Vision-language-action models are typically built on a simple premise: perception is first compressed by visual encoders, and action reasoning begins only afterwards. Yet manipulation often hinges on precisely the visual signals most vulnerable to this abstraction, e.g., local geometry, boundaries, depth, and subtle state changes. This raises a fundamental question: should action policies inherit a fixed visual endpoint, or directly access evolving visual computation? We introduce NEO-VLA, a structurally encoder-free vision-language-action framework. Images are tokenized by lightweight patch convolutional layers and evolve jointly with language throughout the backbone, allowing a flow-matching action expert to draw from instruction-conditioned visual representations at different depths. This simple change reveals a consistent pattern. Native architectures outperform encoder-based counterparts under controlled comparisons; full fine-tuning reduces but does not erase the gap, and additional scale alone cannot recover it. More strikingly, layer-wise probing shows that the action-critical signals often peak well before the final visual representation. Across diverse manipulation settings and distribution shifts, NEO-VLA consequently achieves stronger and more robust control. Together, these results suggest that embodied control benefits from access to visual representations across the computational hierarchy, rather than a fixed encoder output.
Nicolas Tricard, Linzheng Wang, Jason Chen, Sili Deng
Reconstructing full flow fields from sparse sensors requires prior assumptions about unmeasured regions and physical variables. Yet it remains unclear when learned models outperform classical estimators and which aspects of their predictions can be trusted. FlowReconBench evaluates ten learned and five classical methods on three regimes: a laminar cylinder wake across Reynolds numbers, two-dimensional Kolmogorov turbulence and a large-scale three-dimensional (3D) isotropic turbulence. Methods share leakage-controlled splits and sensor draws and are scored on per-channel error, ensemble spread, calibration, spectra and cost. The leading method changes for each regime. Classical methods perform strongly in low-order and dense-sensing regimes. On the cylinder wake, linear stochastic estimation and gappy proper orthogonal decomposition lead, including for unobserved pressure. In 3D, inverse-distance weighting matches the best learned methods on observed channels. Data-driven methods become superior at decreasing sensor density or poor sensor observability. On Kolmogorov flow, continuous ranked probability score and error yield different method rankings: generative models offer broader but underdispersed ensembles, while deterministic models prioritize pointwise error. Generative models balance this tradeoff through their sampler budget, with a larger budget improving ensemble quality at higher computational cost. In 3D, scaling tests demonstrate that queryable methods are flat in memory, while full-grid construction methods run out of memory. Observing only 2 of 4 fields leaves the problem kinematically underconstrained, yet a latent model can reproduce structural similarities to the ground truth.
Fei Lin, Tengchao Zhang, Ziyang Gong, Xiaotong Yu, Bohan Zhang, Yifan Zhou, Qihao Yang, Ji Dai 等 18 位作者
Self-Driving Chemical Laboratories (SDCLs) are moving chemical experimentation toward embodied automation, where machines must perceive laboratory scenes, reason about experimental states, and act under physical, procedural, and safety constraints. Although Multimodal Large Language Models (MLLMs) are increasingly considered as embodied brains for such systems, existing laboratory benchmarks mainly focus on safety, anomaly detection, or general scientific understanding, leaving their execution-oriented embodied capabilities underexplored. To fill this gap, we introduce BEAKER, a Benchmark for evaluating Embodied Actionability, Knowledge, Experimentation, and Reasoning in SDCLs. BEAKER contains 1,000 expert-curated question-answer samples across 11 image and video task types, covering key judgments required for chemical laboratory execution, including affordance grounding, trajectory planning, equipment, state, and safety recognition, apparatus setup, operation and step reasoning, and temporal procedure reasoning. All questions and annotations are manually created and reviewed by chemistry and embodied Artificial Intelligence (AI) experts, without AI-assisted generation or annotation. We evaluate 28 general-purpose, science-oriented, chemistry-specialized, and embodied MLLMs in an offline, single-turn setting. Results show that current models can follow coarse experimental procedures in videos, but still struggle to ground actions into operable regions and feasible trajectories, to recognize fine-grained laboratory apparatus, states, and safety norms, and to order steps in mid-length procedures. BEAKER thus reveals a clear gap between coarse procedural understanding and the precise, action-grounded perception required for laboratory execution, providing a diagnostic benchmark for future MLLMs in SDCLs.
Bingran You, Wenjun Ke, Ria Rosenauer, Bo Liu, Wenbo Chen, Min Chen, Shuqi Xu, Yue Ma 等 72 位作者
AI agents are playing an increasingly important role in physics research. However, existing physics benchmarks only evaluate fixed questions, isolated sub-tasks, or paper reproduction with given solutions, and rarely check research quality alongside correctness. We introduce FrontierPhysics, an end-to-end benchmark with 56 tasks made by 42 physicists based on their own research, covering nine domains across theoretical, computational, and experimental physics. Each task requires agents to review the literature, develop a research plan, conduct research and experiments, and produce structured outputs and scientific manuscripts. Combining deterministic verifiers with expert-authored rubrics, FrontierPhysics evaluates whether agents' research is correct and publishable. Across 22 models in 26 configurations, no configuration passes more than 46 of its 168 rollouts. Manual review of agent trajectories shows that frontier agents' main weaknesses are insufficiently rigorous reasoning and limited expert intuition, rather than gaps in physics knowledge or execution errors.
Fei Lin, Tengchao Zhang, Ziyang Gong, Fei-Yue Wang
In recent years, generative artificial intelligence (GenAI) has demonstrated remarkable capabilities in high-stakes domains such as molecular science. However, challenges related to the verifiability and structural privacy of its outputs remain largely unresolved. This paper focuses on the task of molecular toxicity repair. It proposes a structure-private verification framework - ToxiEval-ZKP - which, for the first time, introduces zero-knowledge proof (ZKP) mechanisms into the evaluation process of this task. The system enables model developers to demonstrate to external verifiers that the generated molecules meet multidimensional toxicity repair criteria, without revealing the molecular structures themselves. To this end, we design a general-purpose circuit compatible with both classification and regression tasks, incorporating evaluation logic, Poseidon-based commitment hashing, and a nullifier-based replay prevention mechanism to build a complete end-to-end ZK verification system. Experimental results demonstrate that ToxiEval-ZKP facilitates adequate validation under complete structural invisibility, offering strong circuit efficiency, security, and adaptability, thereby opening up a novel paradigm for trustworthy evaluation in generative scientific tasks.
Toxicity remains a leading cause of early-stage drug development failure. Despite advances in molecular design and property prediction, the task of molecular toxicity repair, generating structurally valid molecular alternatives with reduced toxicity, has not yet been systematically defined or benchmarked. To fill this gap, we introduce ToxiMol, the first benchmark task for general-purpose Multimodal Large Language Models (MLLMs) focused on molecular toxicity repair. We construct a standardized dataset covering 11 primary tasks and 660 representative toxic molecules spanning diverse mechanisms and granularities. We design a prompt annotation pipeline with mechanism-aware and task-adaptive capabilities, informed by expert toxicological knowledge. In parallel, we propose an automated evaluation framework, ToxiEval, which integrates toxicity endpoint prediction, synthetic accessibility, drug-likeness, and structural similarity into a high-throughput evaluation chain for repair success. We systematically assess 43 mainstream general-purpose MLLMs and conduct multiple ablation studies to analyze key issues, including evaluation metrics, candidate diversity, and failure attribution. Experimental results show that although current MLLMs still face significant challenges on this task, they begin to demonstrate promising capabilities in toxicity understanding, semantic constraint adherence, and structure-aware editing.
Jalaj Bhandari, Dhawal Gupta, Hongli Zhan, Zhenting Wang, Varad Pimpalkhute, Zhengzhong Liu, Eric Xing, Rupesh Srivastava
In this work, we evaluate a simple inference-time scaling approach based on \textit{turn-level best-of-N sampling} (\method) for large language model (LLM) agents on complex, long-horizon tasks. We instantiate this using a \textit{generator-selector} framework with two LLMs; at each turn, a \emph{generator} model generates $N$ candidate actions and a \emph{selector} model selects one of them, which is then executed in the environment. This paper is closely related to the work of \cite{zainullina2025guided} who used critic models for guided search at inference time. Different from their work, we evaluate if off-the-shelf, open-weight LLMs can be effective at comparing different candidate actions directly. Experiments on long-horizon terminal tasks show that \method can lead to significant performance improvements, both in terms of accuracy (5--10\%) and consistency across repeated attempts. Analysis of agent trajectories reveal a shift in agent behaviour towards desirable patterns such as reduction in number of attempts to solve the task, improved task instruction following, debugging and testing.
Yangshen Deng, Xinjing Huang, Yulong Wang, Yunzhuo Liu, Jun Hu, Shuyong Tan, Chun-Chih Tseng, Mingji Han 等 11 位作者
We present HyScale, a long-context training system for learned sparse attention. Learned sparse attention is a promising architecture for million-token contexts, widely adopted in production since DeepSeek-V3.2. Selection, however, grows into the training bottleneck. For each query, it scores every previous token whereas sparse attention reads only $k$; worse, current systems materialize a full score table and repeat selection during recomputation. Our key insight is that the output of selection is not an activation, but rather a set of indices. Building on it, HyScale introduces FlashTopK and selection replay. FlashTopK finds each query's exact top-$k$ on the fly without a score table, while selection replay reuses forward-pass selection in recomputation. Together with holistic system optimizations tackling end-to-end challenges, HyScale speeds up GLM-5.1 training on 1M-token sequences by $2.07\times$ over Megatron-LM on 512 H20 GPUs. It also speeds up the CSA layers of DeepSeek-V4-Flash and the QSA layers of Qwen3.8-Flash-Next by $2.01\times$ and $1.57\times$ at 1M tokens on 32 H800 GPUs.
AI research agents automate the cycle of proposing, implementing, and evaluating experiments, opening a path toward recursive self-improvement. Yet their ability to propose experiments outpaces their capacity to execute them in real environments, making outcome prediction a key capability for sustained self-improvement under limited experimental budgets. We investigate language models as Research World Models (RWMs), which predict the outcomes of candidate interventions across research environments. Our evaluation draws on over 2,600 experimental records from nine research environments spanning pretraining, post-training, and inference, representing more than 171,000 H100 GPU-hours of experimentation. Research knowledge acquired from real experimental experience improves RWM predictions of unseen interventions within the same environment (Spearman +0.27), and can be reused across environments. For example, using only pretraining experience from OLMo3, Marin, and Nanochat, an RWM reduces selection regret in the Qwen3 environment by 78% compared with zero-experience setting. These benefits extend to multi-round Autoresearch under a fixed selection budget: RWMs with in-env and cross-env research knowledge increase the best gain achieved by 15.8% and 11.6%, respectively. Ablations across 13 language models used as RWMs show that adding research knowledge can improve intervention ranking more than changing models or increasing reasoning effort alone. These findings support language models as RWMs and motivate accumulating experimental data for future RWM training.
Recursive self-improvement requires AI systems to improve the models and infrastructure used by subsequent systems. LLM compression offers a concrete efficiency objective: retain a pretrained model's behavior within a smaller model storage budget. We study whether coding agents can carry out this step autonomously. We formulate compression as a terminal task in which an agent receives a source checkpoint, storage limit, inference environment, and time budget, then constructs compressed weights and executable inference code. The agent controls its algorithm and experiments; perplexity feedback guides search, and independent verification checks the final model. Across six agents and six model--size settings, 29 of 36 submissions pass verification. Selecting the lowest-perplexity agent design per task yields an average 3.4% perplexity increase over BF16 and 77.33% downstream accuracy, compared with 79.56% for BF16. Recorded experiments reveal how agents achieve these outcomes and where improvement breaks down. Agents lower PPL in 21 of 24 revisions at nearly fixed storage; complete-model feedback reverses 17 of 79 prior local rankings; and generation and inference tests expose failures that likelihood alone misses. The results identify representation design, experimental judgment, and capability validation as central to agent-driven model compression, a measurable component of efficiency-oriented self-improvement.
Haoxiang Cao, Yuxuan Zhang, Penghui Du, Bo Li, Jiajiong Cao, Yingying Fan, Zimeng Wu, Yifu Luo 等 19 位作者
We introduce LongBanana, to our knowledge the first expert-designed and expert-verified benchmark for long-context multi-reference image synthesis in the age of agent-driven image generation. The task requires retrieving dispersed visual evidence, rejecting distractors, binding entities and relations, and composing one coherent final image from multiple references. LongBanana contains 90 carefully designed tasks organized by a 3-by-3 taxonomy; each task includes about 32 reference images, 20 constraints, and 50 evaluation items, yielding 4,670 expert-specified atomic evaluation items in total. Its graph-grounded evaluation design instantiates each task with reference images and an entity-relation graph, then converts node, edge, and whole-graph obligations into checklist items for Content Fidelity, Relational Binding, and Global Constraint Satisfaction. Reviewed expert labels include a 15% double-labeled subset with 89.4% exact agreement and 0.83 quadratic weighted kappa. We further introduce LongBanana-Gen, an agent-harness framework for long-horizon image generation; experiments show that minimal agent harnesses improve over direct generation, while a ReAct-Plan-Exec-Reflect (RPR) workflow yields harness-dependent gains by strengthening evidence grounding and repair. Together, LongBanana provides an auditable testbed for measuring dense visual evidence use and a shared target for developing stronger agent-driven image-generation systems. Code and data are available at the project page: https://long-banana.com/.
A central yet under-explored question in Multimodal Large Language Models (MLLMs) is where to connect---how visual information from a hierarchical vision encoder should be wired into the layer-wise semantics of a large language model. Mainstream MLLMs inject a fixed single-layer visual representation into the LLM and treat all LLM layers as a single consumer, ignoring that different layers demand different visual granularities. Recent multi-layer fusion methods either compress hierarchical features on the encoder side or rely on predefined sparse or hierarchical layer-to-layer connections, and therefore stop short of modeling adaptive cross-layer interactions between the ViT and LLM hierarchies. We address this gap with DGP (Dynamic Gated Pathways), a routing mechanism that establishes all-to-all connections between every ViT layer and every LLM layer, with each LLM layer dynamically gating its preference over visual layers conditioned on its own semantic state. Built on top of LLaVA-1.5-7B and 13B, DGP delivers consistent gains over the LLaVA-1.5-7B baseline (e.g., +4.2 on SQA$^{I}$, +3.3 on MM-Vet, +3.1 on VizWiz and LLaVA$^{W}$) and surpasses prior multi-layer visual fusion methods on most benchmarks. Layer-wise routing analyses further show that different LLM layers do select different visual granularities on demand, supporting both the effectiveness and interpretability of the proposed dynamic gated pathways. The code is available: https://github.com/YingyingYan1/DGP.
Large language model (LLM) coding agents develop software through experimentation, offering an opportunity to design the serving systems supporting their own operation. Effective serving requires coordinating implementation and operating decisions under session dependencies, latency constraints, and correctness requirements. We introduce ServerBench, a benchmark for agentic design of complete, specialized LLM serving systems. Agents experiment against a common objective, dollars per dollar: latency-qualified token value per dollar of GPU cost. Stateful replay preserves session dependencies, and independent verification checks the submitted system. Across 11 coding-agent models and three deployments, Claude Sonnet 5.5 achieves 2.97 dollars per dollar on Gemma E4B, $4.87\times$ the vLLM reference; 22 of 33 runs yield verified systems, and 13 exceed the strongest baseline. Our analysis identifies three findings: high-value designs coordinate capacity improvements with load selection; agents interpret performance feedback through workload-relevant experiments; and reliable delivery requires verification of the execution paths introduced by optimization. Overall, ServerBench supports systematic study of how agents translate service objectives and feedback into effective, verified serving systems for given task specifications.
Autoregressive image generation has recently emerged as a competitive visual generation paradigm, yet its generation order is often predefined or heuristic, such as next-patch prediction, random-order prediction, or one-step generation for specific scales. We observe that these strategies can cause models to overlook important structural dependencies in images, thereby limiting both structural robustness and reasoning capability. In this work, we propose a context-aware autoregressive generation framework, where the model dynamically determines the next token location based on global context and current local information. Rather than following a fixed or random order, our method allows the generation process to adapt to image structure and semantic dependencies. Extensive experiments show that context-aware generation significantly improves structural robustness across different image AR paradigms, including masked autoregressive and scale-wise autoregressive models. Beyond quantitative improvements, we further observe emerging reasoning properties: the model can better understand complex semantic instructions and generate images that more faithfully follow the given guidance. These findings suggest that generation order is a critical yet underexplored factor in image autoregressive modeling, and that context-aware decoding can unlock stronger reasoning-aware visual generation capabilities. Code is available at https://github.com/Hpesojyjx/Saccades-AR.
Juncheng Ma, Jiaxin Li, Jiankai Tu, Bo Liang, Guodong Lv, Yufan Deng, Junzhe Tang, Liangzhu Yang 等 31 位作者
Egocentric human video offers a scalable source of manipulation priors for robot learning, yet translating raw data scale into generalizable robot policies demands rigorous data curation and training recipes. We introduce Open-HumanScale, the first fully open recipe for scaling egocentric pretraining to 50,000 hours, anchored by Open-HumanScale-50K---a curated egocentric video corpus densely annotated with hand motion and natural language. Leveraging this corpus, we systematically investigate pure human pretraining via closed-loop simulation evaluations and uncover an empirical log-linear scaling law linking pretraining data hours to downstream robot success rates. Grounded in these findings, we develop HumanWAM-Vanilla, a world--action model pretrained exclusively on 50k hours of egocentric human video. Upon robot post-training, HumanWAM-Vanilla reaches 92.08% success on RoboTwin 2.0 Clean2Clean and 68.88% on Clean2Random, alongside a 14.17% mean success rate and a 19.00 mean progress score on RoboDojo. To push the capability frontier further, we explore joint human--robot pretraining with HumanWAM-Goal, a visual-goal-conditioned world--action model. Hybrid data scaling unlocks broad zero-shot generalization across both simulation and physical benchmarks, propelling post-trained RoboTwin 2.0 performance to new state-of-the-art levels: 92.70% on Clean2Clean and a substantial leap to 84.60% on Clean2Random. On physical robots, HumanWAM-Goal further exhibits emergent embodied in-context learning, dynamically adapting its manipulation behavior to unseen visual goals without parameter updates. To support reproducible scaling research in embodied AI, we open-source the Open-HumanScale-50K corpus, data curation and pretraining pipelines, alongside the pretrained and post-trained HumanWAM checkpoints, inference code, and simulation evaluation suites.
Diffusion and rectified flow (RF) models generate high-fidelity images and videos, but their iterative velocity-field evaluations are computationally expensive. Existing caching methods accelerate sampling by skipping timesteps, yet their coarse approximations introduce accumulated errors over long skip intervals and degrade quality under aggressive acceleration. We propose TACache (Trajectory-Aware Cache), a training-free acceleration framework following a skip-then-compensate paradigm. TACache performs an orthogonal decomposition of discrete velocity acceleration along the RF trajectory into a parallel component and an orthogonal residual, isolating the magnitude and directional sources of per-step approximation error. The framework operates in two stages: offline, cumulative variation thresholds on the magnitude and direction indicators yield the skip schedule and bound how far each skip interval may extend; online, at each skipped step the offline statistics are combined with the sample's historical orthogonal direction to reconstruct the skipped velocity without additional model evaluations. Main comparisons on BAGEL, FLUX.1-dev, and Wan2.1-1.3B show that TACache achieves up to $4.11\times$ speedup on text-to-image generation and $2.11\times$ speedup on text-to-video generation, with consistent improvements over prior cache-based methods on most quality metrics. Our code is available at https://github.com/XiaoLiu1237/TACache.
Lingfei Qian, Yan Wang, Xueqing Peng, Yupeng Cao, Yuechen Jiang, Zhuohan Xie, Anke Xu, Wenbo Cao 等 15 位作者
Financial backtests of large language models (LLMs) are meaningful only if the model did not already know the outcomes it is asked to predict. Practitioners guard against this by trusting a vendor's stated knowledge cutoff, by asking the model what it knows, or by instructing it to answer as of a past date. To verify whether any of these safeguards can be trusted, we introduce FinDendro, a benchmark of 2,226 dated U.S. facts from 1999 to 2026, covering monthly macroeconomic series, month-end stock prices and Federal Reserve decisions. Each fact yields four prompts: a value question, Recall and Predict questions that pair each candidate value with a claim of knowing or guessing without a declared date, and a Predict question with a historical declared date. Controls for persistence and dominant outcomes keep chance accuracy near 0.27. Results show that for most frontier models, self-reports understate knowledge. GPT-5 answers 76% of controlled facts correctly but claims to know 38%, and answers 61% of the facts it says it does not know. Knowledge claims can persist beyond the measured knowledge boundary, by up to seven quarters, and a declared date removes most of the claims of prior knowledge while leaving accuracy unchanged. Backtests inside a model's training window therefore cannot be certified leak-free by self-report, instruction or stated cutoff. These findings motivate behavioral measurement of temporal knowledge boundaries, and FinDendro provides a benchmark and procedure for selecting backtest periods based on demonstrated factual knowledge.
High-quality teleoperation datasets are costly to collect, particularly for hard tasks. We observe that many tasks exhibit directional asymmetry: completing the forward hard task is difficult, whereas reversing it by relaxing or disrupting the environment is comparatively easy. This suggests that reversed easy-task trajectories can serve as a scalable supervision signal for the hard task, reducing the cost of manual demonstration collection. However, reversed data can be noisy, and directly training on it may yield suboptimal policies. To enable largely automated acquisition and effective use of reversed data, we propose a teleoperation-cost effective framework for hard policy learning via temporal reversal of easy tasks, consisting of three key components: a closed-loop data collection pipeline that alternates between hard-task and easy-task policies to autonomously reset the environment and generate diverse trajectories; a hierarchical data refinement pipeline that temporally inverts easy-task rollouts and filters low-quality motion using kinematic priors and a critic-guided advantage filter; and an iterative policy learning method that trains the hard-task policy using both initial reversed easy-task demonstrations and the filtered reversed data in a continuous online learning loop. By combining automated collection, hierarchical refinement, and iterative learning, our method enables scalable, reliable training of complex, high-precision manipulation tasks. Across two simulated benchmarks and real-robot experiments, we demonstrate that our method improves hard-task success rates with higher data efficiency and more stable training compared to reversal-based and reinforcement-learning baselines, without requiring extensive hard-task teleoperation.
Objective: To develop the Cardiologist Annotation Tool (CAT) for frame-level cardiac view and image-quality annotation in fetal echocardiography and compare it with spreadsheet-based annotation. Methods: CAT provides cine playback, frame navigation, and keyboard-based labeling without manual frame-range entry. Experts annotated 25 studies using both methods, and spreadsheet annotations were audited for predefined data-entry errors. In a randomized crossover study, six medical students annotated two normal studies, one with each method, against an expert reference. Results: Among 1,746 spreadsheet-annotated files, 6.2% contained major or critical data-entry errors, more frequently in congenital heart disease than normal studies (10.9% vs 3.4%). CAT yielded more annotated view segments per clip (3.6 vs 1.5). Rates of labeled and correctly labeled frames per minute were 1.77 and 1.68 times those with spreadsheets, respectively (both P = 0.031), with gains in all six students. No detectable difference in view-label accuracy was observed. Conclusion: CAT eliminates manual frame-range entry and enables finer-grained, structured annotation of fetal echocardiography, providing a practical approach to preparing datasets for AI research.
A fast acceptance classifier could select revisions for a slower evaluator. We ask whether agreement on existing papers supports that use. We compare a classifier of word and document features with two released text-only acceptance predictors fine-tuned from the same language model with three billion parameters. For 198 papers from the International Conference on Learning Representations, we compare model rankings of the original papers and of paragraph rewrites screened automatically for unchanged facts, claim strength, and attribution. The classifier ranks original papers closely with one predictor and moderately with the other, but shows no detectable agreement with either on which rewrites raise scores. The tests specified in advance used reduced numerical precision. Rounding error dominated the small score responses. We therefore rescored the released weights in full 32-bit precision as a post hoc check; edit agreement remained near zero. At this precision, when each predictor chooses among five rewrites per paper, its average raw score is under two points out of 100 higher than with random choice. The evidence covers presentation rewrites and two open predictors from one family; transfer to costly evaluators remains untested. Revision selectors should therefore be checked against the intended evaluator on the revisions they rank.
Reinforcement learning with verifiable rewards (RLVR) often adopts GRPO-style group-relative updates, sampling multiple rollouts per prompt to construct normalized learning signals. However, merely increasing the number of rollouts does not reliably strengthen learning: under GRPO-style group normalization, per-rollout policy-gradient features can concentrate into a low-rank, signed geometry, causing substantial cancellation during aggregation and weakening the effective update. We address this failure mode with SALT, a Subspace-Adaptive geometry pLug-in componenT that uses sample-wise gradient geometry to reweight the coefficients of group-relative updates. SALT estimates a dominant shared subspace from the mini-batch Gram geometry, decomposes group-relative coefficients into shared and residual channels, and adaptively amplifies the residual channel when signed cancellation is severe. Across diverse reasoning-oriented RLVR benchmarks and model scales, SALT improves effective update geometry and performance without modifying the reward model or the rollout sampling procedure.
Coding agents powered by models such as GPT-6-Astra connect spatial reasoning with executable robot behavior, raising the question of how such agents can produce lasting improvements in robot capabilities from physical interaction. In this work, we study embodied recursive self-improvement (RSI) through real-world exploration, code refinement, and deployment using PyRUA-REAL, a framework in which coding agents retain interaction experience as persistent memory and reusable executable skills that inform subsequent exploration and improvement. Our empirical study places physical manipulation at the center of the improvement process, examining how agents collaborate, what kinds of execution feedback support effective code refinement, and how task specifications and previously acquired skills influence robustness and task transfer. We further investigate the efficiency and capability of the resulting programs when deployed without involving an online agent for decision-making. Across eight real-world tasks, PyRUA-REAL achieves an average success rate of 91.1%. In complementary evaluation on 42 tasks from the RoboDojo benchmark, it achieves an average success rate of 84.4%. This study provides empirical guidance for turning physical exploration into reusable robot skills that support continued improvement and robust, efficient deployment.
Yuchi Wang, Rongyao Fang, Weihao Yu, Xilin Wei, Tiancheng Gu, Dingkang Yang, Hang Zhang, Dayiheng Liu 等 10 位作者
Beyond basic tool use and coding skills, effectively incorporating visual references into agentic generation is equally essential for applying AI agents to real-world productivity tasks. We study this crucial capability through webpage development, a practical and readily evaluable setting in which existing benchmarks remain largely idealized: they typically focus on reconstructing visual targets or generating webpages from textual instructions alone, overlooking the more common workflow of developing from visual examples. Specifically, we introduce RefCraft, a benchmark comprising 300 realistic multimodal webpage-development tasks that combine natural-language instructions with screenshots from one to five reference webpages. Agents must selectively interpret and utilize these references while rendering, visually inspecting, and iteratively refining their outputs. We evaluate the resulting open-ended artifacts using independently verifiable keypoints and complementary programmatic and VLM judges. Across nine frontier (M)LLMs, the strongest system achieves an overall score of only 52.12 out of 100, revealing that visual-reference-conditioned agentic generation remains a substantial challenge for current models in practical webpage-development settings.
Shahan Ali Memon, Anna-Maria Gueorguieva, Anukriti Kumar, Lucy Lu Wang, Jevin D. West, Nic Weber
In the years following the public release of ChatGPT in late 2022, more than half of U.S. states have published formal guidelines around AI usage in K-12 education, creating a record of largely independent state education departments responding to a shared technological disruption. In this work, we analyze publicly available AI usage guidance documents from 35 states in the U.S. and Puerto Rico. We compare the guidance in these documents across various dimensions, including pedagogy and instruction, audience, actionability, workforce and career readiness, responsible use and equity, and alignment with state-level demographics and administrative capacity. Finally, we investigate who authored these documents and their cited sources, identifying epistemic foundations underlying their claims. We find that guidance priorities converge on topics anchored in laws such as data privacy and academic integrity, each addressed in 97% of documents. While most documents mention equity of access (86%) and algorithmic bias (97%), fewer discuss specific marginalized populations. Across documents, there is a common framing of urgency around AI integration to prepare students for workforce readiness (86%) and promote loosely-defined personalized learning (86%), which is not matched by specific guidance for capacity-building, actionability, or accountability. State priorities show no relationship to state demographics, capacity, or partisanship. Instead, we identify notable interdependencies between states with shared citations to non-profit organizations like TeachAI and Code.org, with some states reproducing TeachAI's model guidance near-verbatim, and peer citations concentrating on a few early states such as Oregon and Washington. No state includes parents as contributors on these documents, and only five include students.
Shijue Huang, Hangyu Guo, Guanting Dong, Chenxin Li, Junting Lu, Xinyu Geng, Zhaochen Su, Zhenyu Li 等 11 位作者
Multimodal deep search requires an agent to solve open-world problems by chaining search, tool use, and visual reasoning over evolving textual and visual context. Two bottlenecks limit current systems. First, existing tool-use harnesses treat images returned by search, browsing, or transformation as transient outputs, so intermediate visual evidence cannot be re-consumed by later tools. Second, training data is usually built by fixed curation recipes that cannot track the target agent's evolving capability. To address these challenges, we first introduce a visual-native agent harness centered on an image bank reference protocol, which registers every tool-returned image as an addressable <image:N> reference and makes intermediate visual evidence reusable by later tools. On top of this harness, On-policy Data Evolution (ODE) runs a closed-loop data generator that refines itself across rounds from rollouts of the policy being trained. This per-round refinement makes each round's data target what the current policy still needs to learn. The same framework supports both diverse supervised fine-tuning data and policy-aware reinforcement learning data curation, covering the full training lifecycle of the target agent. Across 8 multimodal deep search benchmarks, ODE improves the Qwen3-VL-8B agent from 24.9% to 39.0% on average. At 30B, ODE raises the average score from 30.6% to 41.5%. Further analyses validate the effectiveness of image bank reuse, especially on complex tasks requiring iterative visual refinement, while rollout-feedback evolution yields more grounded SFT traces and better policy-matched RL tasks than static synthesis.
Action-conditioned robot world models must respond precisely to robot trajectories while preserving realistic visual dynamics, yet learning both from heterogeneous robot videos remains challenging. Simulation offers structured motion supervision, but appearance differences hinder direct transfer, and inaccurate simulation predictions can misguide real-video generation. We present SimForcing, a simulation-guided framework that uses simulation both as a source of transferable motion knowledge and as a controllable reference for prediction. First, we transfer motion knowledge from a simulation teacher through latent-motion distillation, aligning temporal changes in latent space to internalize motion priors while mitigating the influence of appearance differences. Second, we introduce multi-block simulation conditioning with condition dropout to exploit predicted simulation trajectories without relying excessively on their accuracy. Our simulation-conditioning classifier-free guidance scheme unifies these two ideas by balancing predictions based on internalized motion knowledge with those additionally guided by simulation latents. The jointly trained student generates both simulation conditions and real-domain videos, requiring no additional world model at inference. On Bridge, SimForcing achieves the best PSNR, SSIM, LPIPS, and FVD among the compared methods without external embodied pretraining. Evaluation on InternData-A1 further supports its applicability across robot datasets. Moreover, using our trained world model to initialize a vision-language-action model improves LIBERO success, suggesting its utility for downstream policy learning. https://github.com/Wang-Xiaodong1899/SimForcing
Strong single-turn editing performance does not guarantee reliable multi-turn interaction: as edited outputs are repeatedly reused, small unintended changes can accumulate into severe visual drift. This long-term drift is often mitigated through training on multi-turn data or cross-turn rollouts, potentially altering pretrained priors, or through specialized inference-time mechanisms that are difficult to scale to foundation models. We introduce TailSteer, a final-stage training framework that improves multi-turn robustness while retaining the generative and editing capabilities acquired through earlier training. TailSteer follows a simple principle: keep early denoising dynamics close to those of the base model and concentrate preservation corrections near the trajectory tail, where fine-grained deviations can be suppressed. These corrections are learned from the editor's own on-policy denoising rollouts, using unchanged source content as a preservation reference without requiring curated ground-truth editing pairs. Experiments across the FLUX, SenseNova, and Qwen families show that TailSteer substantially reduces cumulative drift over 15-20 editing turns while largely preserving base-model performance on extensive text-to-image and single-image editing benchmarks. These results demonstrate that long-horizon editing robustness can be unlocked in foundation models after TailSteer, rather than by relearning their existing capabilities.
Streaming video understanding requires models to repeatedly decide whether to respond or continue observing. Recently, Group Relative Policy Optimization (GRPO) has been applied to this setting, with existing approaches emphasizing reward design while retaining a single trajectory-level advantage across turns. Such shared credit can reinforce redundant replies after earlier successes and discourage useful follow-ups after earlier failures. We propose TurnWise, a turn-level GRPO framework that aligns learning with individual streaming decisions. TurnWise addresses the mismatch between trajectory-level feedback and turn-level interaction, promoting informative responses and reducing redundancy. Specifically, turn-level credit assignment uses silence-completed prefix scoring to construct group-relative advantages from each decision's remaining return, separating its learning signal from earlier rewards. Turn-mean policy optimization averages token losses within each turn and then across turns, giving response and waiting decisions equal weight regardless of output length. These complementary designs improve streaming policy learning without a learned critic or additional rollouts. Extensive experiments show our method outperforms prior RL approaches for streaming video understanding.
Training an agent to play Pac-Man with only up, down, left and right actions raises a basic question: when performance is poor, what should be improved? Difficulty can arise from reading the screen or from choosing actions over a long sequence. We examine these problems in a Qwen3.5-9B-based agent through separate perception checks and primitive-policy reinforcement-learning studies. In the policy study, which uses textual boards to remove image transcription from the decision interface, completion on eight held-out mazes increased from 0% before this optimization to 80.2% at a selected checkpoint, with three extra lives. The perception checks show that accurate static screenshot reading can coexist with errors during interaction. A separate policy comparison also shows that further optimization can accompany worse action choices. Together, these findings motivate concrete checks for post-training: evaluate perception in use, distinguish training success from full-game validation, and examine decisions and later performance rather than relying on one selected score. The study provides empirical lessons for improving future agent training and evaluation.
Deep learning methods for pathology are constrained by the limited availability and high cost of large-scale pixel-level annotations. Synthetic data can alleviate this limitation, but existing pipelines often lack reliable quality control and remain weakly connected to downstream analysis. To address these issues, we propose a three-stage framework that integrates automated data screening, generative representation learning, and efficient adaptation to downstream tasks. First, we develop a pathology agent system to assess image quality and filter samples, resulting in TCGA-200K, a dataset containing 217,030 pathology patches selected from the public TCGA repository. Second, we train a generative model on the screened data. In addition to image synthesis, its backbone learns structural and semantic representations that can be transferred to downstream tasks. Third, we adapt the pretrained generative backbone to semantic segmentation using a lightweight task-specific head. We further introduce PathSeg-380K, a semantic segmentation dataset with expert annotations for approximately 380,000 cells across 12 cell types. Experiments on multiple pathology datasets show that the proposed framework achieves state-of-the-art segmentation performance. Our code is publicly available at https://github.com/script-Yang/Agent-GenPath.