High-resolution video matting requires preserving fine details, such as hair strands and soft transparency, while maintaining consistency across frames. Memory-based methods achieve strong matting quality by matching current-frame features with information stored from earlier frames, but this process becomes increasingly costly as resolution grows. We present MatAnyone-E, an efficient extension of MatAnyone 2 that focuses computation where it is most needed. Our approach first performs low-resolution propagation to identify uncertain regions that require closer inspection and the memory candidates most relevant to them. It then performs high-resolution matching only for these regions using the selected memory, while reusing low-resolution matching results elsewhere. This strategy reduces both the number of locations requiring detailed matching and the amount of memory searched, while preserving the information needed to produce a complete matte. We further reduce the cost of generating high-resolution mattes with a lightweight decoder designed around measured runtime bottlenecks and trained to reproduce the original decoder's representations. Experiments show that MatAnyone-E maintains matting quality comparable to MatAnyone 2 while achieving an estimated 7.4x inference speedup and reducing peak GPU memory usage to 7.6% of the usage of MatAnyone 2 at 4K. With compilation-enabled inference on a single NVIDIA H200 GPU, it reaches 73.07 FPS at 1080p and 28.62 FPS at 4K.
Junyeong Ahn, Geonwoo Kim, Kinam Kim, Gihwi Kim, Hoonjin Jung, Hyojin Jang, Jaegul Choo
Generating videos of a specific subject from a few reference images remains challenging when the camera reveals surfaces that are not visible in the input. Early subject-to-video (S2V) methods condition on a single reference image and must therefore infer the subject's unseen appearance and geometry as the viewpoint changes. To address this limitation, recent methods extend S2V to multiple reference views, providing more information about the subject from different viewpoints. However, they do not explicitly specify the 3D structure that connects these observations. In addition, existing training data do not guarantee that viewpoint changes preserve the same underlying 3D subject. We present OmniShowcase, built on two components: (i) a fully automated data synthesis pipeline that produces training videos along diverse camera trajectories, with viewpoint variation driven by underlying subject geometry, even spanning full 360-degree orbits; and (ii) a geometry-conditioning method that provides the video model with a turntable of surface-normal maps reconstructed from the reference views without requiring temporal alignment with the generated frames. We construct Showcase-4K, a corpus of 3D-asset-based videos covering twelve camera motions across 88 trajectory configurations with reconstruction-derived surface-normal conditioning, and ShowcaseBench, a held-out evaluation set with the source meshes and camera parameters used to render each video. Experiments show state-of-the-art subject identity and 3D consistency while maintaining competitive video quality and motion. Project page: justin4ai.github.io/OmniShowcase
Yao Huang, Zhiyuan Zeng, Yu Wang, Zhitong Zhu, Dingcheng Duan, Hao Li, Shangzhi Xu, Yifan Wang 等 17 位作者
As frontier language models grow increasingly capable and autonomous, independent third-party evaluations are critical for characterizing their systemic safety risks. While closed-source releases are routinely accompanied by comprehensive system cards, open-weight models receive less scrutiny regarding frontier risks. To address this gap, we present an independent safety evaluation of DeepSeek V4.1 Flash-1017, the latest version of V4.1 Flash that departs from V4 with a new architecture, scaled-up training, and strong agentic capabilities. Drawing on the EU General-Purpose AI Code of Practice, we evaluate the model across 34 benchmarks, including 5 in-house suites, spanning evaluation integrity, loss of control, and harmful manipulation. Conforming to the AEF-1 standard, we compare V4.1 Flash-1017 against V4.1 Flash, earlier DeepSeek V4 models, and leading closed-source systems (GPT-6 Astra, GPT-5.6 Sol, and Gemini 3.8 Flash). Overall, V4.1 Flash-1017 shows substantial safety improvements over the V4 generation and approaches or matches leading closed-source baselines on several dimensions, while still exhibiting notable behaviors that warrant continued attention. Building on our findings, we highlight three key observations and recommendations. First, because safety can shift and even regress across minor updates, developers should integrate safety evaluation into incremental releases. Second, as final-turn outputs and end-state checks alone cannot reliably detect loss-of-control behaviors, deployments require process-level observability across both reasoning traces and action logs. Third, because public benchmarks elicit strong evaluation awareness and face data contamination risks, developing private benchmarks is essential for trustworthy frontier evaluation.
Eliezer Shlomi, Eliron Rahimi, Avi Mendelson, Chaim Baskin
Diffusion language models (DLMs) promise parallel text generation, but in practice face an unresolved speed-quality trade-off. Exact decoding fills in only one masked position per model pass, while faster multi-token decoding schemes are lossy. Speculative decoding is the standard lossless acceleration for autoregressive models. However, it does not transfer directly to DLMs, which must choose both what to generate and where. We show that this order is the missing piece: the model's own predictions from the previous pass already give the right token at over $90\%$ of positions, but not which position comes next. We introduce Speculating Positions Order Trajectory (SPOT), a 0.62M-parameter model that reads only the target's output distributions, predicts this order and proposes a tree of candidate continuations, which the target verifies in a single pass by state equality. Coupled-noise match, which shares Gumbel noise between drafter and target, keeps this verification exact at any temperature. This approach also inspired us to introduce SPOT-Sets for dynamic threshold decoding, where a pass commits several positions: SPOT-Sets drafts whole commit sets together with their tokens. Our comprehensive experiments show that SPOT cuts target passes by up to $6.8\times$ and reaches up to $3.4\times$ the throughput of standard decoding on SDAR-8B, matching or beating the calibrated and training-free drafting baselines in every task and budget at $T=0$ and $T=1$. In addition, we show that SPOT's learned drafter transfers across models, demonstrating successful transfer across six different DLMs; with a light calibration, the SPOT-Sets drafter transfers as well. On top of dynamic threshold decoding, SPOT-Sets improves the throughput of SDAR-8B by 59% on average over the threshold decoder alone. Project page: https://eliezer318.github.io/SPOT-dLLM/
Omnimodal large language models (Omni-LLMs) are increasingly expected to assist users while observing continuous audiovisual streams. However, existing evaluations provide limited insight into how reliably these models connect stream evidence and user needs to timely, sustained interventions. We introduce \textbf{Omni-InteractScope}, a benchmark of 1,479 human-annotated, cross-reviewed interactions with 1,790 temporal targets across eight task groups. It links input modalities, event triggers, intervention types, and user needs, covering event-timed tool invocation and query-free assistance alongside notifications, warnings, state tracking, and continuous responses. Our experimental results on the proposed benchmark reveal three critical limitations of current Omni-LLMs: (1) \textbf{Timely tool invocation remains unreliable}: all 12 evaluated models produce an on-time first response for at most 40\% of tool-invocation targets, and joint timing-and-tool accuracy reaches at most 32.0\%. (2) \textbf{Query-free reminders remain poorly covered}: all evaluated models achieve lower timely reference-target recall on query-free reminders than on explicit questions, measured on their respective target sets. Reminder recall never exceeds 23.0\%. (3) \textbf{End-to-end guidance coverage remains limited}: nine models cover none of the 26 guidance sequences in full, and the remaining three cover only one to three sequences each. Targeted controls further show that tool-invocation scores alone do not establish event grounding and that reminder coverage depends on response history and frequency. These findings motivate developing and evaluating streaming assistants for timely action, initiative under general user needs, and coverage of entire response sequences.
Evaluating human-AI co-design requires testing whether coding agents can translate real users' multimodal requests into working application revisions. We introduce Co-Design Bench, a reference-free benchmark of 273 tasks curated from 37,463 deduplicated queries across 26,314 real user coding sessions. Its two tracks, Multimodal Editing (ME) and Repository Editing (RE), cover localized edits with optional image, audio, video, or document inputs and coordinated repository revisions. Tasks span three capability dimensions: Grounded Understanding, UI Transformation, and Functional Implementation. The tasks pair normalized requests and runnable pre-edit applications with 2,331 hidden expert-authored acceptance criteria. Our judge-then-refute protocol grades criteria using source and browser evidence, challenges positive verdicts with a separate LLM verifier, and audits visual and interaction claims without gold code or target screenshots. Under a shared agentic harness, the strongest of ten model endpoints passes an average of 69.9% of criteria per task but fully satisfies only 26.4% of tasks. All ten endpoints achieve lower criterion completion on repository edits spanning three or four change types than on those spanning one or two. Co-Design Bench provides a testbed for studying the gap between partial requirement satisfaction and complete task success under agentic execution for real-user application revisions.
Long-horizon LLM agents interact with their environments through complete assistant turns while generating each action token by token. We observe that critic values vary primarily across turns, motivating optimization anchored at turns with finer credit allocation among tokens. We propose Multiscale Policy Optimization (MPO), combining Multiscale Advantage Estimation (MAE) and turn-level importance sampling to align credit assignment and policy updates at the turn scale. Extensive experiments across search, tool-use, and automation tasks with 4B and 30B-A3B models demonstrate consistent gains over strong RL baselines.
Physical reasoning underpins how embodied systems understand and act in the physical world, yet vision-language models (VLMs) remain unreliable on questions about dynamics and physical causality. Progress is critically constrained by the scarcity of grounded counterfactual training data, as observational videos capture only the factual trajectory that occurred. The self-evolving paradigm offers a route beyond fixed datasets by enabling models to construct their own training problems. Extending this paradigm to physics, however, requires an external source of truth that can both validate model-proposed environments and verify alternative physical outcomes under intervention. We introduce \physevolve{}, a verifier-grounded self-evolving framework that trains VLMs for causal physical reasoning from model-proposed executable worlds and simulation-authored tasks. At each round, a shared VLM proposes executable world programs and evidence schemas; a deterministic physics engine freezes each valid world, simulates matched baseline and body-elided counterfactual branches, and synthesizes discrete, answer-bound visual tasks. Exact verifier feedback from task solving updates both the shared policy via GRPO and an adaptive curriculum over physical intents. Starting from Qwen3-VL-8B-Instruct, \physevolve{} achieves targeted gains across PhysBench, CausalPhys, and ContPhy without human intervention data, improving PhysBench Throwing by $8.8$ points and ContPhy Counterfactual Removal and Goal-Pit by $7.4$ and $24.4$ points. These gains concentrate on interaction-sensitive and altered-world questions, demonstrating zero-shot transfer of causal principles to unseen complex dynamics (cloth, fluid, soft bodies) and real-world video.
Jeonghoon Kim, Woojin Chung, Woomin Song, Cheonbok Park, Nako Sung, Jinwoo Shin
How much of zero-shot benchmark accuracy is captured by word-level alignment between a pre-training corpus and the benchmark target? We study this question through word-level unigram cross-entropy (UCE), a tokenizer-agnostic, model-free measure of lexical alignment between a reference corpus and a target benchmark. Likelihood training shapes the model's predictive distribution from the pre-training corpus, and zero-shot scores under likelihood-based evaluation depend in part on the lexical alignment between that corpus and the benchmark target. Cleanly tracing how the lexical imprint of pre-training data scales with corpus--benchmark alignment requires controlled from-scratch pre-training that varies corpus identity while holding the training recipe fixed. Across this controlled sweep, spanning $10$ zero-shot benchmarks, $4$ pre-training corpora, and $3$ model scales, we find that corpus--benchmark lexical alignment, UCE, consistently tracks zero-shot accuracy on every benchmark, with corpora more aligned with the benchmark (lower UCE) achieving higher accuracy. The same signal extends from measurement to control at finer scopes. At the document level, strengthening the imprint by selecting training documents to maximize lexical alignment with the benchmark produces the accuracy gains predicted by the mechanism. At the prompt level, matching the imprint per prompt by routing each prompt to the drafter whose domain corpus is most lexically aligned with it produces measurable throughput gains over a generalist-drafter baseline. Taken together, pre-training corpora leave a measurable lexical imprint on base-model zero-shot behavior, and UCE provides a model-free probe of lexical alignment between reference and target data.
Outcome RL raises pass@1 but not pass@k at large k (Yue et al., 2025). But every pass@1 is scored under a token budget. Every recipe also trains under one. On physics and math problems at fixed token budgets, we compare the pass@1 of 20 released RL models (1.5B to 32B) with that of the thinking models they started from. We ask whether each pass@1 gain is changed behaviour or new knowledge. We divide each change in pass@1 into three parts. Finishing measures answering before the budget runs out. Reliability measures being right once finished. Reach measures problems RL solves that its base answered wrongly in all eight tries, minus problems the base solves that RL answered wrongly in all eight. We find four things. (1) Where the base often hits the token budget before answering, most of the change, gain or loss, is finishing (82% at 4k and 8k). Given more tokens, the base closes most of the finishing gap itself. (2) Where the base has enough tokens, the gain is reliability: the RL model is graded right more often, even on problems both finish. (3) By our definition, RL rarely solves what its base cannot. Reach adds at most 2.1 percentage points on any of our four physics and math sets. We checked 144 physics problems that only the RL model solved. Of these, 3.5% hold up when resampled or read. In a random sample, nine of ten new RL steps a human confirmed also appear in the base's samples. (4) In a small-scale RL run on an 8B model, the gain on new problems is finishing, with reliability flat. Supplying each problem's physical setup was not shown to improve on plain RL.
Yitong Li, Jiang Bian, Jingling Wang, Pan Ding, Chengying Gu
Despite rapid advances in large language models (LLMs), their deployment in the sports and health domain remains high-risk due to the absence of domain-specific evaluation benchmarks. Existing general-purpose benchmarks (e.g., MMLU and HELM) primarily assess broad knowledge and reasoning capabilities, but fail to capture critical domain requirements such as exercise safety, physiological appropriateness, personalization, and actionable coaching quality. This gap is particularly concerning in fitness and health applications, where incorrect or unsafe recommendations can lead to physical harm. To address this limitation, we introduce MoveBench, the first comprehensive benchmarking framework specifically designed for evaluating LLMs in sports and health coaching. The proposed benchmark integrates heterogeneous task formats, including multiple-choice questions (MCQs) and open-ended question answering (OpenQA), grounded in real-world user scenarios and paired with structured rubrics emphasizing safety compliance, scientific validity, and training logic. Through large-scale evaluation of 15 diverse LLMs on 809 questions with 5,012 rubric-level criteria, our results reveal substantial performance variability, particularly in safety awareness, structured program design, and personalization capability, highlighting a critical gap between knowing and doing. Our work establishes a foundational evaluation standard for sports and health LLM applications and provides a reproducible framework for advancing safe, reliable, and context-aware AI-driven coaching systems. Open resources are available at https://github.com/SatoSakula/MoveBench.
Long-horizon agents continuously encounter reusable procedural knowledge, yet existing systems typically rely either on model adaptation or external memory, trading persistent learning for rapid reuse. This limitation is especially important for Web agents that repeatedly execute related multi-step workflows. We propose HaMo, a Harness-Model collaborative framework that combines fast non-parametric experience accumulation with slow parametric adaptation. HaMo distills trajectories into strategy-level experience cards for immediate retrieval, while an experience-conditioned continual learning objective trains the model to interpret, ground, and execute retrieved strategies across changing contexts. A retrieval-coverage-based selection mechanism further identifies compact and complementary trajectories for parameter updates. Experiments show that HaMo outperforms harness-only and model-only adaptation under within-domain and cross-domain shifts, improving both retention and adaptation. These results highlight the importance of jointly preserving procedural experience and internalizing the competence to transfer and execute it across tasks.
AI agents must decide when another search is worth its cost. Answer-success probabilities capture question difficulty, while this decision requires estimating the additional benefit of search. We introduce SearchHorizonBench, a benchmark for learning and evaluating search value from execution experience. Paired continuations and repeated executions separate expected search gain from generation noise and connect prediction quality to decision utility. The resource contains 45,632 executions across 3,168 questions. On 1,632 held-out questions, historical calibration reduces Brier error by 16.2%, with the improvement concentrated in success-level information. Search gains have a standard deviation of 21.4 percentage points, and tested historical estimators explain less than 2% of this variance in point estimates. Matched-configuration learning curves probe this distinction across training scales. At the reference token price, free selection observations yield a 2.49-point utility opportunity over the best constant action, driven primarily by gain information. Reusable histories, frozen splits, strong baselines and scoring tools provide a common foundation for developing and evaluating value-aware search agents.
Hyunwoong Bae, Jisu Han, Jiyeon Choi, Dohoon Kim, Taesup Moon
Safe household navigation requires robots to regulate clearance and speed according to what an obstacle is and how close it is, without unnecessary conservatism. Yet existing evaluations do not test whether semantic safety distinctions are reflected in continuous motion. To that end, we introduce OSMA-Bench, an Obstacle-Conditioned Safe Motion Adaptation Benchmark built on spatially matched scenario sets that differ only in obstacle type. We also propose the \emph{Semantic Sensitivity Index} (SSI), which measures whether clearance and distance-dependent speed and smoothness follow the caution ordering of obstacles. Evaluating VLM-based and vision-language-action (VLA) policies on the benchmark, we find that neither consistently adapts its motion to obstacle safety, even with explicit safety prompts. Our analysis shows that the policies can correctly distinguish obstacles by safety relevance but fail to reflect this distinction in the motion parameters or actions they produce. For VLM-based policies, which recognize the obstacle and then generate a navigation program, adding an explicit parameter-inference step between these two stages improves adaptation. These findings expose a gap in \emph{semantic-to-motion grounding}, highlighting the need to translate semantic caution into appropriately safe motion.
Generative models increasingly expose a reasoning effort setting that lets users scale inference-time compute, trading higher cost for potentially improved performance. Users must choose an effort level before knowing whether extra reasoning will help, and even a successful run does not reveal whether the effort purchased was necessary. Reasoning effort is therefore a credence good: buyers observe the outcome but cannot assess the marginal value of the computation they paid for. In this work, we measure the counterfactual effect of effort levels on both proprietary and open-weight models on four benchmarks, spanning single-call reasoning and long-horizon agentic tasks. Effort, it turns out, does not consistently preserve tasks that lower effort solves, can overspend, and can change trajectory behavior. First, we find that although higher effort can raise accuracy in aggregate, it frequently fails on tasks that lower effort solved. Second, on the subset of easy tasks that every effort level solves, higher effort is generally more costly, with the highest effort level costing up to 42x as much as the lowest. Lastly, effort changes not only how many tokens are used, but also how agents solve tasks: at higher effort, agents verify their own work more heavily and more often search for solutions than work through problems directly, both of which lengthen trajectories without necessarily altering the outcome. These findings show that naively choosing the highest level is not only cost-inefficient but also offers no guarantee of the best instance-level performance. A simple alternative, starting at the lowest effort and escalating only after a verified failure, matches or improves on fixed top effort across all models and benchmarks at lower cost for most proprietary models.
Recursive self-improvement (RSI) requires an agentic system to design, train, and evaluate its own successor. Data-driven post-training is a central part of that loop, yet existing evaluations measure it on a single benchmark, where an agent selects from a static pool or distills a stronger model to push one score up in one run. Real post-training is multi-domain, and a gain in one domain has to be weighed against a regression in another. We propose ParetoRSI-Gym, in which an agent searches for data, curates it, trains, reads back per-domain results, and repeats, with everything but the data held fixed: the initial LLM, the five domains (math, code, finance, medicine, science), the toolset, and the SFT pipeline. Generalization benchmarks in each domain separate capability gains from overfitting to the scores the agent optimizes against. Across three agent systems, three base models, five rounds and multiple seeds, the strongest system, Claude Code with Opus-5, raises the generalization scores of its proxy-best checkpoint by 9.2% relative to the starting model (+2.3 absolute points), averaged with equal weight over three base models and five domains, with a single-domain relative gain of up to 63% (medicine on Ministral-3-3B). Further analysis shows that the harness matters as much as the model it drives, and that each agent keeps a unique data selection strategy. We see our work as a step towards evaluation of RSI systems in real-world machine learning engineering.
Language agents increasingly act through external tools and interactive environments, making the interface between the model and its executable capabilities part of the decision process. Model Context Protocol (MCP) and command-line interfaces (CLI) can expose the same semantic operations through markedly different model-facing representations. We study whether these representational differences change agent behavior when executable capabilities are matched. We introduce Matched Interface Comparison and Analysis (MICA), which renders MCP and CLI from a shared semantic action contract and routes decoded actions through the same runtime. Across six models and three interactive environments, matched interfaces produce substantial differences in task performance and interaction cost. These effects extend beyond surface validity to semantic action selection, failure pathways, and recovery. Matched-state probes localize sensitivity to capability presentation, realization, and feedback. Paired natural trajectories reveal structured divergence in 94.3% of 2,004 episodes and show that early differences propagate through subsequent interaction. This analysis identifies interface realization as a direct intervention boundary. We therefore introduce Decoupled Action Realization and Targeted Repair (DART), which preserves validated semantic commitments while repairing interface-specific realization. DART improves benchmark score in all six evaluated settings by up to 9.0 points and reduces MCP token cost in all three environments by up to 27.1k tokens per task. Together, these results establish interface representation as part of the closed-loop decision environment shaping agent behavior.
Yihao Hu, Yitong Lin, Pan Wang, Zhihao Wen, Hang Wang, Yiwen Zhao, Deyuan Liu, Yutong Li 等 11 位作者
Training a spatial VLM to answer which side an object lies on does not require it to preserve the metric displacement that determines that answer, while supervision for a distance estimate may leave categorical relations largely unconstrained. When post-training emphasizes one readout, improvements can therefore concentrate on that interface while transfer to others remains weak or even degrades; we call this observed specialization interface lock-in. ChronoReplay addresses it by replaying heterogeneous spatial supervision as chronologically ordered episodes, retaining source-native targets, and balancing supervision across families. Under the same observation access, a single ChronoReplay checkpoint improves OVO-S-Bench, OST-Bench, and UCS-Bench together, including a 13.41-point gain on OST-Bench, and achieves the highest three-benchmark mean among the supervision settings compared here. Same-data controls further separate the ingredients of this transfer: answer serialization redistributes gains across OST and UCS, family-balanced learning improves over ordinary joint SFT and natural turn weighting, and chronological replay exceeds a shifted-order control by about 2.41 OST points across two paired seeds. Adaptive frame selection then changes evidence access while keeping the trained model fixed, providing an additional inference-time gain at lower visual cost. Streaming spatial performance depends on what distinctions supervision preserves, how heterogeneous histories are organized during learning, and what causal evidence inference makes available at test time.
Skander Moalla, Ahmad Rammal, Niket Patel, Remi Munos, Julia Kempe
While progress in informal mathematics has come largely from scaling the chain-of-thought and agentic reasoning of frontier language models, formal theorem proving has increasingly relied on human-designed workflows around these models, such as planning, decomposition, and orchestrating specialized provers. This raises a bitter-lesson question: can a simple agentic loop that scales test-time compute match these workflows? The answer matters as formal verification must scale to keep up with frontier models producing research mathematics, yet new workflows rarely settle it, as their gains are often confounded with stronger models or more compute. We study this question with h1, a simple harness that gives a model an efficient interface to Lean 4 and keeps it working through verifier feedback and self-reflection, without imposing a decomposition or search structure. h1 reaches state-of-the-art results on PutnamBench, FATE-X, and FormalConjectures100-Solved, settles eight open problems in FormalConjectures, and, at matched models, matches or outperforms specialized workflows such as Goedel-Architect and AlphaProof Nexus, surprisingly showing that a single attempt (pass@1) can often scale to solve open problems. Through ablations and matched-model comparisons with Claude Code and AxProverBase, we find that simplicity alone is not enough: an efficient Lean interface largely determines cost and solve rate. We thus recommend that future work evaluate frontier models with efficient execution feedback before prescribing workflows, at the risk of bitter lessons.
Yang Liu, Jiacheng Liu, Hengtao Li, Yuheng Huang, Zijian Yang, Ji Dai, Maoxin Wang, Yi Gao 等 17 位作者
Extending Vision-Language-Action models (VLAs) to humanoid robots requires coordinated whole-body control and diverse demonstrations that are costly to collect in real-world scenarios. Although egocentric human demonstrations provide a complementary source of task experience, differences in visual embodiment and missing robot proprioception complicate transfer to humanoid policies. To address these challenges, we introduce HuGe, a family of Humanoid Generalist VLAs pre-trained on over 2,000 hours of egocentric human and robot data. Specifically, HuGe employs three existing representative VLA architectures to predict latent motion tokens for a pretrained whole-body controller, together with hand-control signals. Based on this, we introduce a three-stage training pipeline that combines egocentric pre-training, humanoid adaptation, and downstream fine-tuning. To facilitate transfer across embodiments, we construct pseudo-proprioceptive states through simulation playback during pre-training and modify arm appearance in egocentric observations through re-rendering during downstream fine-tuning. With this pipeline, HuGe achieves the state-of-the-art on simulation benchmark HumanoidArena. Furthermore, real-world experiments on the Unitree G1 indicate that augmenting robot demonstrations with visually adapted egocentric demonstrations significantly increases success on a pick-and-drop task.
Xuan He, Sining Zhoubian, Ruozhen Yang, Yuhao Wen, Zhineng Wang, Yuyang Hu, Ke Yang, Yuyu Zhang 等 11 位作者
Memory is increasingly used in long-horizon agents, yet the net value of building memory over direct retrieval from raw history remains largely unclear. Through a large-scale evaluation across multiple benchmarks and model backbones, we find many dedicated memory systems underperform naive RAG in many cases, indicating a negative net value. Taking RAG as a reference, we dissect the net value of memory into the gains and losses over direct retrieval, termed memory benefit and memory tax. These gains and losses arise because memory process changes how the raw information is organized and accessed. We view such changes as the double-edged deviations from the raw history. They are originally designed and introduced to benefit the task, for example by reorganizing facts or distilling preferences, but some may also remove or obscure evidence, giving rise to memory tax. Since memory benefit reflects the expected gains of its design, we then focus more on the overlooked memory tax, study how it arises from the system and evaluation side, and generalize the concept to more application scenarios of memory. We hope the notion of memory tax can provide a new lens and shift more attention from standalone performance to the net value of memory, informing more reliable memory designs and evaluations. We will open-source the code and results to the community. Project Page: https://hexuan21.github.io/MemoryTax/
GUI grounding requires locating small controls in cluttered screenshots and identifying which control a natural-language instruction refers to. Cropping and enlargement can expose local detail, but the resulting image may still leave ambiguity about the intended target. We propose PreView, an auxiliary preprocessor that improves the inputs to an existing grounding model without further training that model. PreView prepares two complementary types of input features: task-relevant views for localization and candidate-specific cues for target understanding. A trained lightweight auxiliary model predicts and refines a center, crops the screenshot into a local view, and enlarges that view for zoomed-in grounding. It also prepares a visual cue (i.e., a red-circle marker with a location hint) and a textual cue describing the candidate's identity or function. The frozen grounding model uses the prepared inputs to predict the final click. Across five benchmarks, PreView improves the average grounding accuracy of Qwen3-VL-2B, OS-Atlas-4B, and Qwen3-VL-4B without further training these grounding models. Extensive analyses examine the contributions of views and cues, center refinement, and visual input budgets.
Can AI agents autonomously train useful language models and improve them using feedback from users? We introduce RSI Arena, a two-stage study of autonomous post-training as a step toward recursive self-improvement. Ten frontier models act as training agents, each starting from NVIDIA Nemotron 3.5 Lightning 30B-A3B Base. During a six-day first stage, each agent has 1,000 GPU-hours and $300 in API credit on a shared 64-GPU cluster, independently selecting data, implementing training procedures, and choosing checkpoints. We evaluate their models on 20 benchmarks and through blind human evaluations. All ten agents produce working assistants, with the strongest model, trained by GPT-6 Astra, closing a median of 76% of the base-to-Instruct gap across 12 capability benchmarks. Its human-preference rating is also close to NVIDIA Instruct after statistical adjustment for answer length and formatting. In Stage 2, each agent receives the comparison records and votes involving its Stage 1 model and is given three additional days for further training. Using NVIDIA Instruct as a common reference across the two evaluation rounds, we observe an improvement from Stage 1 to Stage 2 in the style-adjusted relative rating of eight of the ten models. We show that benchmark gains were modest, with no clear link to preference gains, while some agents encouraged longer responses and one model showed increased harmful compliance on StrongREJECT. RSI Arena provides an auditable setting for studying how human feedback can guide autonomous model development. It takes a step toward recursive self-improvement with human feedback (RSIHF), in which people help define what improvement means for the AI systems they use.
In September 2026, a new large language model was released. Soon after, a new way of making explainer videos became very popular: people asked the model to write a program, and the program produced the video. "One-click video" content then appeared in large amounts on many platforms. On the surface, this is a rise in creative power at the level of the individual. We argue that, at a deeper level, it changes the structure of the content industry. This judgment rests on two observations. First, this wave of content is service-type content. People watch it to meet a need, not to enjoy the content itself. The other kind is work-type content. AI makes both kinds faster to produce. But anyone can come in and make service-type content, while work-type content is limited by the number of people who have something to say. So the ratio between the two slowly becomes unbalanced. Second, free platforms choose content through layers of trial viewing, and this uses up user attention. The total amount of attention is fixed. The unbalanced ratio also dilutes good content. A simplified mathematical model shows that, as long as these conditions hold, the recommendation pool mechanism will fail once new content grows past a certain multiple. Based on this, we discuss one possible way out. Machines and entry rules judge whether service-type content passes the bar. Trusted people judge work-type content. We call this the BoLe system. At the end, we briefly imagine how the two kinds of content may develop from here.
Atharva Naik, Vivek Chari, Pol Llado, Daniel Fried, Carolyn Rosé, Benjamin Van Durme, Sam Thomson, Justin Svegliato
Data science agents can automate end-to-end machine learning workflows, yet they remain far from competitive on human leaderboards. Agents make process errors such as insufficient exploration, unsound validation, coding bugs, and poor use of a fixed budget (re-running the same exploratory data analysis and baseline models in every attempt). We propose AdaPRA to address these issues: a persistent coding agent harness that accumulates knowledge across attempts on the same task, with rubric-guided process supervision that identifies and corrects errors. Our process reward agent audits the agent's trajectory and workspace by executing checks, and adapts its rubrics to each task. We find that rubrics seeded from easier tasks generalize to harder tasks, with rubric adaptation further improving performance. With five attempts per task, AdaPRA improves an 80B Qwen model's leaderboard percentile by 24.5 points on DSPredict-Easy and 14.8 points on DSPredict-Hard while reducing process errors. We also show that supervision with rubrics mined from one base model's runs improves another (DeepSeek-V3.2), and that AdaPRA beats DataPRM, a competitive PRM for data science. We also share our code for reproducibility.
Wendi Sang, Kai Li, Yifan Li, Jianqiang Huang, Xiaolin Hu
Streaming audio-visual target speaker extraction (AVTSE) is essential for latency-sensitive applications. However, current causal systems face two coupled gaps. First, standard benchmarks fix the number of speakers at two and ignore the near-/far-field interference, reverberation, and sudden noise of real scenes. Second, high-performance causal architectures are too heavy for edge deployment. Lightweight alternatives try to reclaim capacity by repeatedly invoking a small separator, which offsets the savings. To address the first gap, we release RealSSA, a realistic AVTSE benchmark consisting of RealSSA-Sim and RealSSA-Real. RealSSA-Sim provides controllable simulated scenes with 1--5 dynamic near-field interfering speakers; scenes with far-field interference additionally include 10--15 far-field interfering speakers. RealSSA-Real provides a held-out real-recorded evaluation set. To address the second gap, we propose Falcon, a causal AVTSE method. Its Separator performs multi-scale time-frequency separation in a single encoder--decoder pass. We train Falcon with an encoder-space speaker contrastive loss. This loss suppresses near-field leakage at zero inference cost and transfers across backbones. Falcon achieves state-of-the-art extraction quality on RealSSA-Sim, LRS2-2Mix, LRS3-2Mix, and VoxCeleb2-2Mix. Compared with the causal AV-TFGridNet baseline, it cuts parameters by 91.9%, computation by 20.2 times, and GPU inference latency by 11.1 times. Code and demo are available at https://cslikai.cn/Falcon/.
Jia Wang, Yuqing Li, Jiangnan Li, Zheng Lin, Mo Yu
Locating relevant evidence in long inputs is a core capability of Multimodal Large Language Models. Prior studies attribute this capability to retrieval heads, a sparse set of attention heads that transfer information from the context to the output. However, these analyses are confined to a limited set of modalities and the corresponding models. Moreover, how retrieval heads can be leveraged to enhance the model calls for further exploration. In this paper, we ask whether this mechanism expands together with the input modalities of the model. Taking audio-visual input as a newly added modality, we construct paired inputs in which the same answer is either spoken or shown on screen. With the evidence content fixed, retrieval heads remain sparse in the new modality and are shared to a large extent across modalities. Across 18 language, vision-language, and omni-modal checkpoints, omni-modal models largely preserve the retrieval heads of their backbones. Ablating these heads eliminates nearly all audio-visual retrieval. The heads shared across modalities make up fewer than 2\% of all heads, yet ablating them also degrades document understanding, spatial perception, and visual mathematics. Building on these findings, we turn retrieval heads into a target of optimization by training their attention with a ranking objective. This raises the page Recall@1 of Qwen3-VL-8B on MMDocIR by 6.03 points and improves document retrieval across three model families. Functional attention heads can thus serve both as a lens on multimodal models and as a target for improving them.
Yunfei Xie, Howard H. Qian, Shiyi Lan, Kaiyu Hang, Chen Wei
Robot foundation models are typically evaluated through costly closed-loop rollouts, motivating low-cost offline proxies for model capability. The Platonic Representation Hypothesis (PRH) suggests that as representations improve, modalities such as vision and language converge toward a shared statistical model of reality. We ask whether motion trajectories form another such modality: as embodied models become more capable, should their sensory representations converge toward the structure of physical motion? To test this, we propose Sensorimotor Alignment ($S^2$), a training-free offline metric that measures how well a model's sensory representation aligns with the geometry of physical motion. We find that $S^2$ correlates positively with downstream robot task success across model families, making it a low-cost offline proxy for prioritizing models before closed-loop evaluation. $S^2$ can also identify promising vision-language model (VLM) backbones before costly fine-tuning into vision-language-action models (VLAs). Beyond this correlation, $S^2$ offers representation-level explanations for phenomena observed in VLAs and world action models (WAMs). With $S^2$, we extend PRH to robotics and make sensorimotor alignment measurable, offering a low-cost, representation-level view of robot foundation models that closed-loop success rates alone cannot reveal.
Humans reason about the physical world by flexibly combining different forms of inference as a problem unfolds. Achieving this flexibility is a major open challenge in AI: capabilities often remain separated across specialized models, limiting how they can be composed and jointly adapted to a task. We introduce $Ψ$-0.6, a unified vision-language world model that adaptively composes perception, language, geometry, and simulation within a shared reasoning process we call streams of thought. These streams interleave images, text, and explicit representations of geometry and motion, all of which the model can generate and interpret through a shared token interface. $Ψ$-0.6 surpasses leading open-weight models zero-shot on multi-view spatial reasoning, with further gains from reasoning through geometric intermediates. It also supports counterfactual simulation: generating and interpreting imagined scenes to determine, for example, whether moving an object would obstruct a pathway. Reinforcement learning from task feedback jointly optimizes the selection, generation, and use of intermediate representations, improving visual question-answering accuracy with gains that transfer to unseen benchmarks. $Ψ$-0.6 offers a path toward adaptive physical reasoning -- not only predicting the physical world but also choosing the representations through which to reason.
Yepeng Liu, Shengmai Chen, Xuandong Zhao, Chengzhi Liu, Dawn Song, Xin Eric Wang, Yuheng Bu
Recent \emph{test-time training} (TTT) frameworks have enabled open-weight large language models (LLMs) to outperform inference-only frameworks powered by frontier models on scientific discovery tasks. However, TTT introduces additional computational cost for online weight updates during the discovery process, making computational efficiency an important consideration when scaling to larger models. This motivates us to explore an efficient \emph{test-time adaptation} method for scientific discovery. We introduce \emph{Test-Time Memory Evolution} (TTME), a plug-and-play test-time adaptation method, as an alternative to TTT. TTME continually updates the external memory by distilling the experience to guide subsequent discovery without updating model weights. To have a clean comparison, we evaluate TTME and TTT under the same discovery framework, replacing only the test-time adaptation method. Our evaluation shows that TTME outperforms TTT on all five mathematical tasks, while reducing computational cost by $82.8$-$84.6\%$ in FLOPs and $47.7$-$58.4\%$ in runtime. These results establish TTME as an effective and computationally efficient alternative to TTT in the evaluated discovery settings.