arXiv subjects

Weinan Zhang

Publications and source records attributed to Weinan Zhang.

At least 55 records · Page 3Linked to original sources

arXiv2026

SkillGate: Training In-Policy Skill Selection in Long-Horizon Agents

Agent frameworks increasingly package procedural knowledge as skills: instruction files an agent reads on demand, while public libraries now hold thousands of them. Which skill to read has thus become a decision the policy itself makes in the middle of an episode, yet no existing signal trains it. We show that the default remedy, outcome-rewarded RL over the candidate slate, cannot teach it, for a structural reason we identify and name selector credit starvation: under a broadcast, sequence-level advantage, the few tokens that name the chosen skill carry a vanishing share of the loss, and the credit they inherit is increasingly wrong-signed as trajectories lengthen. A correct choice is punished whenever the execution after it fails, even though the choice itself is among the most valuable decisions in the trajectory. Auditing a completed run's own training artifacts confirms all three properties, each worsening monotonically with horizon. SkillGate removes the failure by construction: it partitions the token support into two disjoint credit channels, outcome credit reaching only execution tokens, and a separate action-local advantage reaching exactly the skill-naming tokens, positive only when a trajectory's single read is the correct one. On five agentic benchmarks under a 16-candidate slate, SkillGate lifts a 9B policy from 40.8% to 53.2% trial success, well ahead of the identical budget spent on outcome reward alone, while cutting exposure to misleading candidates by two thirds and reading fewer skills.

cs.AI↗

arXiv2026

SAGE: Self-Evolving Storyboard Skills via Attribution-Guided Rule Evolution

Storyboards turn screenplays into visual shot plans for automated short drama production. Professional storyboarding relies on tacit directorial expertise and remains an industrial bottleneck. Large language models can automate this step, but methods for supplying directing knowledge face three challenges: (1) Knowledge acquisition: the craft remains implicit in exemplars or must be written manually. (2) Knowledge refinement: authored knowledge is not evaluated against execution outcomes, and opaque generation prevents feedback attribution to the knowledge behind each decision. (3) Knowledge injection: injecting all knowledge exceeds usable context, while manual selection for every narrative group does not scale. We present SAGE (Skill with Attribution-Guided Evolution), a deployed framework that learns, attributes, evolves, and routes directing knowledge from expert demonstrations. SAGE derives rules that are independent of episode content by contrasting each training screenplay with its expert storyboard. During generation, the model records each narrative group's adopted rules. Combining these records with localized feedback enables targeted updates to individual rules. Evolved rules form scenario packages with a routing index, so each group retrieves only a bounded set appropriate to its situation without expert intervention. On 18 test episodes across three genres, SAGE scored 77.8 on a rubric validated by experts, versus 77.1 for professional directors. Deployed for 14 days on Virtual Film Studio, SAGE produced 1,344 narrative group outputs; 87.2 percent were accepted without substantive edits, and the production team recorded over 83 percent less authoring time per episode. We release PROSE, the first public dataset pairing screenplays with storyboards by professional directors across 68 episodes: https://github.com/creDreams/PROSE.

cs.AI↗

arXiv2026

Skill2Query: Exploiting Skill Structure to Generate Pseudo-Queries for Agent Skill Retrieval

Pseudo-query generation can alleviate the supervision bottleneck for agent skill retrieval, but existing document-level approaches typically leave the rich internal relations among capabilities, parameters, and usage examples implicit. As a result, generated queries may be topically relevant to a skill while lacking capability grounding and parameter consistency, raising the question of whether explicitly exploiting a skill document's internal structure can produce more effective retrieval signals. We therefore propose Skill2Query, a framework that first parses a skill document into a Skill Knowledge Graph and then generates pseudo-queries through a three-stage process including style mimicking, query template generation, and parameter filling. The generated queries can be used for offline index augmentation, online query expansion, and retriever training. Four benchmarks (TheoremQA, LogicBench, ToolQA, and CHAMP) are used to evaluate Skill2Query with large-scale skill candidate pools across multiple downstream applications, including skill retrieval, retriever training, and end-to-end agent execution. Using nearly 30K skills across diverse domains, we generate 700K category-diverse pseudo-queries. Skill2Query consistently improves sparse, dense, and skill-routing retrieval, with an average Recall@1 gain of 6.70 percentage points across retrieval settings. Skill2Query-generated training data also achieves the best Recall@1 and nDCG@1 among the evaluated generation baselines. Further evaluations with multiple LLM backends demonstrate that improved skill retrieval translates into higher agent task success rates. Code and resources are available at https://github.com/MatZaharia/Skill2Query.

cs.CL↗

arXiv2026

RoboStriker: Latent-Space Strategic Games for Autonomous Humanoid Boxing

Achieving human-level competitive intelligence and physical agility in humanoid robots remains a profound challenge, particularly in contact-rich and highly dynamic tasks such as boxing. While Multi-Agent Reinforcement Learning offers a principled framework for strategic interaction, its direct application to unstructured raw motor spaces inevitably leads to joint-level physical collapse, preventing the emergence of any viable combat tactics. To resolve this fundamental conflict between strategic exploration and physical feasibility, we formulate the humanoid combat task as a novel two-player latent-space zero-sum Markov game. Under standard regularity and approximate best-response assumptions, we show that the latent formulation induces an equivalent game over the decoder-reachable action manifold, providing an approximate-Nash interpretation of the resulting self-play dynamics. To instantiate this theoretical formulation, we propose RoboStriker, a hierarchical framework that decouples high-level reasoning from low-level execution. It first distills the tracking expertise of predefined boxing motions into a topologically bounded latent manifold. This structured latent foundation subsequently drives multi-agent co-evolution via Latent-Space Neural Fictitious Self-Play. Extensive experimental results demonstrate that gaming within this structured latent space substantially outperforms direct exploration. By constraining strategic exploration through a pretrained motion decoder, RoboStriker substantially reduces the catastrophic balance failures observed in raw action-space methods and achieves superior tactical performance in both competitive win rates and striking efficiency. Finally, we successfully deploy and validate our learned combat policies on real-world humanoid robots. Our code and video and supplementary materials are available at RoboStriker.

cs.RO↗

arXiv2026

Learning from Unreachable Rewards: Hint-Conditioned Reinforcement Learning for Generative Recommendation

Semantic-ID generative recommenders represent each item as a short sequence of discrete semantic tokens and predict the next item by autoregressively generating this token sequence. This paradigm enables a unified generation interface for item IDs, histories, and item text, but it also creates a structured optimization bottleneck during reward-based post-training: when an early semantic token enters the wrong branch of the item-token space, finite rollout groups rarely reach the ground-truth item, so group-relative optimization receives identical zero rewards and produces no useful advantage. We propose Hint-Conditioned Generative Recommendation (HCGRec), a semantic-ID generative recommendation framework that recovers learning signal for such hard training instances. HCGRec diagnoses each instance with checkpoint rollouts and supplies a minimal target-prefix hint only when the current generator cannot reach the correct item. The model then generates the unhinted suffix under the hinted semantic branch, turning zero-reward groups into informative comparisons over item-token completions. Hinting also changes token identity: hinted prefix tokens are oracle-provided item context, while unhinted suffix tokens are sampled generation actions. We therefore introduce hint-aware credit decomposition, using supervised learning to preserve item-semantic and prefix-structure alignment for hinted tokens and GRPO to optimize the sampled suffix. Experiments on sequential recommendation benchmarks show that HCGRec substantially improves over supervised fine-tuning and vanilla reward-based post-training, while reducing zero-advantage training samples from over 70% to below 20%. The code is accessible at https://github.com/WncFht/GRec.

cs.IR↗

arXiv2026

RLinf-VLA: A Unified and Efficient Framework for Reinforcement Learning of Vision-Language-Action Models

Recent studies have demonstrated the potential of reinforcement learning (RL) to improve the task performance of vision-language-action (VLA) models through interaction. However, current efforts remain fragmented, lacking a unified platform for fair comparison across architectures and algorithms, as well as an efficient system design for scalable training. Therefore, we present RLinf-VLA, a unified and efficient framework for scalable RL training of VLA models. RLinf-VLA standardizes the integration of diverse VLA architectures, RL algorithms, and heterogeneous simulators through a unified interface, enabling extensibility and reproducibility. To improve efficiency, the framework adopts a flexible resource allocation architecture for rendering, inference, and training in RL pipelines. In particular, RLinf-VLA introduces a hybrid fine-grained pipeline allocation strategy that achieves a 1.61$\times$-1.88$\times$ training speedup on ManiSkill. Using this framework, RL-trained models achieve strong performance across embodied benchmarks, including 98.11% success on 130 LIBERO tasks, 97.66% success on 25 ManiSkill tasks, and 84.63% average success across 6 RoboTwin tasks. In addition, RLinf-VLA distills a set of effective practices for RL-based VLA training. We envision RLinf-VLA as a foundational framework for efficient, unified, and reproducible research in embodied intelligence.

cs.RO↗

arXiv2026

MARA: Flow-Matching-Guided Multi-Agent Resource Allocation for Computational Resource Efficient Learning

Allocating limited computation among concurrent learning tasks is difficult when each task must reach a target loss before a deadline but its required training effort is unknown. Existing approaches combine online loss prediction with adaptive resource allocation, yet commonly treat computation as continuously divisible throughput. We instead study a practical setting in which tasks arrive over time and computation is provided by discrete nodes. This setting introduces both uncertain demand and constrained sequential decisions. We propose MARA, which predicts future loss trajectories with conditional flow matching and coordinates compute nodes through a cooperative multi-agent autoregressive policy. A potential-based progress reward supplies intermediate training feedback while preserving the undiscounted task-completion objective. Across in-distribution, reinforcement-learning, and vision workloads, flow matching reduces remaining-resource prediction error relative to weighted least squares. At the scheduler's training load, MARA completes 63.46% of tasks on average, 8.54 percentage points above strong baseline Learning with Adaptive Resource Allocation (LARA), and remains ahead under unseen heavier workloads.

cs.LG↗

arXiv2026

GAUGE: A Measurement-Grounded Benchmark for Physical Fidelity in Simulation Engines and Video World Models

Physics engines facilitate large-scale training and evaluation for embodied intelligence, while generative video world models are emerging as implicit simulators of future states and interactions. However, existing evaluations of physical fidelity are often conducted in isolation and rely heavily on perceptual similarity or human judgments, providing limited insight into which physical principles or parameters are violated. We introduce GAUGE, a real-world-grounded diagnostic benchmark for jointly evaluating how numerical simulators and generative video world models reproduce or deviate from real-world physics. It comprises 22 controlled task families covering rigid bodies, flexible cables, textiles, and volumetric deformable objects. Grounded in real-world trajectories and paired with calibrated physical metadata, uncertainty annotations, and task-specific observables, these tasks cover fundamental physical processes including collision, friction, momentum transfer, oscillation, self-contact, and deformation across diverse materials and conditions. We benchmark Isaac Sim, Genesis, and Newton on 14 task families using generalized trajectory errors, and evaluate 6 image-to-video models on 5 rigid-body tasks by testing physical-law consistency and the temporal stability of inferred parameters. Our results reveal no uniformly faithful physics engine, with the largest discrepancies arising in impulsive contact, rapid textile motion, and volumetric deformation. We further find that video world models can produce trajectories with the expected equation form while recovering incorrect accelerations, momentum transfer, and oscillation timing. GAUGE lays the groundwork for developing more physically faithful simulators and world models for embodied intelligence.

cs.AI↗

arXiv2026

Belief-Contraction-Driven Active Inverse Source Localization and Characterization

Active inverse source localization and characterization (ISLC) in dynamic fields requires sequential decision making under partial observability, where a mobile sensor must infer latent source parameters from sparse, noisy readings. We introduce a belief-contraction-driven approach that unifies inference, stopping, and control. An attention-augmented particle filter stabilizes Bayesian belief updates through ESS-based resampling, feature-aware sparse attention smoothing, and Metropolis-Hastings rejuvenation that preserves the filtering posterior. Belief contraction (posterior dispersion) defines both a termination rule and a goal-aligned intrinsic reward, enabling reinforcement learning without distance-to-source shaping. Across seven field modalities, spatial out-of-distribution tests, and nonstationary source shifts, our agent (ATT-PFRL) achieves higher completion, faster convergence, and more accurate localization than planning and RL+Bayes baselines under similar computation. Fixed-trajectory studies also show improved ESS and lower RMSE, isolating the benefit of the inference layer.

cs.LG↗

arXiv2026

Harness-R1: Learning to Edit Executable Runtime Harnesses from Agent Failure Trajectories

Agents built around large language models continually accumulate interaction trajectories during deployment, yet their behavior typically remains fixed. Beyond updating model weights, these trajectories can improve the agent harness that constructs context, mediates tools, validates actions, and recovers execution. We introduce Harness-R1, the first method, to our knowledge, that makes failure-conditioned, lifecycle-wide editing of an existing executable runtime a learned capability. It post-trains a dedicated harness engineer with online reinforcement learning so that its edits are optimized for the realized task success they produce, rather than proposed by a fixed editor. A separate 9B engineer converts batches of target-agent failures into validated executable patches; fresh same-batch reruns of the frozen target provide outcome rewards, so training updates only the engineer. Cold-start supervised fine-tuning initializes this editing policy, which is then trained online with group-relative policy optimization. Across WebShop, ALFWorld, and DBBench, Harness-R1 raises vanilla Qwen3.5-9B success from 44.3% to 53.6% (+9.3 percentage points). After direct target-agent fine-tuning, a target-specific engineer raises the average further from 59.2% to 64.2% (+5.0 points); because these gains hold both before and after fine-tuning the target, Harness-R1 points toward co-evolving the harness engineer and the target agent.

cs.AI↗

arXiv2026

Quick on the Uptake: Eliciting Implicit Intents from Human Demonstrations for Personalized Mobile-Use Agents

As multimodal large language models advance rapidly, the automation of mobile tasks has become increasingly feasible through the use of mobile-use agents that mimic human interactions from graphical user interfaces. To further enhance mobile-use agents, previous studies employ demonstration learning to improve mobile-use agents from human demonstrations. However, these methods focus solely on the explicit intention flows of humans (e.g., step sequences) while neglecting implicit intention flows (e.g., personal preferences), which makes it difficult to construct personalized mobile-use agents. In this work, to evaluate the Intention Alignment Rate between mobile-use agents and humans, we first collect MobileIAR, a dataset covering 40 English- and Chinese-speaking users and containing human-intent-aligned actions and ground-truth actions. This enables a comprehensive assessment of the agents' understanding of human intent. Then we propose IFRAgent, a framework built upon Intention Flow Recognition from human demonstrations. IFRAgent analyzes explicit intention flows from human demonstrations to construct a query-level vector library of standard operating procedures (SOP), and analyzes implicit intention flows to build a user-level habit repository. IFRAgent then leverages a SOP extractor combined with retrieval-augmented generation and a query rewriter trained through knowledge distillation to generate personalized queries and SOPs from a raw ambiguous query, enhancing the alignment between mobile-use agents and human intent. Experimental results demonstrate that IFRAgent consistently outperforms baselines on every model, achieving an average absolute improvement of 9.88\% (53.33\% relative) in human intention alignment rate and an average absolute improvement of 8.26\% (32.40\% relative) in step completion rate.

cs.CL↗

arXiv2026

VAD: Attributing Visual Evidence for Target Reconstruction in Multimodal On-Policy Distillation

Multimodal on-policy distillation (OPD) transfers fine-grained visual knowledge by supervising student-generated trajectories with a privileged-view teacher. Yet its next-token corrections are source-mixed, combining visual signals with linguistic priors and teacher-specific effects. The key challenge is to estimate which corrections are supported by visual evidence, not merely where or how strongly to distill. We introduce Visual Attribution Distillation (VAD), a counterfactual target-reconstruction algorithm that estimates the visually attributable part of a teacher correction. At each student-generated prefix, VAD evaluates the same fixed teacher with the relevant evidence present and removed. The corresponding change in centered log-probabilities defines ut, a signed proxy for the visual evidence direction that estimates how revealing the evidence supports or refutes candidate tokens. VAD projects the original correction onto this proxy to obtain an intervention-aligned component and a proxy-unexplained residual, then reconstructs a student-anchored target from the former. During training, this reconstructed target supplies the primary supervision signal, while the privileged teacher contributes a weak regularizer. Across six fine-grained visual benchmarks at 4B and 9B scales, VAD outperforms direct privileged-view distillation and visual-advantage weighting. Token- level and controlled-target analyses show that the proxy-aligned component is enriched in task-relevant visual corrections and yields stronger target shifts, especially when evidence refutes a mistaken answer. These results support counterfactual target reconstruction as an effective alternative to source-mixed supervision.

cs.CV↗

arXiv2026

Skills on the Fly: Test-Time Adaptive Skill Synthesis for LLM Agents

Additional test-time compute can give LLM agents access to more past experience, yet expanding the context or adding rollouts does not necessarily yield greater agent capability. We call this challenge test-time compute-to-capability conversion and propose SkillTTA, which retrieves task-relevant training trajectories and synthesizes a temporary skill conditioned on the visible target context for a solver with fixed parameters. To pursue a higher performance ceiling, SkillTTA further uses meta prompt optimization (MPO) to adapt the policy that writes these skills. MPO evaluates candidate prompts on paired tasks and emphasizes informative transitions. It also confines updates to benchmark-specific atomic slots, reducing the variance caused by observing each edit only indirectly through skill synthesis and solver rollout. Across ALFWorld, SpreadsheetBench, BigCodeBench, and WebShop, SkillTTA outperforms state-of-the-art reuse and optimization baselines. It attains a higher performance ceiling at lower compute cost than baseline reuse and sampling strategies.

cs.CL↗

arXiv2026

Autoregressive Diffusion World Models for Off-Policy Evaluation of LLM Agents

Evaluating large language model (LLM) agents in multi-turn interactive environments is expensive and risky, as it requires online environment interaction. We propose ADWM (Autoregressive Diffusion World Model), an evaluation framework that estimates the performance of a new LLM agent policy purely from pre-collected trajectories. The core idea is to learn a latent diffusion world model that simulates how the environment responds to the evaluation policy, without ever executing it in the real environment. Existing diffusion-based OPE methods guide full trajectories in a single pass by jointly diffusing states and actions, an assumption that breaks down for LLM agents whose actions are discrete text that must be sampled from the policy after observing the environment. Unlike autoregressive world models that suffer from compounding errors, ADWM models each transition as an independent denoising process, enabling reliable step-by-step rollouts where the world model and agent alternate in causal order. Crucially, the LLM agent under evaluation directly guides the diffusion generation at each step via a policy-conditioned score function, ensuring that simulated trajectories accurately reflect its decision-making patterns. Empirically, ADWM achieves accurate value estimates and evaluation reliability across diverse multi-turn agent tasks, demonstrating its promise as a practical framework for offline LLM agent evaluation.

cs.LG↗

arXiv2026

IssueExec: A Test-Driven Approach for Localizing Software Engineering Issues

Issue localization, which identifies code locations requiring modification from issue descriptions, is a critical step in automated software maintenance. Existing approaches predominantly attempt to directly align issue descriptions with code elements, yet often struggle due to the inherent abstraction gap between the issue description and code implementation. Seeking alternative signals, our theoretical analysis suggests that test suites can serve as executable proxies for requirements, reducing localization uncertainty by 7.73 bits of entropy on average. A large-scale empirical study on 18 repositories validates this premise: existing tests cover 96.98\% of ground-truth files, and the two-hop pathway yields stronger semantic connectivity than direct matching in 82.4\% of cases. Despite their potential, leveraging tests for localization faces two key challenges: the semantic gap separating issue descriptions from test identifiers, and the substantial noise in execution traces from infrastructure code. To address these, we propose IssueExec, which bridges the semantic gap through domain-knowledge-enhanced test representations and filters noise via hierarchical trace analysis. Experiments on SWE-bench Lite show that IssueExec achieves state-of-the-art performance, improving function-level Recall@1 by 41.57\% over the strongest baseline. When integrated into the Agentless pipeline, IssueExec resolves 17.72\% more issues, demonstrating practical downstream benefits.

cs.SE↗

arXiv2026

RAD: Retrieval High-quality Demonstrations to Enhance Decision-making

Offline reinforcement learning (RL) learns policies from fixed datasets, thereby avoiding costly or unsafe environment interactions. However, its reliance on finite static datasets inherently restricts the ability to generalize beyond the training distribution. Prior solutions based on synthetic data augmentation often fail to generalize to unseen scenarios in the (augmented) dataset. To address these challenges, we propose Retrieval High-quAlity Demonstrations (RAD) for decision-making, which innovatively introduces a retrieval mechanism into offline RL. Specifically, RAD retrieves high-return and reachable states from the offline dataset as target states, and leverages a generative model to generate sub-trajectories conditioned on these targets for planning. Since the targets are high-return states, once the agent reaches such a target, it can continue to obtain high returns by following the associated high-return actions, thereby improving policy generalization. Extensive experiments confirm that RAD achieves competitive or superior performance compared to baselines across diverse benchmarks, validating its effectiveness. Our code is available at https://github.com/LeahGL/RAD.

cs.AI↗

arXiv2026

Scaling Behavior Foundation Model for Humanoid Robots

Humanoid control requires natural whole-body coordination, precise real-time responses to control signals, and robust generalization across diverse environmental contexts, making it a cornerstone for generalist embodied agents. Behavior Foundation Models (BFMs) have recently emerged as a promising solution to address these challenges by leveraging large-scale behavioral data to achieve superior expressiveness, versatility and generalization. However, despite growing interest in scaling BFMs to further improve their capabilities, it remains unclear how key factors, including the learning paradigm, behavioral data and model architecture should be coordinated to enable effective scaling. In this work, we revisit the scaling recipe for BFMs and demonstrate that substantial performance gains can be achieved through the coordination of three core components: 1) the learning paradigm of motion tracking that reformulates diverse humanoid control problems as the reproduction of integrated whole-body behaviors in the global frame; 2) the strategic synergy between on-policy rollout quantity and reference motion diversity; and 3) the expressive and scalable model architecture termed Humanoid Transformer that facilitates the natural emergence of structured behavioral representations. Through extensive experiments in both simulation and real-world deployment, we demonstrate that our approach yields significant improvements in control fidelity and task generalization, reducing Mean Per-Keypoint Position Error (MPKPE) on the test set by over 10% in local mode and 82% in global mode compared with existing humanoid controllers. These results establish BFM as a principled and effective foundation for scalable and general-purpose humanoid control.

cs.RO↗

arXiv2026

HELP: Human-Efficient Large-Scale Robot Post-Training with Rollout Segmentation

When adapting Vision Language Action (VLA) models to downstream tasks, multiple rounds of post-training are often required to progressively address policy weaknesses. In this report, we focus on maximizing human efficiency during this iterative process, measured by policy improvement and task throughput per unit of human labor and time. We propose HELP, a Human-Efficient Large-scale robot Post-training pipeline in which two specialized operators supervise twelve robots concurrently. A trained Teleoperator provides high-value remote interventions and recovery demonstrations, while a Floor Operator monitors the robot fleet, triggers takeovers, and performs physical resets. This role specialization improves human efficiency by reducing task switching, lowering operator training costs, and expanding robot interaction coverage. Beyond increasing rollout volume, concurrent supervision also broadens the range of policy behaviors observed by the human team, making recurring failure modes easier to identify and enabling more targeted takeovers, resets, and recovery demonstrations. To efficiently utilize the large and mixed-quality rollout data, HELP incorporates \vlac, an automatic rollout segmentation critic specifically designed for this setting. It separates autonomous trajectories into progress-making, idle, failure-inducing, and recovery segments. Useful rollout segments are retained and combined with Human-in-the-Loop data for the next post-training round. Across four real-world manipulation tasks, HELP achieves 80\%--95\% success rates and improves task throughput by 1.7$\times$--4.2$\times$ over the base model. Under matched HITL recovery budgets, VLAC-CUT further amplifies throughput gains by 1.20$\times$--3.43$\times$ and success-rate gains by 1.50$\times$--3.00$\times$ over HITL-only updates.

cs.RO↗