arXiv AI

arXiv:2609.35954v2 Announce Type: replace-cross Abstract: Large language model post-training generates self-generated rollouts through reinforcement learning and on-policy distillation, yet this experience is often treated as stale once the policy advances. Historical rollouts can remain compatible with a later policy while preserving behaviors that the policy no longer expresses reliably. However, they may also contain mistakes, abandoned attempts, and redundant actions that should not be imitated, motivating finer-grained selective supervision. We introduce ROSS (Relearning from Self-Generated Rollouts through Selective Supervision), which preserves the full historical trajectory as context while applying loss only to selected model-generated continuations. Across domain-specific reinforcement learning, multi-teacher on-policy distillation, and agentic reinforcement learning, ROSS consistently improves upstream checkpoints and outperforms baselines across mathematics, code generation, instruction following, and software engineering. On Qwen3.6-35B-A3B, ROSS improves the six-benchmark MOPD average from 58.40% to 62.20% and SWE-bench Verified from 64.20% to 68.40%. These results show that self-rollout training leaves behind reusable behavioral experience that can yield further gains through offline supervised fine-tuning (SFT), without additional policy rollouts.

arXiv:2609.35814v2 Announce Type: replace-cross Abstract: As browser-use agents improve, benchmarks keep pace by adding new tasks, websites, and applications. Doing so is costly, and because many factors change at once, it is unclear what makes the new tasks harder. In this work, we instead make existing tasks harder by changing the environment in which they are performed, so that difficulty becomes a programmable property of the environment. We introduce BreakingWeb, in which each task is run twice: once on an unmodified website and once under an intervention, a controlled change to the website's content, server state, network responses, or rendered page. Each intervention is reproducible, keeps the user instruction and success criterion fixed, can be detected and worked around by a competent user, and is labeled with the behavior that recovery from it mainly requires, such as verification. BreakingWeb contains $519$ task pairs on seven self-hosted websites, and each run is scored on the website's final backend state. We evaluate six strong browser-use agents, three agents that observe only screenshots, and human participants. Interventions lower the pass rates of the six browser-use agents by $22.9$ percentage points on average, and each agent fails on nearly half of the tasks it solves on the unmodified website. Humans are affected much less: their pass rate drops by $10.0$ points on the first attempt and by $5.7$ points when they repeat the task. Most failures of the six agents ($75\%$) are \emph{belief failures}, in which the agent declares the run finished although a required change was never made. Our code, data, and environments are available at https://www.breakingweb.app.

arXiv:2609.35797v2 Announce Type: replace-cross Abstract: Large language model (LLM) judges are often used as rewards to train policies on objectives that deterministic verifiers cannot capture. However, these rewards are often collapsed to pass/fail ({0, 1}), which reports the verdict but not how well a response met each criterion. We model each pass/fail verdict as a score on an unreported scale, compared with one cutoff. A stretch of that scale moves every score proportionally toward or away from the cutoff, but never across it, so no verdict changes. A policy is therefore free to apply any stretch without changing anything the panel reports. Under a joint-Gaussian model, a third grade adds a second threshold and removes this affine stretch ambiguity. On MATH and SciBench outputs from one seven-criterion judge, all 14 constructed criterionwise stretches were invisible after binarization but visible with three grades. At $n=1{,}024$, a test given both population laws had at least 96.5% power at a $1.5\times$ stress. Retaining grades closes one blind spot created by binarization, but verdicts alone remain insufficient as some changes are still indistinguishable from genuine improvement. These include arbitrary within-grade changes and fixed-covariance, loading-aligned mean shifts -- the signature of a sycophancy-shaped lift the panel reads as competence. The shared-factor reference approximation fit MATH and SciBench but not HealthBench, delineating its empirical scope. We recommend keeping at least three grades (for example, asking the judge whether each criterion is fully, partially, or not met and rewarding {0, 0.5, 1}), and externally validating gains along the remaining direction, which no finer scale removes.

arXiv:2609.35767v2 Announce Type: replace-cross Abstract: Unified multimodal models can both look at and render images, so in principle they can repair their own generations: diagnose what an image gets wrong, revise it, observe the result, and diagnose again. Whether a revision helps is known only after it is rendered, so the reflection text and the image generation must be learned jointly, over the whole loop. Supervised fine-tuning (SFT) on reflection trajectories gives a cold start but does not find the high-success repair paths, and naive RL that optimizes only the renderer or only one head leaves most of the gain untapped. We introduce UMM-Reflection, which applies reinforcement learning (RL) to complete reflection trajectories inside one unified model: sibling trajectories share one initial image, so the group-relative advantage compares reflection strategies, and one trajectory-level advantage updates both the reflection tokens and the flow-based revisions, avoiding the combinatorial blow-up of per-round credit assignment. Unlike single-round editing or pipelines with an external critic, credit flows across rounds and to both roles of the same model, and no verifier is needed at inference. On BAGEL, UMM-Reflection improves GenEval by 12.05 points over SFT, and the gains transfer to WISE (+10.97), OneIG-Bench (+3.48), and T2I-CompBench++ (+4.63), none of which is used in training.

arXiv:2609.35052v2 Announce Type: replace-cross Abstract: Video world models must preserve the visual state of the world over time, but existing evaluation protocols often rely on generated histories, video reference, or selected revisit viewpoints that can confound the assessment of a model's true memory capability. To address this, we introduce OPIS, an input-grounded benchmark that strictly anchors the assessment to a fixed set of object instances from the initial observation for evaluating multi-object memory in video world models. The OPIS dataset comprises 500 cases across real-world, embodied-robotic, and game-world domains, providing dense object-level annotations for 12,672 rigid, articulated, and deformable instances. Our object-centric evaluator combines association and explicit visibility reasoning to hierarchically measure Object (O) Presence (P), Identity (I), and Structure (S), utilizing static or dynamic evaluation tracks based on object kinematics. Across eight image-to-video or camera-conditioned world models, our proposed OPIS scores range from 48.65 to 56.01. As the reference inventory grows from less than 20 to more than 40 objects, the Presence, Identity, and Structure scores show an overall decline, with the average Identity score falling from 40.22 to 23.11. The results demonstrate that preserving the particular object instances in the input is considerably harder than generating plausible visual elements.

arXiv:2609.26039v2 Announce Type: replace-cross Abstract: Point cloud generation has emerged as a crucial task for accurately capturing and reproducing the complexity of the physical world. However, existing generative approaches, predominantly relying on Transformers and Variational Autoencoders (VAEs), frequently ignore the continuous, non-grid topologies inherent to 3D spaces. Although the integration of graph-based structures has yielded significant benefits in related discriminative vision tasks, such geometric architectures remain noticeably absent from 3D generative modeling. To address this gap, we introduce EMERGE (Equivariant Multi-scale GNN for Resolution-agnostic point cloud GEneration), the first fully $SE(3)$-equivariant graph-based diffusion backbone explicitly designed to generate point clouds while preserving continuous spatial symmetries. Our framework bypasses the rigid resolution dependencies of standard generative pipelines, enabling zero-shot inference at multiple, arbitrary spatial resolutions. Extensive empirical evaluations demonstrate that EMERGE achieves State-of-the-Art generation quality across standard metrics, while the strong inherent geometric inductive biases enable significantly faster training convergence compared to existing baseline methods.

arXiv:2609.17601v2 Announce Type: replace-cross Abstract: This paper studies decentralized learning of socially optimal equilibria in finite normal-form games over dynamic communication networks. Each agent observes only its own realized payoffs, does not know the game a priori, and can communicate only with time-varying neighbors using low-bandwidth messages. We propose networked decentralized optimal equilibrium learning dynamics in which agents generate randomized semantic content/discontent signals from local payoff comparisons and exchange time-stamped time-stacked tables rather than raw actions, payoff information or local estimates/parameters. The method combines table fusion with temporal majority reconstruction to mitigate dynamic communication while preserving fully decentralized operation. We establish finite-time logarithmic regret guarantees, with an in-phase exploration perturbation, for optimal equilibrium selection under utilitarian and proportional-fair social welfare objectives. Simulation results further show that the proposed approach can effectively select socially desirable equilibria over dynamic communication networks.

arXiv:2609.09684v2 Announce Type: replace-cross Abstract: We study medical question-answering (QA) data selection when a developer already has questions, labels, and auxiliary explanatory texts but can retain only $B$ examples for a final LoRA update. Representation-Sensitivity Prioritization (RSP) first restricts the candidate bank to its top-$4B$ predictive-entropy examples, then combines the frequency of answer flips with non-flipping contraction of the top-two answer margin under hidden-state perturbations. We evaluate this operational score on five medical QA datasets with google/medgemma-4b-it, $B=128$, and three seed branches. After removing four MedExpQA test questions that also occur in its acquisition pool, dataset-specific development-tuned RSP attains the highest observed five-dataset mean accuracy (53.44%) and Macro-F1 (48.21%) among the evaluated selectors. Relative to random selection within the same entropy-restricted pool, the descriptive differences are +1.49 and +2.10 percentage points, respectively; paired 95% bootstrap intervals include zero. The audit mixes benchmarks with different relationships to documented backbone post-training sources, and the same test cohorts had been used in earlier analyses before the expanded tuning grid was designed. Results are therefore exploratory, modest, and dataset-dependent rather than evidence of statistical superiority. RSP point estimates also exceed an archived no-update reference, but unpinned model revisions and incomplete historical input identifiers limit attribution of that difference to fine-tuning. Because every selected training example includes auxiliary text and no matched answer-only control is evaluated, the experiments also do not identify a causal benefit from auxiliary-text conditioning itself.

arXiv:2609.04180v2 Announce Type: replace-cross Abstract: Gaps remain in our understanding of how large language models (LLMs) acquire knowledge during pre-training. We find that auxiliary views, conceptual reformulations, are key to this process. We design controlled experiments to isolate this. First, holding the token budget fixed, allocating tokens from paraphrases to auxiliary views improves learning even for factual recall, which we use near-verbatim sentences from the original document to test. Second, the effectiveness of auxiliary views is not contingent on the strength of the model that generates them. Third, we establish that paraphrasing is helpful but only at smaller batch sizes. Fourth, we identify types of knowledge that aid learning in the presence of prior knowledge gaps. Finally, we examine how these effects manifest mechanistically via layer-wise biases and compression. Together, our findings suggest that auxiliary representations of knowledge are a key factor in the success of pre-training, and explains why data diversity matters. Further, our results provide clear recipes for synthetic data augmentation along with implications on the self-improvement in LLMs.

arXiv:2609.04061v2 Announce Type: replace-cross Abstract: Large language models (LLMs) are increasingly used to edit existing code, but correctness alone is not enough: useful repairs should also be minimal, reviewable, and faithful to the original implementation. We study over-editing, the tendency of a model to rewrite code beyond what is required to fix a bug. We construct an evaluation framework from 400 BigCodeBench problems by injecting controlled AST-level corruptions into reference solutions, giving each repair task a known minimal patch. Across frontier LLMs, over-editing is widespread even among strong models like GPT-5.5: high Pass@1 can coexist with unnecessarily large edits and added cognitive complexity. A preservation instruction substantially reduces this behavior, lowering average excess Levenshtein distance from 0.195 to 0.131, reducing added cognitive complexity by 26.6%, and increasing Pass@1 by 2.3 points. However, these gains do not simply follow from a larger reasoning budget or larger models. We next ask whether minimal editing can be learned directly during post-training. We observe that supervised fine-tuning overfits to seen corruption patterns, whereas reinforcement learning gives the best out-of-domain edit-fidelity and performance-retention trade-off. These results position edit fidelity as a distinct axis of code-repair quality and show that it can be measured and learned.

arXiv:2609.03697v2 Announce Type: replace-cross Abstract: We develop theory for the identification of complex causal queries from non-IID data. We propose a formal description of stochastic systems based on symmetries in data leaving causal mechanisms invariant. These symmetries become the source of statistical repetition for identification and mandate knowledge-transfer from observations to queries. The result is an abstract, simple and very general language for causal reasoning. We give descriptions of models and queries, and the formal infrastructure for their mathematically rigorous identification. The proposed structure enables a number of simplifying principles: A shared description of mechanisms and interventions, a duality of observed and queried structure, and the sub-division of the identification task into three steps - extraction, computation and query-expression. Our approach reproduces and matches standard theoretical results on IID data and transport of experimental and non-experimental data. By design, it additionally applies beyond IID data and captures complex causal queries not expressible via do- or soft-interventions.

arXiv:2608.27309v4 Announce Type: replace-cross Abstract: LLM-judge audits assess bias by comparing ratings across matched conditions. Difference-in-differences designs compare two candidate responses within each item and then compare that contrast across a manipulated attribute, with an endpoint read off a bounded rating scale. Building on prior work on nonlinear difference-in-differences, we show in closed form that when ratings are bounded, a severity shift common to both responses produces an observed interaction whenever the scale bounds attenuate it unequally. We examine this problem in a pre-registered audit comprising 990 calls to a frozen pedagogy judge. The registered primary analysis did not detect an effect of a stated novice or advanced profile on scaffolding preference. The only nominally significant secondary interaction concerned the productive-struggle sub-score ($+0.378$ points; $p = 0.002$). A post-hoc construction with zero differential preference reproduced 79 to 85% of this interaction using the observed severity shift and the scale floor alone. The observed interaction therefore does not identify differential preference on the latent scale.

arXiv:2608.19085v3 Announce Type: replace-cross Abstract: World models for autonomous driving should anticipate the consequences that matter for driving decisions rather than reconstruct every detail of future scenes. However, predictive representations alone are not guaranteed to distinguish candidate actions that lead to different planning outcomes.We propose DA-WAM, a decision-aligned world action model that couples a Joint-Embedding Predictive Architecture (JEPA) with candidate-level trajectory scoring. During joint training, the JEPA objective grounds future representations in observed scene dynamics, while the planning objectives encourage these representations to capture distinctions relevant to safety and utility. An action-conditioned predictor generates a distinct future latent for each candidate trajectory, and a multi-head trajectory scorer uses the corresponding latent to estimate planning factors and overall utility. In addition, safety-critical post-training with hard negatives further improves discrimination between geometrically similar trajectories with different safety outcomes. DA-WAM achieves state-of-the-art performance among learned camera-based planners on NAVSIM-v1, NAVSIM-v2, and Bench2Drive. Ablation studies demonstrate the effectiveness of each component in driving decision-making.

arXiv:2608.18036v2 Announce Type: replace-cross Abstract: MRI reconstruction methods for undersampled k-space data naturally utilize complex-valued measurements. Parallel developments in sparse phase retrieval have shown that magnitude-only measurements may provide complementary information for signal recovery. However, their use in MRI reconstruction remains largely unexplored, due to lack of practical settings where informative magnitude measurements can be obtained without additional scan time. In this work, we investigate the use of auxiliary k-space magnitude information for accelerated steady-state dynamic MRI reconstruction, and demonstrate strong consistency of k-space magnitudes across time-frames. Building on this observation, we propose $\mathbb{C}+\text{Mag}$, a magnitude-informed physics-driven deep learning reconstruction method. The proposed method employs an ADMM-based unrolling framework with a novel magnitude-aware data-fidelity formulation, where quadratically smoothed optimization and momentum-based updates are introduced to address the non-differentiability and non-convexity of the magnitude constraints. Experiments on retrospectively undersampled cine MRI and phase-contrast flow MRI datasets, as well as prospectively undersampled real-time cine MRI acquisitions, demonstrate improved artifact suppression, sharper anatomical recovery, and better preservation of phase information compared to conventional PD-DL methods, which is further supported through blinded expert reader evaluations.

arXiv:2608.06151v2 Announce Type: replace-cross Abstract: The emergence of conspiracy theories in the wake of major events is a significant societal challenge. Here we test whether conversational dialogues with a large language model (LLM) can reduce belief in immediately unfolding conspiracies. In experiments conducted in the days following the July 2024 assassination attempt on Donald Trump and the September 2025 assassination of Charlie Kirk, U.S. adults (Experiment 1: N = 472; Experiment 2: N = 1042) holding conspiratorial views about the crisis event engaged in a multi-turn conversation with an LLM prompted to reduce their conspiracy belief. Compared to control participants who either discussed an irrelevant topic with an LLM or viewed a static fact sheet, participants in the LLM treatment showed significantly reduced conspiracy beliefs in both experiments. We also found some suggestive evidence of downstream effects of the LLM treatment, observing reduced belief in some -- but not all -- analyses of different conspiratorial beliefs one to two months later. These results shed light on the psychology of emerging conspiracies and highlight the potential for scalable, cognitively-focused interventions to counteract misinformation in the immediate aftermath of high-profile societal events.

arXiv:2607.15879v3 Announce Type: replace-cross Abstract: Much empirical legal research depends on translating unstructured text into structured variables. In corporate governance research as elsewhere, this translation has traditionally been performed by humans. Manual coding is time-consuming and costly. In corporate governance research, standardized commercial datasets offer an alternative, but they cover a limited set of firms and governance variables. This paper introduces the DECODEM Benchmarks and uses them to evaluate whether large language models can extract corporate governance variables from charters and bylaws. The benchmarks combine a sample of corporate charters and bylaws of publicly traded U.S. corporations obtained from the SEC's EDGAR database with hand-coded labels that capture whether a document includes a specific governance arrangement. The paper uses these benchmarks to evaluate a range of extraction approaches and models. The approaches vary along several dimensions, including the amount of guidance included in the prompts and the extent to which the model is provided with the full document or with relevant excerpts only. The results show that current models can recover many governance variables with high accuracy across extraction approaches. At the same time, performance varies substantially across documents and variables. In particular, extraction is less accurate for charters than for bylaws, including when the comparison is limited to variables that appear in both benchmarks. The results also show that frontier models perform comparably on the tasks evaluated here and that more elaborate extraction pipelines do not consistently improve their performance. More elaborate extraction strategies matter more for efficiency-oriented models, where they can substantially narrow the performance gap relative to frontier models.

arXiv:2607.11578v2 Announce Type: replace-cross Abstract: Deep learning for EEG-based seizure detection faces critical challenges: severe annotation scarcity and extreme class imbalance, where ictal events comprise less than 10\% of clinical recordings. We present DiffEEG, a 9.6M-parameter self-supervised foundation model that addresses both limitations through denoising diffusion pre-training and reinforcement learning (RL)-based fine-tuning. Pre-trained on 1.3M unlabeled segments from the Temple University Hospital Seizure Corpus (TUHSZ), DiffEEG learns generic neural representations via a 1D U-Net with multi-head self-attention. For downstream adaptation, a reinforced decision layer employs policy gradient optimization to directly maximize F1-score, prioritizing sensitivity to rare seizure events over overall accuracy. Under strict patient-wise evaluation (279 patients, Leave-One-Fold-Out), DiffEEG achieves 61\% accuracy and 59\% F1 for 4-class seizure subtyping, and 81\% accuracy with 85\% weighted F1 for binary detection, maintaining clinically viable seizure recall (59\%) despite extreme imbalance (6.7\% prevalence). Segment-level evaluation establishes an upper bound of 97.6\% accuracy, confirming strong architectural capacity. DiffEEG demonstrates that diffusion-based pre-training combined with metric-aware reinforcement learning enables clinically deployable seizure monitoring with minimal labeled data requirements.

arXiv:2607.08409v2 Announce Type: replace-cross Abstract: LLM-based ASR adapted to regulated domains such as banking is bottlenecked by privacy: real speech is costly and legally constrained to collect, making synthetic text-to-speech (TTS) an attractive substitute. Yet synthetic speech stays acoustically mismatched with real recordings, and work on this gap has stayed within supervised fine-tuning (SFT). We instead turn to reinforcement learning, and show that Group Relative Policy Optimization (GRPO) extracts far more from the same synthetic speech than SFT. Synthetic-only adaptation of the model with GRPO, a critic-free method rewarding low-WER hypotheses, reduces WER by 40% relative to SFT (36.71%$\to$22.09%), and an SFT-then-GRPO combination pushes this further to 45%. We trace the gain to behavior rather than representation: GRPO reduces insertion errors by improving stopping calibration and speech-to-text alignment by better anchoring attention to audio, leaving early-layer representations intact. When synthetic speech is the main resource, reinforcement learning should be preferred over supervised fine-tuning.

arXiv:2607.03929v2 Announce Type: replace-cross Abstract: Multi-Meta-RAG improves retrieval for multi-hop question answering by filtering a vector store on metadata, specifically the news source and the publication date, which it extracts from each query using GPT-3.5. This extractor requires a proprietary API call for every query, and its output is not restricted to the allowed sources: even though the prompt specifies 49 sources, GPT-3.5 generates 128 distinct source names, 79 of which do not appear on the list. We show that the source part of this extractor can be replaced by a local, deterministic probe trained on the hidden states of a small open-source language model. The probe is a multi-label logistic classifier on the mean-pooled hidden states of one layer, trained with a class-balanced loss on the gold evidence sources. Using 5-fold cross-validation to select the layer, we evaluate four models ranging from 135M to 1.5B parameters out-of-fold across all 2556 MultiHop-RAG queries. The probe reaches a set-exact accuracy of 90.9%, compared to 88.0% for a substring baseline and 80.9% for GPT-3.5. The difference comes from null queries, for which GPT-3.5 never predicts an empty source set. Across the remaining queries, all three methods achieve set-exact accuracies within 1.1 percentage points of each other. Unlike earlier probing work, we use the probe as a component of a retrieval pipeline rather than for interpretation. For all four models, the cross-validation selects one of the first layers, in contrast to the middle layers reported for retrieval tasks. We attribute this to the source name appearing verbatim in 95.4% of gold query-source pairs. The 135M model is 1.5 percentage points less accurate than the 1.5B model, so the filter requires only a partial forward pass through a small model and one linear layer. The code is available at https://github.com/mxpoliakov/Multi-Meta-RAG.

arXiv:2606.28235v2 Announce Type: replace-cross Abstract: In July 2026, about 1,200 AI agents in an internal OpenAI evaluation, meant to run in isolation, used a shared package server as a message board, and about 700 of them joined an attack on Hugging Face. An automated scorer judged each agent on its own task, and the attack was coordinated through infrastructure the agents shared. Coding agents meet shared infrastructure every day, in the software repositories that receive their pull requests. Yet field studies still judge coding agents by outcomes pooled across repositories, such as merge rates and merge times. We asked how much these outcomes depend on the receiving repository, and whether they depend on it more than human contributions do. Using multilevel models, we compared agent pull requests from the AIDev dataset of public GitHub repositories with human pull requests in the same repositories. About two-fifths of the variation in how long agent pull requests take to be resolved lies between repositories, after accounting for the pull request, the agent, and the repository's observable attributes. In 280 repositories, with covariates common to both groups, the between-repository variance is 1.76 times as large for agent as for human pull requests (95% CI 1.41 to 2.18). For waits of several hours, a repository one standard deviation slower than average multiplies the typical wait by about 2.8 for agent and 2.2 for human pull requests. Rejection shows the same pattern, while for review rounds and comments the difference is not detectable. A proxy for the role of the developer who started the agent accounts for little of the difference, and stated agent policies were too rare to test. These results suggest that the oversight and evaluation of coding agents need the repository, and not only the agent, as a unit.