The Inference Report

Representation Autoencoders (RAEs) enable diffusion models to operate in the feature spaces of pretrained visual encoders. However, many off-the-shelf encoders are not optimized for faithful reconstruction, discarding fine-grained visual details. As expected, finetuning these encoders for image reconstruction recovers such details. However, perhaps counterintuitively, this procedure reduces the effective dimensionality of the resulting representation, and the altered geometry has downstream effects on generation. Specifically, we show that using the standard velocity prediction in flow matching in this high-dimensional space requires the model to fit orthogonal noise directions outside the low-dimensional signal manifold, making optimization inefficient. This motivates using the clean data parameterization ($\boldsymbol{x}_{0}$-prediction) instead, which focuses learning on the underlying signal manifold. Across experiments with multiple strong-reconstruction encoders, we show that $\boldsymbol{x}_{0}$-prediction consistently improves text-to-image generation performance.

Our results show that contrastive learning outperforms a classification-based approach to authorship verification under the tested settings. We identify loss function, batch size, training duration, pre-trained model, input context length, and random text span data augmentation as important factors of model performance. Based on these considerations, we develop a ModernBERT Bi-Encoder model that achieves 98.4% accuracy on the PAN21 authorship verification task.

Artificial intelligence offers an unprecedented opportunity to augment human capabilities, yet progress at the frontier has focused primarily on advancing model capabilities. We introduce StudentBench, a suite of AI teaching evaluations and a public platform that enables large-scale data collection with over 175,000 student-AI messages to study whether large language models (LLMs) produce learning gains equivalent to human tutoring. Using StudentBench, we measured learning gains on Quantitative and Verbal GRE questions across 2,383 human participants receiving AI tutoring, human tutoring, or no tutoring. We establish that AI tutoring is statistically equivalent to expert human tutoring for GRE learning gains (p = .015), and in five of the seven GRE domains, the best performing AI tutor surpassed the human tutor, on average. In a second study, expert human tutors compared LLM-generated lesson plans and practice problems through 2,028 pairwise rubric evaluations. Together, the two studies clearly separate AI tutors across: (1) lesson planning, (2) practice-problem creation, (3) conversational pedagogy, (4) cost, and (5) engagement. Surprisingly, one AI tutor achieved learning gains equivalent to human tutoring (p = .044) at 918 times lower cost (USD 0.0052 for AI versus USD 4.81 for human, per percentage point gained). For Quantitative GRE sessions, faster AI replies correlated with more student messages, more messages with more correct practice, and more correct practice with larger learning gains (all p < .002). The StudentBench platform is freely available at https://studentbench.org.

Social navigation typically assumes a specified goal and focuses on reaching it while respecting social conventions, whereas robot group joining requires predicting where to join based on the group's real-time activity and formation. This is a highly semantic task, yet an important capability for applications such as robotic guide dogs and autonomous mobility scooters. We formulate language-grounded robot group joining: given an observation and a natural-language description of a target group, the robot identifies the relevant group members and predicts socially compliant joining poses. For grounding, we generate structured candidate subsets through recursive spectral partitioning and rank them with a language-conditioned image--geometry model. Given the grounded group, a goal predictor leverages human-formation priors to produce a multimodal energy--orientation map over feasible robot poses. Experiments on conversations, queues, and audiences across varying group sizes, crowd densities, and visual ambiguities show that our method achieves competitive grounding accuracy with sub-second inference and outperforms all baselines in joining-pose prediction. Real-robot experiments further demonstrate group joining in both static and dynamically changing interactions.

We introduce Sharper Transductive Local Complexity (STLC), a localized complexity method for transductive learning under uniform sampling without replacement. The construction starts from a Bernstein-type concentration inequality for the supremum of the test--train empirical process. Its proof uses the modified log-Sobolev inequality for the swap walk and a two-parameter entropy closure. A peeling argument with a surrogate localization functional then gives excess-risk bounds with the same fixed-point and confidence terms as the classical inductive local Rademacher-complexity bounds, without the additional logarithmic confidence factor in earlier transductive results. For realizable learning over a binary class of VC dimension $\dVC$, with training size $m$, test size $u$, and $u\ge m\ge\dVC$, STLC yields $\cO\{\dVC\log(me/\dVC)/m\}$. This matches the standard inductive rate and, when $m\ge9$, is within a logarithmic factor of the transductive minimax lower bound of order $\dVC/m$. For transductive kernel learning, STLC gives a spectrum-adaptive excess-risk bound without the multiplicative imbalance factors appearing in the earlier local-complexity bound.

Large language models (LLMs) are increasingly used in coding tasks, but their ability to reason about code execution remains unclear. Existing repository-level QA benchmarks mainly evaluate static code understanding and often rely on LLM-based evaluation, while execution-reasoning benchmarks are mostly limited to snippets or functions. We introduce SWE-Flux, a repository-level benchmark for dynamic execution reasoning containing 480 execution-grounded instances across 12 real Python repositories, with gold answers automatically harvested from instrumented test executions rather than written manually or judged by LLMs. The benchmark covers singletest and multi-test questions over control flow, loops, program state, dataflow, exceptions, and program invariants. Evaluating five LLMs shows that this task remains challenging. The best model achieves only 37% accuracy. Models perform better on localized behavior such as invariants, intra-procedural control flow, exceptions, and simple loops, but struggle with dataflow, inter-procedural execution, precise state reasoning, and suite-level aggregation. Finally, we show that the oracle-harvesting pipeline can generate fresh benchmark variants using input perturbation. It successfully harvests valid variants for almost 90% of the selected instances, and the resulting variants are substantially more challenging for the evaluated models.