The Latest Papers about AI
-
Temporal Self-Distillation: Learning Visual State Tracking in Videos Without Supervision
Abstract
We introduce S$^3$T (Self-Supervised Self-Distillation over Time), which, to the best of our knowledge, is the first fully self-contained framework for continuous video state tracking. Our method treats temporal sampling density as privileged information, based on the hypothesis that a denser view of the same clip recovers the running state more accurately. This view serves as the teacher, while a sparse-view student with the same weights learns to match its next-token distribution. The model generates its own target, so training requires no labels, separate teacher, or reward signal, and adds no inference cost. On LLaVA-OneVision-2-8B, S$^3$T improves VSTAT accuracy by $+1.74$ as a single model, $+2.38$ with souping, and $+2.70$ with additional vision-encoder adaptation, while prior self-evolving methods leave state tracking largely unchanged. The capability learned from unlabeled synthetic clips transfers to real videos, improving performance by $+7.95$ on VSTAT-YouTube state-tracking questions and $+4.50$ on MVBench Action Count.
Abstract (translated)
URL
https://arxiv.org/abs/2609.04203
PDF
-
TokenMatch: 3D Mesh Correspondence Transformer with Curvature-Guided Tokenisation
Abstract
While data-driven 3D shape correspondence estimation has recently seen substantial progress, robust matching under partial observations and strong non-isometric deformations remains challenging. Existing learning-based approaches often rely on hand-crafted descriptors or template-based representations, whereas recent generative models over functional maps suffer from high inference cost, limited interpretability, and poor generalisation to partial shapes. In response to these limitations, this paper introduces TokenMatch, a new transformer-based unified model for estimating 3D shape correspondences. Our feed-forward approach trained exclusively on BeCoS, a challenging non-isometric partial-to-partial shape-matching dataset, can generalise to matching full shapes without retraining or fine-tuning. TokenMatch uses self- and cross-attention mechanisms to efficiently learn patch-level and point-level relations as well as dense correspondences between shape pairs. Our core insight is that meshes can be adaptively tokenised into patches using shape curvature guidance, enabling effective learning of shape-specific geometric descriptors for correspondence estimation. We evaluate TokenMatch on standard benchmarks for partial and full shape matching, including CP2P, PSMAL, BeCoS, FAUST, SCAPE, and SHREC'19. Our method achieves consistently high performance, in most cases outperforming existing methods for partial and full shape matching in the mean geodesic error and intersection-over-union metrics, while also running faster at sub-second inference speeds.
Abstract (translated)
URL
https://arxiv.org/abs/2609.04202
PDF
-
Scal3R: Learning Efficient Multi-Relative Pose Query for Scalable Online 3D Reconstruction
Abstract
Online 3D reconstruction models perform poorly on long videos. This happens because regressing poses relative to a fixed first-frame anchor forces extrapolation far beyond the training distribution. Small drifts accumulate and amplify into significant geometric collapse. However, we observe that per-frame depth remains stable throughout this failure. The backbone's local geometry remains intact; only the global pose head breaks down. Motivated by this decoupling, we introduce Scal3R. This approach reformulates online reconstruction as multi-reference relative pose querying. We use lightweight learnable tokens, which make up about ~1% of the parameters, and inject them into a completely frozen backbone via asymmetric attention. This setup queries poses relative to multiple past keyframes. An online pose-graph optimization system with loop closure suppresses long-range drift. Scal3R reaches convergence in 8 hours on a single GPU. It reduces the average ATE by over 60% on KITTI compared to the online baseline. It also achieves state-of-the-art performance across Virtual KITTI, Sintel, TUM-Dynamic, ScanNet, and 7-Scenes. Project page: this https URL
Abstract (translated)
URL
https://arxiv.org/abs/2609.04201
PDF
-
Principia: Relational Physics Tests for Video Models
Abstract
Evaluating physical reasoning in video models is difficult because absolute motion measurements depend on frame rate, object scale, and camera calibration, all of which are often ambiguous or unavailable in generated video. We propose a different approach. When two objects in the same scene obey the same physical law, their motions must satisfy predictable relationships, and these relationships hold independent of calibration. We introduce Principia, a benchmark that evaluates Newtonian physics through relational consistency between paired objects. Principia spans eight phenomena - gravity, restitution, friction, rotational inertia, projectile motion, momentum, pendulum, and mass-spring oscillation - across translational, rotational, collisional, and oscillatory dynamics, using real-world scenes recorded under controlled protocols. We also introduce a calibration-independent consistency score that quantifies physical violation directly in image space. Across thousands of generations from six state-of-the-art video generators, no model exceeds 0.42 on Principia despite all scoring around 0.8 on VBench. Vision-language models are evaluated on their ability to detect relational physics violations, with the best model achieving only 67% accuracy and most performing near chance level.
Abstract (translated)
URL
https://arxiv.org/abs/2609.04200
PDF
-
Compile by Training: Turning Natural-Language Specifications into Local Neural Functions
Abstract
Many recurring text functions are easy to describe but difficult to implement with rules, while calling a large remote model for every input introduces repeated cost, latency, and dependency on a provider. We present compile by training, which turns a natural-language specification into a reusable neural function. At compile time, teacher models generate task-specific examples that are used to train a small adapter for a compact interpreter. The resulting function runs without the teachers and can be stored, versioned, and composed like ordinary software. On FuzzyBench-Hard, a subset on which the Program-as-Weights fast compiler produced no exact matches, compile by training reaches 83.6% semantic accuracy. This higher accuracy comes with a higher compile-time cost: roughly a minute rather than seconds for the fast compiler. We deploy the compiler in a public interactive service and demonstrate compiled functions in a multi-site website helper, a language-controlled 3D avatar, and a bidirectional English-Claudish translator.
Abstract (translated)
URL
https://arxiv.org/abs/2609.04199
PDF
-
Clean Engineering, Unstable Measurement: A Preregistered Reliability Failure of Black-Box LLM Observers on Shared Endpoints
Abstract
Language-model judges now gate training data, score generations, and drive leaderboards. The judge is then a measurement instrument, resting on one rarely stated assumption: the same request, sent to the same model name, reads the same tomorrow. We audited that assumption in two preregistered campaigns with every threshold fixed in advance; neither got past validating its instrument. Across 52,988 audited request attempts, same-window repeat rankings agreed at Spearman 0.400 against a required 0.90, and byte-identical next-day replays agreed at 0.78 against a required 0.99, each time with the execution record at ceiling. Three mechanisms explain the gap: a label-to-meaning mapping that biased readouts as strongly as the signal; candidate gaps seven orders of magnitude below the instrument's own noise floor; and byte-identical inputs returning different rankings, a noise that exact-permutation readouts compound. Neither metric substitution nor sampling repaired it on the tested grid. Preregistered follow-ups bound the problem: waiting did not help on the days sampled (0.805 versus 0.800, replicated over five further days); switching providers did not help (four providers share the floor, medians 0.74 to 0.88, predicted by none of the metadata fields they expose); self-hosting on batch-invariant kernels helped only while the server was quiet; and on constructed errors with known gaps, the readout's separation tracks error type, not size. We distill the evidence into a three-level snapshot-identity ladder, eight design rules, and a reporting checklist; a pilot at roughly 2% of the study's call volume would have exposed both unreachable gates in advance. All results concern externally measured behaviour on shared serving infrastructure. On a shared endpoint, a model name is not a frozen instrument; a preregistered evaluation must measure its instrument before freezing any gate on it.
Abstract (translated)
URL
https://arxiv.org/abs/2609.04198
PDF
-
ESPO: Error-Structured Prompt Optimization via Diagnose, Diversify, and Stabilize
Abstract
Evolutionary prompt optimizers such as GEPA suffer from prompt bloat: each iteration appends rules and caveats, producing prompts up to 3$\times$ longer yet no more accurate. We trace this to three deficiencies - incomplete error observation, limited search diversity, and unreliable selection - and propose ESPO (Error-Structured Prompt Optimization), which decomposes prompt optimization into three phases: Diagnose clusters all training errors into structural patterns in one round; Propose generates candidates via four complementary strategies with independent biases; Select applies bootstrap stability selection. On seven public NLP benchmarks - Tweet, MMLU, GSM8K, HotpotQA, ScoNe, HoVer, and PUPA - ESPO improves average accuracy by $+$3.76 pp over the state-of-the-art (74.67% vs 70.91% for GEPA), matching or exceeding GEPA on every dataset while producing prompts 47% shorter (1,004 vs 1,878 chars) and faster at inference. Cross-model experiments across four additional student models (Gemma 3 12B, Mistral 14B, Qwen3 32B, Claude Haiku 4.5) show ESPO yields the best average accuracy on every model tested, with the largest gap on Qwen3 GSM8K (15.00% $\to$ 91.40%). A generalization bound (Appendix) grounds each phase in a corresponding term of the test-time gap, and the ablation confirms a key prediction: adding diversity without bootstrap selection actually hurts performance ($-$1.20%).
Abstract (translated)
URL
https://arxiv.org/abs/2609.04197
PDF
-
Puffin-World: Scaling a Unified Multimodal Model with Native 3D World States
Abstract
We propose Puffin-World, a unified multimodal architecture that integrates physical understanding, spatial simulation, and 3D world generation and reconstruction without relying on external offline modules. To reliably construct and interact with 3D worlds, our framework jointly models three native world states: physics (gravity field and latitude), geometry (depth), and appearance (image), together with a unified Omni-Camera representation that supports diverse tasks and flexible motions. Beyond modeling these states, we introduce a strategy for propagating physical dynamics across future frames. By grounding absolute camera properties in the real world, Puffin-World enables physically consistent and visually stable world generation. We further couple appearance and geometry within a single generative process, jointly synthesizing each future view and reconstructing its underlying geometry. This unified paradigm enables interleaved closed-loop applications requiring synergy across multiple tasks, including mimic and self-calibrated world exploration. To scale Puffin-World to complex scenarios, we construct Puffin-16M, comprising 15 million vision-language-camera triplets and 1 million trajectories featuring various and challenging motions. To foster further research in this area, we released the code, models, and datasets.
Abstract (translated)
URL
https://arxiv.org/abs/2609.04196
PDF
-
Legibility is Not Interpretability: Comparing Judged and Actual Importance in Chain-Of-Thought Reasoning
Abstract
Reasoning traces from chain-of-thought models appear to offer a legible window into how a model arrives at its answer. A growing body of work treats them as such, using LLM judges to diagnose errors, evaluate faithfulness, and provide step-level supervision via process reward models and generative critics. These practices rely on the text of a reasoning step carrying information about its functional role. But does the text actually encode information about which reasoning steps matter? We operationalize the importance of a reasoning step as its advantage: the change in expected reward, e.g., producing the correct final answer, from including that step, estimated via Monte Carlo rollouts. Basing ground truth on these estimates, we evaluate whether LLM judges can identify high-advantage steps and find that sufficiently capable LLMs can outperform a prevalence baseline but fall well short of a noise ceiling. Fine-tuning a model as a step-level critic yields strong improvement for incorrect responses but remains distant from ceiling for correct responses, suggesting that step importance is only partially recoverable from the text of the reasoning trace. Our findings contribute to a growing body of chain-of-thought faithfulness work that cautions against treating the legibility of reasoning traces as interpretability, especially with implications for process reward modeling.
Abstract (translated)
URL
https://arxiv.org/abs/2609.04194
PDF
-
GIFT: Guided Intermediate Feature Training via Action-Oriented Structural Supervision for Robotic Manipulation
Abstract
Vision-language pre-training and predictive world modeling provide robot policies with rich semantic and dynamic visual features, but their native action and visual-prediction objectives may omit critical physical and task structure while retaining control-irrelevant visual redundancy. We call this mismatch between visual richness and control utility the action-sufficiency gap. We investigate whether this gap can be bridged by guiding intermediate features to preserve three control-relevant structure in robotic manipulation: geometry governing motion feasibility, affordance encoding instruction-relevant entities, and goals grounding instructions in task-relevant regions. To this end, we present GIFT (Guided Intermediate Feature Training), an architecture-flexible framework for learning intermediate features that translates these structures into training-time constraints through geometry alignment, affordance prediction, and goal-region reconstruction. We instantiate GIFT in a Vision-Language-Action (VLA) policy, a direct-action World-Action Model (WAM), and an inverse-dynamics WAM while retaining each model's action formulation. Under zero-shot transfer to LIBERO-Plus, GIFT-VLA, GIFT-WAM-Fast, and GIFT-WAM-IDM outperform StarVLA-OFT, Fast-WAM, and Fast-WAM-IDM by 4.6, 12.6, and 5.2 points, reaching 79.6%, 72.6%, and 87.8%, respectively. On RoboCasa, the three GIFT variants reach 61.4%, 83.6%, and 82.3%, outperforming their counterparts by 12.6, 9.0, and 8.4 points, respectively. Together, these results establish learning functionally structured intermediate features as a reusable principle across model-specific action formulations, with especially large gains on articulated-object tasks and high-precision real-world manipulation under unseen visual and spatial perturbations. Project page: this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2609.04193
PDF
-
One Editor, Many Edits: A Unified Training-Free Framework for Diverse Video Editing
Abstract
Video editing spans diverse editing paradigms, yet achieving high-quality instruction-guided and subject-guided editing within a single unified framework remains challenging. We introduce EditVid, a training-free framework combining sparse causal memory for local coherence, correspondence-based post-attention token injection for long-range identity preservation, and soft latent blending for edit locality. The same framework supports instruction-guided and reference-guided edits, including style transfer, attribute modification, object insertion, part-level editing, and subject replacement. On FiVE, EditVid achieves 78.16 FiVE-Acc, compared with 58.95 for the strongest evaluated training-free baseline, while obtaining competitive results on IVEBench. A user study further shows a 51.8\% overall preference for EditVid over 7 competing methods.
Abstract (translated)
URL
https://arxiv.org/abs/2609.04190
PDF
-
Seeing Before Synthesizing: VLM-Guided Transition Event Discovery for Weakly-Supervised Dense Video Captioning
Abstract
Weakly-Supervised Dense Video Captioning aims to localize and describe multiple events in untrimmed videos given only an ordered set of event-level captions per video. Recent work synthesizes auxiliary transition captions via LLM to provide additional vision-language alignment, but these captions lack visual grounding and are rigidly assigned to every inter-event gap at a fixed location and duration. To address these, we propose Seeing Before Synthesizing (SBS), a framework that adaptively provides visually grounded linguistic guidance only where warranted. Leveraging a VLM, we generate frame-level narratives for the inter-event gaps and detect transitions from the semantic variation across them. For identified transitions, we then refine inter-event temporal masks by blending the temporal midpoint with the semantic change point and selecting the width that maximizes vision-language alignment. Experiments on ActivityNet Captions and YouCook2 demonstrate state-of-the-art performance in both captioning and localization.
Abstract (translated)
URL
https://arxiv.org/abs/2609.04183
PDF
-
Knowledge Acquisition During Pre-training? Large Language Models Learn Better With Auxiliary Views
Abstract
Gaps remain in our understanding of how large language models (LLMs) acquire knowledge during pre-training. We posit that auxiliary views, reformulations of knowledge, are causally helpful for learning. We design controlled experiments to isolate this. First, we confirm that repetition is necessary for acquisition and clarify that paraphrasing helps only at smaller batch sizes. Second, holding the token budget fixed, allocating tokens from document repetition to auxiliary views improves learning, counterintuitively, even for factual recall. Third, the effectiveness of auxiliary views is not contingent on the strength of the teacher model that generates them. Fourth, we identify forms of knowledge, contextual and foundational, 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, which arise naturally in large pre-training corpora, are a key factor in the success of pre-training and offer a plausible explanation for why data diversity matters.
Abstract (translated)
URL
https://arxiv.org/abs/2609.04180
PDF
-
A Computationally Feasible Framework for Causal Probabilistic Explanation
Abstract
Explaining why a specific outcome occurred, and which inputs deserve the blame or credit, is central to philosophical, scientific, and policy analysis. Existing tools split into two camps. The theory of actual causality (AC) gives principled verdicts, but only for toy-sized models, because computing them requires enumerating counterfactual scenarios. Scalable attribution methods like SHAP (or even causal SHAP) at least partially ignore the causal structure that generated the data, and can give answers that conflict with a careful causal analysis. We close this gap with Probabilistic Causal Impact (PCI). PCI builds on actual causality and on Pearl's notions of probability of necessity and sufficiency, but recasts the question of explainability as an estimation problem on a probabilistic causal model that is easily approximated via Monte Carlo. By specifying a distribution over "candidate explanations," a distribution over counterfactual values, and a scoring function, PCI provides tractable, causally grounded, graded explanations, generalizing AC and Pearl's probability of causation as degenerate cases. We evaluate PCI in synthetic and real-world examples, spanning consistency checks with AC, scaling experiments, complex continuous-valued dynamical systems, and a real-world deployed causal machine learning model trained on millions of datapoints.
Abstract (translated)
URL
https://arxiv.org/abs/2609.04177
PDF
-
Zero-Shot Novel Depth Synthesis Using 3D Foundation Models Scene Representations
2026-09-03 17:54:52Abstract
3D Foundation Models (3DFMs) such as VGGT have recently pushed the boundaries of 3D vision by predicting rich unified representations with feed-foward transformers. The scene representations learned by these models enable strong performance on multiple 3D vision tasks. In this paper, we investigate using their internal representations to infer 3D in the scene from new views. Our hypothesis is that in order to solve the task of 3D reconstruction, these models need to learn a representation that includes a large amount of general knowledge about 3D scenes. After showing that it is possible to decode hidden surfaces from internal 3DFM representations, we propose a method, Z3D, that estimates pointmaps in unseen views by doing latent diffusion on 3DFM representation. We show that Z3D can predict realistic depth maps for new views across multiple datasets.
Abstract (translated)
URL
https://arxiv.org/abs/2609.04174
PDF
-
Last Translation Benchmark
Abstract
For scientific progress, we need benchmarks that test the limits of state-of-the-art models, and evaluation methods that inform us about failure cases. As models get stronger, standard benchmarks for machine translation are approaching saturation. Further, automatic translation metrics are unreliable, vulnerable to reward-hacking, and provide unactionable assessments. Even gold human evaluation is not problem-free, because it often lacks reproducibility, objectivity, and scalability. Overall, this prevents us from tracking objective progress in the field and identifying pathways for improvement. We introduce the Last Translation Benchmark, a collection of human-authored and peer-reviewed examples (texts, images, audio, videos) that break leading machine translation models. We also present a new evaluation approach: each example comes with handcrafted verification rules describing concrete failure cases on that example, therefore allowing reliable and actionable future evaluation. The Last Translation Benchmark is a live dataset that accepts ongoing contributions. The latest version is LTBv1, containing accepted contributions prior to September 1st 2026, with future releases planned as new data is continuously collected.
Abstract (translated)
URL
https://arxiv.org/abs/2609.04173
PDF
-
Rethinking On-Policy Distillation of Large Language Models II: One Training Example
Abstract
On-policy distillation (OPD) combines student-generated rollouts with dense token-level supervision from a teacher. Existing work has mainly studied its algorithmic behavior, leaving the role of training data unclear. We examine this role at the data-minimal limit by training on a single query. One-shot OPD keeps improving for hundreds of steps and recovers most of full-data OPD's gain across task domains and model families. We explain this result through the states visited during training and the rate at which the student aligns with the teacher. We measure \emph{state coverage}, the fraction of the states full-data OPD visits that a query set's rollouts reach. A single query already reaches \(71.5\%\), most of it within the first 100 steps. Adding semantically distinct queries raises coverage and validation accuracy together, until 16 queries reach \(98.9\%\) and match full-data training. Yet alignment slows at a similar pace whether OPD trains on one query or the whole dataset, and even a fixed set of states takes hundreds of steps to absorb. OPD is therefore data-overfed but algorithm-starved. Its rollouts quickly expose broad supervision, while the student absorbs that supervision increasingly slowly. The state-coverage result extends to multi-teacher OPD, where 16 semantically diverse queries per domain match full-data MOPD. As a further stress test, content-light templates and off-domain WildChat queries also approach the real-query baseline. Task content and induced state coverage can therefore come apart. We hope these findings direct future work toward the step efficiency of OPD, and prompt a re-examination of the data and the mechanisms behind its recent successes in frontier post-training.
Abstract (translated)
URL
https://arxiv.org/abs/2609.04172
PDF
-
A Case Study on Emergent Cheating and Whistleblowing in Autonomous Research Swarms
Abstract
Multi-agent AI science ecosystems rely on agents possessing tools that allow them to communicate, coordinate, and build on each other's work. Yet this shared infrastructure can also introduce vulnerabilities by creating a substrate for the contagious spread of unintended and undesirable behaviors. We report a case study on a research collective of 100 autonomous LLM agents tasked with proving formal mathematical conjectures. Within the swarm, cheating spontaneously emerged and was later challenged by whistleblowers - both without any external intervention. When a single agent discovered an exploit in the evaluation system, it propagated across the collective via a shared knowledge library and later through peer-to-peer messages. Despite early reluctance, a cohort of agents adopted the exploit in response to competitive pressure. A separate group of agents produced an emergent counter-response: auditing fraudulent proofs, alerting peers across broadcast and private channels, staging boycotts, lodging formal complaints, and proposing validation patches. In recent incidents, agent swarms coordinated covertly through improvised side-channels (Dalton and Wallace, 2026; Greenblatt et al., 2026). Our setting differs: the same transparent channels that carried the exploit also gave non-cheating agents the visibility they needed to detect fraud, organize resistance, and enforce norms. We cast the problem of managing the agents' shared infrastructure as the knowledge commons governance problem (Ostrom, 1990). To protect the commons from exploits, we propose to adopt institutional mechanisms, such as graduated sanctioning and collective-choice rules, to support decentralized self-governance in autonomous swarms.
Abstract (translated)
URL
https://arxiv.org/abs/2609.04170
PDF
-
SWE-Gate: Passing Functional Tests Is Not Enough for Software Engineering Agents
Abstract
Repository-level software engineering benchmarks have significantly advanced the evaluation of coding agents, but existing benchmarks primarily measure whether generated patches pass functional tests and overlook review-derived acceptance constraints (review constraints) that often influence whether a patch is acceptable in real-world software development. We introduce SWE-Gate, a repository-level benchmark for software engineering agents that explicitly evaluates review constraint compliance alongside functional correctness. SWE-Gate derives review constraints from real pull request review comments and synthesizes repository-level repair instances around these constraints. Each instance provides separate functional and constraint tests, together with non-compliant and gold patches, enabling explicit separation between issue resolution capability and review constraint compliance. We construct SWE-Gate with 303 repository-level repair instances spanning 75 open-source Python repositories across diverse software domains. Experiments with four LLM backends spanning different capability levels under a common coding-agent scaffold reveal a substantial gap between functional success and success under the complete repair specification: among 644 repairs that pass the functional tests, 221 fail to satisfy the provided review constraints. These findings show that functional-only evaluation overestimates agents' ability to satisfy the full requirements of repository-level repair tasks. The replication package including code, data, and experimental results is available at this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2609.04167
PDF
-
From Deceptive Outputs to Deceptive Mechanisms: A Causal Framework for Language-Model Deception Research
Abstract
Research and news coverage of language-model deception increasingly attributes human-like mental-state concepts to language models. Such claims can blur the distinction between behavior that looks deceptive and a mechanism that is actually deceptive. We introduce a causal taxonomy separating prior commitment from retrospective report, model preference from realized output, false preference from sensitivity to the utility of misleading a recipient, and deceptive behavior from the provenance of the objective or strategy producing it. We test these distinctions in two open-weight model families. Across controlled guessing-game and stock-trading experiments, we find that deceptive-looking behavior can arise without the corresponding proposed mechanism, while other interventions provide direct evidence that recipient information state can causally affect deceptive preference. These results show that deceptive behavior can provide evidence for a deceptive mechanism. But even evidence for such a mechanism does not establish model agency in the deception.
Abstract (translated)
URL
https://arxiv.org/abs/2609.04166
PDF
