap-based, duration-based and boundary-based segmentation metrics. Experiment results demonstrated that BreathGRU achieved the highest breath event recall (0.83), the lowest onset-localisation error (0.14s) and the highest Mean Match Intersection over Union (0.81), with competitive overall segmentation performance compared to large pretrained VAD models like Silero. Qualitative evaluation on manually annotated recordings further showed close agreement between BreathGRU and manual annotation, with better breath detection compared to Silero. These findings demonstrate that explicit breath event modelling provides advantages over general-purpose VAD models and establish BreathGRU as an effective speech-breath segmentation framework which can be applied for respiratory audio analysis and pulmonary healthcare applications.

Recommender systems engineer serendipity to foster active exploration and break predictable consumption cycles. The problem with existing offline beyond-accuracy metrics is that they often either isolate historical similarity or global popularity. We aim to design an evaluation metric that examines similarity, popularity, and actual user relevance. To achieve this, we introduce SPADE (Serendipitous Pareto Distance Evaluation). SPADE maps all items into a two-dimensional space to directly calculate a user-specific Pareto frontier of maximally popular and historically similar items. The final serendipity score is then computed by averaging the minimum Euclidean distance from this boundary strictly for the correctly recommended test-set items. Evaluating SPADE across five datasets and five baseline algorithms confirms its effectiveness; our results show that the metric successfully prevents algorithms from exploiting beyond-accuracy measures with irrelevant or non-personalized recommendations, reliably isolating serendipitous discoveries.

Time series forecasting is typically framed as learning a direct mapping from historical to future observations in the observation space. However, sequences of observations generally provide only a partial view of the dynamics of the underlying system, with future observations being shaped by latent dynamics. Recent latent-space forecasting methods thus achieve improved performance by predicting future observations from latent-space representations of historical observations rather than directly forecasting future observations in the observation space. Next, while future observations are also shaped by external factors, how to incorporate external, often multimodal, information into forecasting, so that it can shape latent-state formation and evolution directly, remains underexplored. We propose WorldTS, a world-modeling based forecasting framework that integrates multimodal covariates directly into the forecasting to further improve forecasting performance. Specifically, WorldTS employs a two-stage training strategy. First, it learns forecasting-relevant latent state dynamics conditioned on multimodal covariates, yielding encoded future states. Next, the learned state dynamics are frozen, and an observation decoder is trained to map the predicted future states back to future observations. Extensive experiments on 21 real-world datasets offer insight into WorldTS and its effectiveness.

Recent empirical and theoretical advances suggest that joint-embedding predictive architectures (JEPAs) may learn meaningful representations for action-conditioned prediction of future outcomes, thus becoming one of the foundational structures for world models. However, accurate prediction does not, in general, necessarily imply recovery of underlying causal states that give rise to the observed dynamics. This work investigates when and how JEPAs can recover the underlying causal states from observations. We first introduce a latent variable model, in which high-dimensional observations are generated from latent causal states whose dynamics are governed by action-conditioned transition mechanisms. Based on this formulation, we develop a general information-theoretic objective that combines conditional likelihood maximization for learning transition dynamics with entropy maximization for preserving latent state information. We then establish identifiability conditions under which representations learned by this general objective recover the underlying latent causal states up to component-wise invertible transformations and permutation. One key condition for such identifiability is sufficient action-induced variation in the transition mechanisms. Guided by this finding, we instantiate the general objective with an action-modulated Gaussian additive-noise model, yielding action-modulated JEPA (A-JEPA). Experiments on synthetic environments verify the theoretical findings under the identifiability conditions and robustness to moderate violations, while visual benchmarks demonstrate improved state recovery and transfer to unseen transition mechanisms.

Vessel segmentation in medical images is essential for many clinical tasks, ranging from diagnosis to treatment planning. However, it remains challenging due to complex vascular morphology and diverse imaging conditions. Existing deep learning methods rarely aim at building a generalizable vessel segmentor across anatomies and modalities. While the Seg- ment Anything Model (SAM) has shown promise for med- ical image segmentation, its original design does not fully exploit vascular morphology and struggles with fine-grained vascular structures, leading to suboptimal performance. In this paper, we propose ReG-SAM, a SAM-based framework tailored to 2D vessel segmentation that leverages reference graph set for enhancing vascular representations. Specifically, we introduce two modality-aware representations derived from the reference masks: graph prompt embeddings (GPEs) that encode global spatial features from graphs, and vascu- lar prototype embeddings (VPEs) that capture fine-grained modality-specific vessel characteristics from multi-scale fea- ture maps and vascular masks. Since both require vascular masks that are unavailable during inference and require robust modality-aware vascular feature representations, we construct a modality-wise vascular database and develop two reference graph-guided representation learning schemes for estimating GPEs and VPEs using samples from the database rather than ground-truth masks. Extensive experiments across 19 datasets demonstrate that ReG-SAM consistently outperforms existing baselines, even those using manual prompts, particularly on challenging thin vessels.

We propose a momentum-guided federated split distillation framework for personalized, efficient, and autonomous temporal edge intelligence. We introduce TeRR-SAtt, our novel temporal reservoir student attention design that combines fixed reservoir representations, a lightweight temporal student, and personalized output modules. We also present AMGF, our anticipatory momentum-guided fusion mechanism that clusters clients through learning momentum and derives specialized teacher updates. On real-world smart-building data, TeRR-SAtt reduces edge training latency by 65.50%, inference latency by 44.70%, training memory usage by 18.40%, and inference CPU usage by 33.10% over the considered baselines. At the same time, AMGF improves local learning by up to 35.31% in RMSE compared to global updates.

Multi-dimensional time series, inherently tensorial, are common in practice. Despite great progress in time series anomaly detection, most existing methods are confined to uni-/multi-variate time series. When handling multi-dimensional time series using these methods, reshaping operations are required, which inevitably break the intrinsic correlations and thus lead to performance degradation. In uni-/multi-variate time series anomaly detection, AutoEncoders (AEs) are widely adopted and generally categorized into reconstruction-based and prediction-based AEs. The reconstruction-based AE utilizes the current observation for reconstruction, while the prediction-based AE utilizes the historical information to predict the current observation. Thus, the two AEs utilize different information. To bridge the gap between reconstruction-based and prediction-based AEs, so as to fully leverage the available information and thus further enhance performance, we propose a predictive prior and incorporate it into the reconstruction-based AE. It may not be very difficult to conceive this idea, but designing the predictive prior so that it can work for tensor anomaly detection is non-trivial. Specifically, to avoid breaking the intrinsic correlations within the multi-dimensional time series, we use the tensor AE as the backbone. To incorporate the predictive prior into the reconstruction-based AE, we propose a Bayesian fusion approach and our analysis reveals that this approach can enhance the modeling capability of the model for normal data. To mitigate the over-generalization problem of AE, we incorporate physical laws, i.e. tensor low-rank decomposition rules, into the neural networks in the predictive prior, leading to the Physics-informed Predictive Prior Tensor AE (PPPTAE) framework. Experimental results on real-world datasets demonstrate the effectiveness of the proposed method.

Compressing a trained model yields a family of deployment candidates, and under domain shift the most compressed one need not be the one to deploy. We study selection over such a family, with candidates and teacher fixed and target labels absent or scarce. Two findings organize the label-free case. Minimum teacher distortion behaves almost as a constant rule, selecting the same eight-bit, per-channel, unclipped configuration in every run, which does not minimize empirical target cross-entropy. Established estimators divide sharply: in the overconfident-collapse regime of the CNN families, confidence-based estimators order the family close to backwards, and the diagnostics that identify it need the labels the setting denies, while output-distribution estimators match the teacher-relative anchor and on one architecture beat it. Distortion is nonetheless stable, so a supervised term can move selection away from it. Combining the two, we give exact quadratic identities for a canonical quadratic analogue of the family. We also show that under symmetric corruption the label-dependent part of a criterion linear in the label indicator is multiplied by one common factor whenever its coefficient sums are candidate-invariant, a class holding teacher contrasts and accuracy but not cross-entropy. These characterize the score's components without bounding selection regret. Across one hundred and thirty-four candidate families, one per independently trained convolutional or Vision Transformer teacher, anchoring reduces mean regret at the smallest label budget in every setting, an advantage that fades beyond twenty-five labels.

Cross-site lung histopathology classification must account for stain variation, non-IID client data, missing classes, and the cost of adapting large pathology encoders. This study evaluates FedHisto-PAST v2 for three-way classification of adenocarcinoma (ACA), Normal, and squamous cell carcinoma (SCC). FedHisto-PAST v2 combines a frozen HIBOU-B foundation model with parameter-efficient adaptation, stain-conditioned paired-view prediction and feature consistency, reliability-aware prototype learning, and adaptive federated aggregation. Experiments used a five-client, non-IID, raw-data-local simulation with fixed internal evaluation, client-level analysis, component ablations, communication accounting, and a development-influenced exploratory LungHist700 cohort. All principal methods achieved near- ceiling internal performance, which limited discrimination on the fixed split. On LungHist700, FedHisto- PAST v2 achieved a Macro-F1 of 0.728560 and a balanced accuracy of 0.730454. Higher recognition of Normal and SCC was accompanied by lower ACA recall, and calibration remained imperfect. Prediction-level consistency was the only component with a clearly supported independent contribution in the external ablation analysis. Feature consistency and prototype regularization showed no conclusive independent overall gains in Macro-F1. The framework updated 1.253841% of the model parameters. The results provide exploratory cross-dataset evidence for stain-aware, parameter-efficient federation; they do not establish formal privacy, patient-level independence, prospective deployment, or clinical validation.

Scientific measurements such as single-cell RNA (scRNA) sequencing often take the form of nonnegative integer counts, whereas continuous-state diffusion models approximate this discrete structure using continuous coordinates. Building on stochastic chemical reaction networks (CRNs), a class of count-native Markov jump processes, we introduce CRNDiff, a structured framework that combines count-space diffusion with inference-time conditioning on rare subpopulations. An independent birth--death instantiation yields a closed-form transition kernel for forward noising. This kernel enables reverse sampling via forward-filtering backward-sampling (FFBS) and supports data-driven selection of the terminal noising time, eliminating the need for a validation sweep. This tractability also lets us introduce tilted Feynman--Kac (FK) steering, a method for sampling target subpopulations from a frozen generator without retraining. By tilting posterior marginals before FK particle correction, steering mitigates importance-weight concentration when the target population is rare. Using scRNA-seq data from the human heart cell atlas, we test the ability of CRNDiff to generate cell-type-specific distributions. Across the three evaluated target populations, CRNDiff achieves the highest conditional fidelity among the evaluated generative models, with larger mean purity margins for rarer target populations. Generated cells preserve marker-level differential-expression structure. Replacing real training cells for the target classes with generated cells yields downstream classification performance approaching that of the real-data reference.

Models trained with reinforcement learning for calibrated decisions (RLCD), such as Jev, answer a typed question about an input, the state, with a probability, a choice, or a score, and software acts on the answer without a person reading it. Their robustness has not been measured: adversarial benchmarks score what a model generates or executes, whereas a typed model generates nothing and returns a well-formed answer even when manipulated. Measurement is also hard, because identical requests can return different answers, most available labels come from the model itself, and the API preprocesses each request out of view. Our key idea is to score each attacked decision against the model's own clean decision rather than against labels, and to read it against the change caused by an identical re-run. Building on this, we introduce JevAdvBench, to our knowledge the first adversarial benchmark for RLCD models, with 812 typed questions over 66 scenarios, and a black-box attack suite of 9,744 single-edit variants that each edit one part of a request, with billed input tokens confirming that the edit reached the model. On jev-1.13.0, rewording stays within 1.2 percentage points of the re-run baseline, and fields outside the schema never reach the model. In contrast, one unverified opinion appended to the state flips 12.1% of decisions, statistically tied with the strongest injected command (10.1%), and pushes 38% of confident answers below the 0.8 confidence threshold that routes them to human review. Applications built on RLCD models should therefore treat the state as untrusted, argued input. Project website: https://JevAdvBench.github.io/JevAdvBench/

Knowledge distillation is the standard way to compress encrypted-traffic classifiers for the edge, and almost all such work judges students by accuracy alone. We ask what else a student inherits: unknown-traffic detection, calibration, shortcut reliance, and whether any survives a year of drift. Resemblance proves little on its own, since soft targets also regularise. We therefore distil one 101k-parameter student from two teachers of equal accuracy but different construction, a five-member ensemble and a single wider model, so that following one rather than the other is attributable to it. The design was pre-registered before any test result was seen. We tested ten hypotheses on CESNET-TLS-Year22, a year of real TLS traffic, across 18 test windows over 35 weeks. Two are supported: a student's per-flow unknown-scores shift toward its own teacher, but only at a conventional temperature, not the accuracy-optimal one; and a shortcut-reliant teacher passes its over-confidence to a student that never sees the feature. The drift prediction is reversed under both scores, the gap narrowing rather than widening and the student overtaking under the energy score in two of three replicates, as is the prediction that such a teacher harms its student's detection, which improves slightly. Shortcut reliance is set by model size, not distillation. Under the logit-based scores nothing else transfers: distillation beats neither a temperature-scaled direct student nor label smoothing. Exploratory analysis shows this turns on the scoring rule: with a feature-space detector the teacher detects unknown traffic 0.073 AUROC better than the direct student, where the energy score sees 0.000, and the conventional-temperature student inherits most of it. Label smoothing, with no teacher, recovers more. Distillation transfers the teacher's habits; what looks like an inherited ability is available without one.

Most Vision-Language Models (VLMs) are built by extending pretrained Large Language Models (LLMs) with visual modules and multimodal alignment. However, this multimodal scaling often degrades the language-side reasoning ability originally encoded in the base LLM. While the base LLM retains usable reasoning after scaling, the aligned VLM itself cannot reliably access this ability. Therefore, recovering the degraded reasoning capability in VLMs would benefit more from seeking help from the base LLM than from the VLM alone. Motivated by this, we propose LIFT (Language-side reasonIng Facilitation and Transfer), a lightweight vector-intervention method that transfers reasoning capability from the base LLM to the VLM without retraining the backbone. LIFT defines Reasoning Vectors as answer-token hidden-state differences between a Reasoner path with an explicit reasoning trace and a Solver path without it, and injects these vectors into language-side activations of the target VLM. LIFT further supports learnable vector adaptation while keeping the VLM backbone frozen. We evaluate LIFT on two VLMs across six reasoning benchmarks, comparing Reasoning Vectors extracted from the base LLM and from the aligned VLM under matched protocols. Results show that LLM-derived vectors consistently outperform VLM-derived vectors, confirming that the base LLM is a more effective source for recovering reasoning. LIFT partially recovers degraded reasoning through lightweight language-side interventions. Further analyses show that Reasoning Vectors influence intermediate reasoning behavior rather than merely altering final answers. The source code will be released soon.

The growing availability of Artificial Intelligence (AI) tools creates new opportunities to support engineering design processes, yet their current use often remains limited to isolated tasks such as coding, documentation, or information retrieval. Less attention has been given to how AI can support cooperative engineering workflows at the process level, where teams must coordinate requirements, tasks, communication, knowledge transfer, and subsystem integration. This paper investigates this challenge in the context of the European Rover Challenge (ERC), where student teams design and integrate complex rover systems within a single academic cycle under strict time constraints and high subsystem interdependence. We conducted a role adaptive 40 question survey with ERC 2025 teams, yielding 104 responses from 14 teams. The survey examined team structure, knowledge transfer, task management, integration practices, communication patterns, and current AI usage. The results reveal recurring workflow bottlenecks, including limited documentation, unclear requirements, fragmented communication, informal task monitoring, and substantial integration rework. Based on these findings, we derive requirements for AI augmented cooperative engineering work-flows and propose an initial assistant system architecture that connects user facing interfaces, credential management, service selection, specialized AI services, and external engineering tools. The proposed architecture aims to support task clarification, requirement and compliance management, communication summarization, integration risk detection, and continuous knowledge capture. In doing so, the paper contributes empirical requirements and an architectural direction for AI augmented cooperative engineering workflows in hybrid human AI team settings.

Spatio-Temporal Video Grounding (STVG) aims to localize the spatio-temporal tube in a video corresponding to a natural language query. While recent methods achieve strong performance in fully supervised, weakly supervised, and zero-shot settings, they typically rely on computationally expensive architectures, complex training pipelines, or multimodal large language models. We present Pocket-STVG (P-STVG), a lightweight cascade architecture that addresses STVG by combining efficient pre-trained components instead of large end-to-end models. P-STVG integrates a temporal-aware video encoder based on MobileViCLIP, a spatial encoder-decoder derived from MDETR, and a shared aligned text encoder. Temporal localization is performed through either a lightweight 1D U-Net or a simple thresholding strategy, enabling the same framework to operate in both weakly supervised and zero-shot settings. Furthermore, video representations are precomputed independently of the query, yielding an indexing-friendly pipeline for efficient inference and large-scale video collections. Despite requiring fewer than 90M parameters, P-STVG performs on par with weakly supervised methods and improves on earlier zero-shot approaches at a fraction of their memory and computational cost, establishing a favorable performance-efficiency trade-off for STVG.

High-fidelity atomistic evolution over long timescales requires more than observing the current crystal configuration. Instantaneous atomistic snapshots are often incomplete: locally similar configurations can correspond to different hidden dynamical contexts, future event preferences, and waiting-time scales. We argue that this snapshot ambiguity makes long-horizon atomistic evolution fundamentally a memory-based world-state restoration problem. To address this, we introduce AtomWorld-Mem, a memory-restored atomistic world model that recovers the latent world state missing from instantaneous crystal snapshots. AtomWorld-Mem treats the evolving alloy as an AtomWorld: spatial encoders write multi-scale atomistic keyframes from dense local topology and sparse long-range defect context, while short-term event memory and long-term structural memory integrate these keyframes across time to restore a future-predictive evolutionary state. The restored state is used to prioritize legal vacancy-mediated events under single-event Kinetic Monte Carlo (KMC) constraints, while event legality, physical execution, and residence-time updates remain governed by the underlying simulator. Empirically, AtomWorld-Mem improves long-horizon atomistic progress under fixed microscopic event budgets while maintaining high-fidelity evolution across energetic, structural, and vacancy-transport observables. It further transfers zero-shot across diverse unseen alloy-temperature AtomWorlds, suggesting that the learned memory-restoration mechanism captures reusable principles of hidden-state inference rather than a system-specific local energy heuristic. These results position memory-restored world-state modeling as a promising route toward efficient, physically grounded, and transferable atomistic evolution.

We empirically investigate Question Under Discussion based modelling in naturalistic dialogue by studying whether the salience of generated potential questions predicts their subsequent resolution. Building on Wu et al. (2024), we construct a dataset of 7,124 questions automatically generated from utterances and preceding context from the British National Corpus, and annotated for salience and answerability. We find a robust but low positive correlation between salience and answerability in dialogue, indicating that more salient questions are more likely to be addressed. However, this effect is markedly weaker than in monologic text, suggesting that conversational structure is less predictable. We further observe that structured interactions exhibit stronger alignment between annotators than less organised dialogues.

Frequency-based adversarial attacks have recently grown popular by exploiting spectral sensitivities shared across neural architectures. Unlike spatial perturbations, frequency-based attacks expose deeper vulnerabilities, making them especially valuable for robust evaluation of safety-critical and security-sensitive applications. Yet, existing approaches are typically not derived as solutions to an optimization problem that explicitly captures transform-domain structure. In this paper, we propose a methodology for crafting principled frequency-based adversarial attacks, via a dedicated optimization framework. A cornerstone of our method hinges on the introduction of a perturbation constraint set, tied to highly structured non-orthogonal transforms, well-known for their flexible, non-predefined frequency handling. We prove that the attacks emerge as weighted $\ell_2$-projections onto this set, yielding a general and controlled attack generation mechanism. By this, we provide a clear geometric attack characterization, ensuring alignment between the optimization objective and the perturbation constraint. We assess our framework on standardized datasets, for pretrained and adversarially robust models. Results highlight that our attacks, being solutions to an optimization problem, over a structured perturbation set, are highly effective, even across different, unseen architectures. Our methodology could serve as a theoretical baseline for designing and analyzing transformed-based attacks, targeting fundamental model vulnerabilities, instead of mere architecture-specific artifacts typically studied in the robustness literature.

Activation steering is a powerful training-free paradigm for controlling large language models at inference time. However, standard approaches estimate a per-layer steering direction from contrastive data and apply it on the layer's entire representation space, which may couple the intervention to off-target properties present in the contrastive data and degrade unrelated capabilities. To mitigate this issue, we introduce LocUS (Localized Unembedding Steering), a method which grounds activation steering to the model's own output vocabulary subspace. By identifying a property-specific linear subspace within the unembedding matrix, LocUS enforces a geometric constraint that restricts the steering transformation to a specific subspace and at the same time localizes its application to a sparse subset of attention heads. Extensive evaluations across three model families on toxicity mitigation, sentiment redirection and sycophancy suppression show that LocUS matches or outperforms state-of-the-art baselines while intervening on under 6% of parameters and better preserving general capability.

Chain-of-thought (CoT) monitoring is a promising safety technique for reasoning models, enabling detection of problematic reasoning before models act. A key concern is encoded reasoning, where models hide their true reasoning in ways that monitors and humans cannot interpret. Optimization pressure from CoT monitors during reinforcement learning is considered a likely driver of such behavior. We investigate this by training reasoning models to perform a main task and a side task, while penalizing them when a monitor detects reasoning about the side task. Surprisingly, models learn to evade monitors without encoding their reasoning. Instead, they learn to phrase and format their chains of thought such that monitors fail to flag side task reasoning, while the reasoning remains completely transparent to human readers. We call this phenomenon monitor jailbreaking. We find that monitor jailbreaking arises across different model sizes, monitors, and tasks. Jailbreaks generalize to monitors not seen during training, including both less and more capable monitors, and transfer across different monitor prompts. While jailbreaking strategies appear simple, manually replicating them does not reliably fool monitors. Finally, we show that paraphrasing is an effective defense: paraphrasing a jailbroken CoT allows the same monitor to correctly flag it, while still allowing the model to perform both tasks.

Neural algorithmic reasoning aims to train neural networks to follow known algorithms and generalize beyond the input sizes seen during training. However, correct final outputs and intermediate supervision do not necessarily show that a model follows the intended execution. We study this problem using insertion sort. Our analysis of the CLRS30 baseline NAR shows that the hint objective is weakly optimized and that hint accuracy remains low. Moreover, many intermediate representations can already be decoded into sorted sequences before the reference insertion-sort execution terminates, suggesting that the model learns a shortcut to the final output. Motivated by these findings, we introduce Discrete Neural Insertion Sort. Our model represents the sequence as a chain, separates scalar exchanges from control-state transitions, and projects node representations back to discrete states after every processor step. When trained only on sequences of length 16, the model achieves $100\%$ sorted-sequence accuracy on sequences of length 64 and 128. However, an ablation shows that discretization and graph structure alone are insufficient: without additional supervision of the global inner-loop state, the model fails even at the training length. Our results show that discrete execution can support strong length generalization, while also highlighting the problem-specific inductive bias required to learn a faithful algorithmic execution.

We study Bayesian optimization (BO) through the lens of information geometry. Pulling back the Fisher information metric through the surrogate posterior map yields a local sensitivity tensor on the input space, which leads to an upper bound on the gradient of reparameterizable acquisition functions. This view explains vanishing-gradient behavior in high-dimensional BO and provides a common interpretation of heuristics such as RAASP and dimension-scaled lengthscales. Building on this analysis, we propose FITR, a trust-region-based BO method that replaces lengthscale-based scaling by local pullback-Fisher weights. FITR is not restricted to GP kernels with explicit lengthscales. On GP benchmarks with an SE kernel, experiments show competitive performance using FITR. The proposed method also easily generalizes to non-isotropic surrogates, although the gains are more task-dependent in that setting.

Spatial reasoning with metric constraints requires linking objects to geometric measurements and preserving their numerical content during language reasoning. We present DepthEvidence, a 4B model that uses its own dense metric predictions as object-grounded evidence for language generation. A camera-conditioned decoder predicts full-resolution metric depth using multi-scale visual features and high-resolution RGB refinement. A dense-to-language interface converts predicted depths and decoder features into object-aligned continuous geometry tokens anchored to object identifiers. Geometric supervision encourages metric information to remain recoverable before and after language-context interaction, while instruction tuning supports object measurement and compositional reasoning. We introduce a Depth-VQA benchmark evaluating object-depth queries, relative comparisons, and decisions combining spatial and numerical constraints. Across nine datasets, DepthEvidence achieves the highest average dense $δ_1$ among evaluated methods, competitive with specialized estimators. It also leads the evaluated methods in instance-level metric depth estimation and overall accuracy on both relative and metric reasoning tracks, while broadly preserving general VQA performance and improving spatial understanding relative to the base model.

The flatness of the loss landscape at a minimizer is a widely used heuristic for reasoning about neural-network generalization, yet evidence for this relation is mostly empirical and controversial. We study this relation in a teacher-student tree committee machine, where both the ERM estimator and the Hessian spectrum are analytically tractable in the proportional high-dimensional limit. First, we use a zero-temperature Gibbs formulation to obtain predictions for the observables of the typical minimizers of the empirical loss. Secondly, we use Edwards-Jones formalism to derive the limiting Hessian resolvent around these typical minimizers. All predictions agree with finite-size gradient-descent simulations. Finally, we study three measures of flatness, namely the left and right edges and the spectral mean, and check if a decrease in generalization error as the dataset size is increased corresponds to an increase in flatness. We find that the answer strongly depends on the learning task and on the ratio of the number of parameters to the number of data points. In regression, the spectral mean and right edge correlate with the generalization error, while the left edge does so only in the overparametrized regime. In classification this correlation reliably holds only in the highly overparametrized phase, while for underparametrized networks it can even reverse.

We introduce \emph{effective depth} ($\Deff$), a scalar diagnostic that treats the layer-wise residual stream of a transformer as a discrete-time process, measures how representation similarity decays with layer distance, and aggregates that profile into one number. Across sixteen decoder-only language models, $\Deff$ separates a structural consequence of residual accumulation from an empirical one: even maximally diverse orthogonal updates have the closed-form reference $F_L = 2L/(L+1)<2$, yet fifteen of sixteen default measurements lie below $F_L$ (Qwen3.5: 32--44\%, OLMo-2: 40--41\%, Pythia: 23--28\%). Matched references show that the gap is not caused by the persistent initial state or update-size imbalance, but is largely a calibrated signature of correlated residual updates rather than evidence that depth is unused. Symmetric position-0, token-normalisation, and top-PC controls show the regime is not reducible to BOS or top-PC artefacts: the lone above-reference default outlier joins the same regime, and all sixteen models are sub-reference after token-normalisation or top-1-PC removal. Intermediate checkpoints show that the regime is established early in OLMo-2 and stable through 5T tokens, while Pythia-1.4B follows a distinct decreasing trajectory. A controlled residual-carry intervention supports the mechanism, and $\Deff$ is best read as a \emph{global} accumulated-state diagnostic, not as a capability score or pruning method.

Scaling language models to long contexts is limited by the quadratic cost of self-attention. Block sparse attention offers an efficient alternative, but selecting the retained blocks remains a bottleneck. Conventional block selection requires scoring all query-block pairs and therefore remains quadratic in sequence length. To address this issue, we propose PISA, a block-sparse attention mechanism that employs a pyramid Top-$K$ selection strategy. The main idea is to gradually narrow down the candidates across different levels, making it more efficient to find the most relevant keys. Specifically, we construct a coarse-to-fine hierarchy of keys and perform selection from the coarsest level. At each level, LogSumExp scoring is applied to a bounded candidate set to select candidates for the next finer level, continuing until the finest level is reached. Through pooling, we construct $O(\log N)$ levels of keys, yielding an overall complexity of $O(N\log N)$, where $N$ denotes the sequence length. We develop hardware-aware Triton kernels for both training and inference, fusing hierarchical routing and LogSumExp scoring without materializing the query-key score matrix. We further evaluate our method on language modeling tasks. Compared with the baseline, our method achieves comparable performance on benchmarks such as commonsense reasoning while delivering better results on retrieval tasks.

Knowing where species occur is fundamental for biodiversity research and conservation. Species distribution models (SDMs) link species observations to environmental conditions to estimate their spatial distribution. However, accuracy varies with the underlying data and models, making it essential to know for which species models can be trusted. Deep-learning-based SDMs ("DeepSDMs") now jointly model thousands of species, drawing on hundreds of millions of community-science records. At this scale, averaging performance hides substantial species-level variability, particularly for rare species, often of greatest conservation concern. Records are also strongly biased, making occurrence counts misleading. Accounting for these factors is essential for a reliable and informative evaluation of multi-species SDMs. Here, we introduce a Sampling-Aware Global Evaluation (SAGE) benchmark, combining GBIF records for training with sPlotOpen vegetation plots for presence-absence evaluation across 5771 plant species. We propose an evaluation framework that groups species based on two properties, sampling effort and relative prevalence, which describe how densely a species' range is sampled and how frequently the species is recorded. Evaluating single-species SDMs and multi-species DeepSDMs, we find that Random Forests and DeepSDMs perform best overall, but neither dominates: DeepSDMs outperform single-species SDMs for infrequently recorded species while offering no consistent advantage for well-sampled ones. Crucially, this advantage emerges only when established bias-correction practices, such as spatial thinning and reweighting, are carried over to the deep-learning setting. SAGE helps identify the species and data conditions for which a given approach is beneficial, thereby supporting the development of more transparent and ecologically credible SDMs. Data and code: https://earens.github.io/sage/

Debates amongst agents driven by large language models (LLMs) have demonstrated vast potential in various applications, but when these interactions include humans and take place in high-stakes environments, e.g., in public policy deliberations, they are beset with issues such as sycophancy and a lack of faithful explanations. To tackle these issues, we present OmouAI, an interactive and inclusive deliberation system that uses LLMs in combination with computational argumentation, a field which excels in representing and reasoning within debates. OmouAI allows a human user to deliberate policy claims for real-world challenges with simulated personas, e.g., representing stakeholders, domain experts or devil's advocates, towards reducing sycophancy. Each persona generates its own arguments, and the arguments of all parties form a shared argumentation framework. Users can then contest, add and revise arguments, providing crucial human oversight. Then, arguments are evaluated using deterministic argumentative semantics against external goals, such as the UN Sustainable Development Goals, guaranteeing faithful explanations. The advancement or worsening of the goals thus serve as indicators for the policy recommendations.

Language agents struggle to act and learn in environments that require long sequences of low-level actions. Code-based abstractions can make these agents more productive by letting them invoke reusable skills instead of repeatedly selecting individual actions. The code handles recurring local decisions, while the language model decides which skills to use and how to combine them. Yet abstractions are leaky, and situations beyond a skill's capabilities may require a return to primitive actions. Motivated by this tradeoff between productivity and flexibility, we systematically study how code-based action abstraction affects the performance, inference cost, and learning of language agents. We study this in NetHack, a challenging, long-horizon game environment, using CodeHack, our library of code-based skills with natural-language descriptions. We use this library to compare agents restricted to primitives with those using semantic skills alone or in combination with primitives. We evaluate these agents in three settings: zero-shot prompting, supervised fine-tuning, and reinforcement learning. Across a broad zero-shot evaluation on NetHack, we find that compared with primitives, skills nearly triple game progression, while reducing inference cost per episode by 86%. Combining skills with primitives retains much of this benefit while preserving a path back down to low-level actions. Finally, in RL, we find that skill-based agents learn significantly faster than agents acting on primitives, achieving a 7.2x larger average gain in dungeon level over the same training budget. These results show that a supplied skill library can improve performance, efficiency, and learning, while retaining primitives provides flexibility when the library is insufficient. We release CodeHack together with training and evaluation code.

Corr2Cause asks whether a causal claim holds in every DAG compatible with observed correlations and conditional independencies. We frame this as latent-object reasoning: the label is defined by a CPDAG query, but free-form chain-of-thought often collapses the Markov-equivalence-class problem into local pattern matching. We propose Structured Thinking, a two-turn pipeline that first externalizes a typed, schema-constrained CPDAG summary and then answers against that graph state. On the Corr2Cause full test, Structured Thinking raises Qwen3.5-27B from $73.0$ to $86.4$ $F_1$(Yes) over a strong PC-instruction baseline in the primary paired run ($+13.4$ pp; McNemar $p=2.4\times 10^{-6}$; bootstrap $95\%$ CI [$+8.4$, $+18.6$]); across three full-ID seeds, the mean gain is $+8.1 \pm 5.3$ pp. A PC-scaffolded two-turn prose control reaches only $67.6$ $F_1$, indicating that a detailed PC scaffold plus a schema-free prose intermediate is not sufficient. The same pattern holds on Qwen3.6-27B, Paraphrase-OOD, and GPT-5.4-mini. Scrambling the emitted CPDAG costs $12.0$ pp $F_1$, and a full-split audit shows close agreement with the reference CPDAG (ID skeleton $F_1$ $0.960$; exact match $75.9\%$). These results support a bounded design principle: externalize the latent object that defines the label, constrain its form, and test whether downstream answers use it.

Quantum Machine Learning is a novel field of research aimed at devising machine learning approaches exploiting principles of quantum mechanics, such as superposition, entanglement and interference. In this context, we present a scalable hybrid Quantum Diffusion Model, and evaluate its use for medical image analysis. Specifically, our method is based on a Discrete-Time Quantum Walk algorithm, executed on a real quantum device, to model the forward dynamics of the diffusion model. For the backward step of the diffusion model, we devise and evaluate a classical learning model, which is used to reversely denoise the data. In contrast with other existing attempts at applying quantum machine learning for image analysis tasks, severely limited by the size of existing quantum devices, our method allows to process real-world large size medical data. In particular, we present results on grayscale and RGB images, as well as 3D volumes of moderate sizes. We benchmark our results by reproducing an alternative classical counterpart model, based on diffusion models on discrete state spaces. By doing so, we compare the generation capabilities of both models in terms of three distinct state-of-the-art metrics in the field of image generation, showing the competitive, promising results of our approach.

Patient-facing AI assistants promise valuable support to patients, but incoming queries can pose medical risks. To create guardrails, we work with oncologists to define three ordinal risk axes: Medical Urgency, Psychological Urgency, and Topic Sensitivity. We propose Clinical Guardrail Probes (CG-Probes) to measure the risks from query embeddings. We probe for each axis in the normalized embedding space of frozen embedders via the difference-in-means method, treating each axis as a potential linear direction. To train the probes, we cluster 79,658 Czech oncology search queries with BERTopic and use these clusters to generate pairs of queries with contrastive risk levels via few-shot prompting. We evaluate the approach on 200 queries (90 real, 110 synthetic), each graded by two oncologists, against two open-weight LLMs and a frontier LLM. We find that urgency-based axes are recoverable as linear directions, and the probes are competitive with open-weight LLMs (no significant differences in quadratic-weighted kappa) at a fraction of the latency. Each axis yields a scalar score that clinicians can inspect and use to set escalation thresholds. The pipeline requires only search logs, axis definitions, and black-box access to the embedding model, suggesting transferability across healthcare domains. Robust validation on new queries and axes remains future work.

Application developers of distributed learning services face challenges that a typical federated learning loop does not address. Specifically, the model updates can still leak private data, devices might not be able to participate in the training due to limited resources, a single aggregator might not be able to scale, and the transmissions of model weights induce a considerable bandwidth cost. This paper demonstrates DLaaS (Distributed Learning as a Service) from the developer's vantage point. Using a single admin dashboard, the developer initiates a distributed/federated learning job and is able to activate Differential Privacy (DP), Split Learning (SL), Hierarchical Aggregation (HA), and Knowledge Distillation (KD) as declarative options, with no change to the clients' code. We demonstrate the complete service lifecycle on an industrial smart-home Wake-up Word (WuW) task, using the "Ok Aura" dataset. Once the developer initiates a distributed learning job by toggling DP, SL, HA, and KD in the admin dashboard, the system dispatches the job to a set of Android clients and Dockerized helper aggregators. In the demonstration, these mechanisms run live across configurations. Then, the clients train the model locally and return their updates. The trained model is served to a consumer-side Android application that performs on-device WuW detection on a live microphone stream. In particular, the conference attendees will be invited to speak the trigger phrase and monitor in real time the per-class confidence and inference latency. Finally, we release the source code and short video walkthroughs of these configurations.

Machine unlearning aims to remove targeted information while preserving a model's other abilities. In realistic settings, such as privacy requests under the EU GDPR, the target may be narrow, for example information associated with a single person. Behavioral forgetting alone may be insufficient, motivating interventions directly on internal representations. However, standard mechanistic-interpretability extractors are poorly selective for such targets. We identify an energy bias in reconstruction-based extraction, which favors dominant background structure over low-energy target-specific components. We introduce SCALPEL, a contrastive sparse autoencoder designed to learn more selective forget features. We show theoretically that contrastive training promotes target-selective features and that our selection score controls expected background knowledge perturbation. We validate SCALPEL experimentally on TOFU across Qwen, Llama, and Gemma, where it substantially improves over NMF and standard SAE interventions and is competitive with Gradient Difference and RMU, bridging mechanistic interpretability and fine-grained unlearning.

Large Language Model (LLM) pollution occurs when synthetic responses contaminate data intended to capture human behavior. High deployment costs have so far limited the risk posed by autonomous survey agents. However, open-weight models paired with open-source agentic frameworks may have removed this barrier. We compared the performance and detectability of nine agent configurations, ranging from fully open variants to closed commercial ones. Each agent autonomously completed a survey containing multiple response types yielding various detection checks. Fully open agents ran locally without usage fees and performed competitively with commercial alternatives. Open and commercial agents failed different sets of checks, and no single check reliably detected all agents, but open-text responses discriminated best between agents and humans. These findings identify fully open agents as a distinct risk for LLM pollution and support multilayered detection strategies emphasizing open-text analysis.

Agentic LLM workflows decide their execution paths at runtime. Downstream computation may be predictable, or may have run before, yet it cannot begin until the model or the user resolves the branch. We call this serialization the branch-resolution barrier. Caching alone does not hide it: the key that identifies a reusable result is not known until then. In this paper, we propose DynBranch, which makes an unresolved branch addressable before it resolves. Its stable coordinate lets candidate subgraphs run during resolution and completed subgraph results be reused across later requests. A two-level controller admits this work when its expected benefit exceeds the load price. DynBranch sits at the model-API boundary and requires no changes to agent harnesses or model execution engines. Across four agentic workloads with Qwen3-32B on 4x H200 GPUs, DynBranch reduces mean latency by up to 32% over each workload's strongest prior system and by 46-66% against a no-reuse floor, while preserving workflow results. The benefit persists across backbone families and on a commodity Qwen3-8B/RTX 4090 deployment.

Collaborative science learning requires nuanced interpretation of student dialogue to characterize how learners identify knowledge gaps, build explanations, and work toward resolution - a theory-driven analysis that is labor-intensive and difficult to scale. We investigate whether instruction-tuned large language models (LLMs) can support multidimensional analysis of collaborative sensemaking without task-specific training, and whether structured knowledge-state information improves model inference. We evaluate two mid-size LLMs on 23 richly annotated, expert-labeled episodes across prompting conditions that vary definitional scaffolding, reasoning mode, and turn structure. Without reasoning, models tend to overpredict successful sensemaking; reasoning-enabled prompting improves identification of unsuccessful cases. Knowledge-state diagnostics provide additional grounding, improving detection of unsuccessful sensemaking and increasing agreement with expert annotations. No single configuration performs best across all sensemaking dimensions, underscoring the multidimensional nature of the task.

Conversational agents, generative recommenders, and personalized advertising all rest on one capability: understanding each user from raw behavior. Prevailing industrial practice is task-specific: for each task, a relevant subsequence is extracted from the full history and a dedicated model trained on it. In production it hits two bottlenecks. First, even after filtering, a single-task sequence stays extremely long: content-interest summarization reads several hundred items per user, tens of thousands of tokens once serialized as prompt text. Second, profiles are refreshed routinely: a billion users weekly, roughly 100K QPM in aggregate, which under a fixed GPU budget sets a hard throughput floor. Compression is therefore mandatory, yet truncation or coarse compression can silently distort the profile, introducing four hallucination types (fabrication, omission, date misattribution, broken logic) that, with no way to evaluate the compressed representation itself, surface only as diffuse degradation in downstream metrics. We present KuaFu, a unified behavior-compression layer whose minimal unit is one behavior item. A two-axis projector compresses each item into 2-4 tokens of width 128-256 (about 10x along the token axis, 20x along width; per-item cache 10 KB to 0.5 KB), with fidelity-oriented four-stage training and layered intermediate evaluation. Across four production profiling tasks it matches or exceeds uncompressed single-task production models on all five headline metrics, raises per-GPU throughput by 37%-350%, and saves 190 GPUs. On public benchmarks it nearly always beats prior compressors at the same compression ratio (up to +17.7 EM on out-of-domain MRQA); on RecBench, a 4B model surpasses its 8B counterpart by 1.90 points. KuaFu has run on the Tencent advertising and recommendation platform for ten months, lifting overall GMV by 1.37%.

The rise of autonomous AI agents equipped with tools has introduced significant security risks, ranging from unintended tool misuse to adversarial manipulation through Indirect Prompt Injection (IPI) attacks. In practice, deployed agent systems such as OpenAI Codex and Claude Code protect tool invocations through a combination of coarse-grained permission rules and LLM-based judgments about individual proposed actions. Both components, however, have important limitations: static policies must anticipate possible user intents and therefore do not scale to open-ended tasks, while LLM-driven authorization supports dynamic decisions but produces inconsistent outcomes and remains vulnerable to targeted IPI attacks. To provide scalable and more consistent authorization, we propose MetaPermit, a policy-based tool access-control framework that decouples semantic inference from security enforcement. By analyzing agent-user interactions, we derive a compact, task-independent set of meta-attributes that capture the relationships among the user's intent, the execution context, and the proposed tool call. These meta-attributes allow MetaPermit to authorize tool use without enumerating user intents. At runtime, an LLM infers the meta-attribute values for each proposed tool call, while a fixed policy evaluates these values to allow or deny the call, making each decision auditable through the inferred values and the applied policy rule. We evaluate MetaPermit on the AgentDojo and AgentDyn benchmarks, across seven task suites and five attack methods, using two widely deployed open-weight LLMs. The results show that MetaPermit produces 31% more consistent authorization decisions than LLM-driven authorization and outperforms the state-of-the-art defenses CaMeL and IPIGuard in both task completion, with improvements of up to 109%, and robustness to IPI attacks, with no malicious tool calls executed.

Time series foundation models (TSFMs) enable cross-domain forecasting, but their development as general-purpose forecasters remains constrained by underexplored training potential and limited architectural versatility. To address these challenges, we introduce Aurora-X, a billion-scale TSFM with a progressive curriculum and a unified architecture. We first use channel-independent pretraining to learn temporal patterns, then introduce cross-variable dependencies, varied context and horizon lengths, and future covariates if available during midtraining. Variable-resolution post-training further enables an adjustable temporal span per token at inference. With fixed model weights, this supports longer histories under a fixed token budget or fewer tokens for the same history, enabling test-time scaling. With a versatile architecture, Aurora-X supports cross-variable modeling, covariate conditioning, and parallel decoding of future patches for probabilistic forecasting. These are supported by a novel pattern-guided mixture-of-experts that expands model capacity through sparse activation and uses shallow patch similarities to constrain deep-layer routing, guiding expert specialization across heterogeneous time series. Furthermore, we propose an implicit quantile network head that predicts arbitrary quantiles to characterize predictive distributions, enhancing probabilistic forecasting flexibility. Comprehensive experiments on GIFT-Eval, TIME, FEV-Bench, TFB, and DAG-Bench demonstrate state-of-the-art forecasting performance against pretrained TSFMs and task-specific supervised models.

Clustering on graphs where both node attributes and structural links are partially missing remains a challenging task. Existing methods typically rely on imputation-then-clustering on single-view missingness incomplete graphs, which are vulnerable to cross-view error propagation and cluster-boundary blurring under simultaneous attribute and structure missingness. To address these limitations, we propose a Robust Graph Clustering Network for Multiple Missing Data (RGCN), which is designed to handle simultaneous node attribute and graph structure incompleteness. RGCN introduces three key innovations: First, we design a view-decoupled dual-branch imputation to mitigate interference and enable mutual enhancement in recovering missing data. Second, we employ a multi-hyperspherical mixture prior to enhance intra-cluster compactness and inter-cluster separability on a directional latent manifold. Third, a boundary-aware contrastive enhancement objective mitigates the blurring of clusters caused by imputation bias. Extensive experiments on real-world datasets demonstrate that RGCN consistently outperforms state-of-the-art baselines under various missing patterns.

Deep ensembles improve robustness in mobile sensing, but repeatedly executing many models over continuous sensor streams is costly. Selecting only a few members reduces this cost, yet adaptive selection often requires additional model execution to obtain reliable evidence about inactive candidates. We present MetaSE, an active ensemble framework that exploits short-term persistence in per-model reliability. MetaSE maintains a small active set across windows, uses post-execution evidence to reject unreliable members, and invokes lightweight routing only when replacement is needed. This stateful design accesses the diversity of a larger pool without repeated full-pool evaluation. Across four HAR datasets and four model architectures, MetaSE consistently improves over a fixed three-model ensemble and achieves accuracy comparable to substantially more expensive adaptive and full-ensemble inference. On a Raspberry Pi 4B, MetaSE is 2.7x faster and uses 69% less memory than full ten-model inference.

Reasoning systems usually treat premise use as a question of relevance: if a fact is available and useful, it may be selected for inference. Authorization imposes a different constraint: a premise may be represented and logically usable but not permitted for a particular local transition. We formalize this distinction as Governed Deduction (GD), with a transition-local admission predicate admit(p, tau, S). From an independently produced RBAC-augmented Spider benchmark, we construct 4,461 matched authorization pairs in which the same query premise and policy state support permitted and denied consuming transitions. An initial joint controller reaches 99.19% held-out accuracy, but a transition-only control reaches 100%, exposing a role-name shortcut. After a frozen, label-independent context-local role permutation removes that shortcut, premise/state-only, transition-only, and joint linear controllers all score exactly 50% on 1,856 held-out edges, while a symbolic policy oracle remains at 100%. The result is a controlled negative finding: the benchmark instantiates policy-grounded authorization beyond relevance, but the frozen linear representation does not recover the relation. Matched one-sided controls and leakage audits are therefore essential for evaluating learned policy-sensitive reasoning.

Precise, high-speed control remains challenging for robots with complex actuation dynamics. Learning directly on hardware is further constrained by the cost of real-world interaction. We present an online model-based reinforcement learning framework that learns a probabilistic dynamics ensemble model from scratch for sampling-based model predictive control. A precision-gated contouring objective conditions the progress reward on path accuracy, prioritizing precision over speed. In a data-driven excavator simulator, the framework achieves higher sample efficiency than the evaluated model-based reinforcement learning baselines. We validate the framework by learning directly on an 11.5-ton Menzi Muck M445 hydraulic excavator, without demonstrations or simulation pretraining. After 20 minutes of interaction, the controller reaches tracking accuracy comparable to prior learned controllers trained on 100-150 minutes of data. After 40 minutes, it sustains sub-centimeter mean path error at high operating speeds.

Sound matching can be formulated as optimizing synthesizer parameters against an audio-domain objective. However, objectives derived from generic audio representations are often difficult to optimize, while direct search requires rendering every candidate. We introduce Synth-JEPA, which learns mutually predictive audio and parameter representations from paired synthesizer data. At inference, candidate parameters are scored directly in this learned space, yielding a renderer-free objective whose audio geometry is shaped by parameter correspondences rather than generic audio similarity. We evaluate Synth-JEPA on Surge XT using held-out synthesizer sounds and out-of-domain NSynth and FSD50K targets, against inverse models, direct search, and learned proxy objectives. Synth-JEPA outperforms all baselines in-domain and remains competitive out-of-domain. Its matching quality continues to improve with additional test-time search, allowing compute to be traded for match quality. In pairwise listening tests, listeners preferred Synth-JEPA in 85% of trials overall. Together, these results show that an audio representation with a parameter-induced geometry allows synthesizer sound matching to be approached as an effective renderer-free search problem.

Generalization in Reinforcement Learning (RL) refers to the ability to execute close-to-optimal policies in unseen tasks after the agent has been trained on a different set of tasks. Building on the seminal work of the successor representation and further adaptations with function approximation, Transfer in RL has traditionally focused on generalizing to tasks that only differ in the reward function. A decade after the introduction of the successor representation, Robust RL emerged simultaneously from several articles in the field of operations research. In Robust RL, the transition kernel is unknown, and the goal is to maximize the expected reward under this uncertainty. Our work unifies these two paradigms through robust successor features, which generalize across both the reward function and the transition kernel, under the assumption that tasks are linear Markov Decision Processes. We derive a bound on Generalized Policy Improvement (GPI) that explicitly quantifies how performance degrades with the mismatch between transition kernels, recovering existing successor-feature guarantees when dynamics are shared. Finally, the generalization capabilities of robust successor features are validated on several grid-based benchmarks and compared to previous alternatives that focus solely on either the reward or the transition kernel.

Researchers increasingly use generative large language models (LLMs) to convert corporate text into empirical variables. We examine the extent to which LLM-based textual measures are invariant to model choice using thirteen measures, including sentiment, management clarity, uncertainty, answer specificity, and climate and political risk. Seven LLMs from different providers score earnings call transcripts of S&P 500 companies on these constructs. Cross-model rank correlations average only 0.52, and transcript-level differences common across providers account for only 34% of total score variation. Cross-model disagreement does not predict subsequent analyst or market disagreement, consistent with a substantial model-specific component rather than common ambiguity in the underlying disclosure. Model choice significantly affects downstream inference, with coefficient magnitudes, signs, and statistical significance varying substantially across models. Averaging across providers makes transcript rankings more stable for most constructs, but score levels remain sensitive to the models included in the ensemble. LLM-generated variables should therefore be treated as model-contingent measurements and validated across providers.