arXiv subjects

Xiyang Wu

Publications and source records attributed to Xiyang Wu.

At least 19 recordsLinked to original sources

arXiv2026

Failure Recovery via Belief Updates and World Modeling for Long-Horizon Agents

Long-horizon agents often struggle to recover from failures whose causes lie several steps earlier, beyond what local reflection or backtracking can resolve. We introduce BELIEFWEAVER, a framework that combines persistent evidence across attempts with prospective reasoning over repair directions. It maintains a relational belief state over frontiers, attempts, observations, and checkpoints. Execution evidence updates related frontier priorities across attempts and rollback, while diagnostic evidence gates execution. An implicit world model (IWM) evaluates repair hypotheses, predicts progress, and identifies relations among frontiers to guide exploration. We further distill these recurring IWM reasoning components into a 9B student through supervised fine-tuning and on-policy distillation (OPD). Across three terminal-based benchmarks, BELIEFWEAVER improves aggregate final score by 5.1 percentage points over native Terminus-2 continuation, with positive trends in exploration efficiency. The distilled 9B student improves final score by 9.8 percentage points over the reported native baseline while reducing estimated teacher API cost by 54.8% relative to the matched teacher. These results show that persistent evidence and prospective reasoning can improve recovery in long-horizon agent tasks while substantially reducing reliance on teacher-model inference.

cs.LG↗

arXiv2026

Long-Horizon-Terminal-Bench: Testing the Limits of Agents on Long-Horizon Terminal Tasks with Dense Reward-Based Grading

AI agents have become capable of autonomously completing short, well-specified tasks. However, existing terminal benchmarks largely focus on simple problems that finish within minutes and are evaluated only by their final outcome. This setup overlooks intermediate progress and partial solutions, yielding sparse reward signals and an incomplete picture of agent capability. We introduce Long-Horizon-Terminal-Bench, a terminal benchmark of 46 long-horizon tasks spanning nine categories, including experiment reproduction, software engineering, multimodal analysis, interactive games, and scientific computing. Each task follows a Terminal-Bench-style setup with a reference solution or simulation engine, but is further decomposed into fine-grained graded subtasks. This design enables dense intermediate rewards and partial credit, allowing evaluation to capture not only whether an agent reaches the final goal, but also how far it progresses on open-ended workflows. Tasks in Long-Horizon-Terminal-Bench typically require hundreds of episodes and minutes to hours of execution, stressing long-horizon planning, long-context management, and iterative debugging rather than one-shot problem solving. We evaluate 15 frontier models and find that agents consume on average 9.9M tokens per task, with roughly 231 episodes and 85.3 minutes of execution time per run, making Long-Horizon-Terminal-Bench more demanding than prior terminal-based benchmarks. Even the strongest tested model achieves 15.2% pass@1 at a partial-reward threshold of 0.95 and 10.9% at a perfect-reward threshold of 1.0, while the mean pass rate across models is 4.3% and 1.7% under the two thresholds, respectively. These results reveal headroom for improvement. We further analyze failure modes and error patterns, and release Long-Horizon-Terminal-Bench to support future progress on long-horizon terminal agents.

cs.AI↗

arXiv2026

Long-Horizon-Terminal-Bench 2.0: Automating Diverse Task Construction and Measuring Sustained Agent Progress

Long-horizon terminal benchmarks measure whether an agent can sustain progress through extended workflows with process rewards, not only whether its final state passes a binary check. We present Long-Horizon-Terminal-Bench 2.0 (LHTB 2.0), an extension of Long-Horizon-Terminal-Bench that preserves its 46 original tasks while correcting task and evaluation-infrastructure defects and extending 31 tasks. LHTB 2.0 contains 77 tasks across 13 domains and eight task types. The extended tasks are produced through a systematic reconstruction-task pipeline. We start from a working scientific or engineering artifact, then use a parameterized reference implementation to generate multiple scenarios, reveal six as behavioral evidence, hold out three for evaluation, and use dense progress ladders to reward partial progress while preventing memorization of public outputs. We also extend the evaluation horizon from 90 minutes to three hours and record verifier checkpoints every 30 minutes, enabling direct measurement of sustained progress within a run. Across 1,145 scored model-task runs from 15 frontier models, GPT-6-astra leads with dense mean reward 0.69 and 37 tasks at R at least 0.95, followed by GPT-5.6-sol with dense mean reward 0.67 and 32 tasks at the same threshold. On the 30 reconstruction tasks, binary verifier scores average 0.221, whereas dense ladder scores average 0.653, showing that outcome-only evaluation discards substantial intermediate progress. LHTB 2.0 provides a broader, hardened, and process-aware testbed for studying reliable agent execution over hours.

cs.AI↗

arXiv2026

ASSEMBLE: Atomic Skills for Evidence-Grounded Video Reasoning

Complex video reasoning often depends on evidence scattered across distant moments, entities, and events, yet a correct answer alone does not reveal whether a model relied on the right parts of the video. We introduce ASSEMBLE, a framework that makes supporting evidence explicit throughout long-video reasoning. ASSEMBLE organizes local observations and cross-clip narratives into timestamped evidence catalogs traceable to the source video. A grounding-aware reader then composes question-specific atomic skills whose structured outputs contain explicit evidence references and support assessments. We use correctness-gated citation alignment as a direct grounding signal: after teacher-supervised fine-tuning, Group Relative Policy Optimization (GRPO) jointly optimizes answer correctness and citation alignment. This produces inspectable intermediate traces while keeping final predictions linked to explicit supporting evidence. Using a 9B reader supervised by a 235B teacher and shared precomputed evidence catalogs, ASSEMBLE achieves 59.2% macro-averaged answer accuracy across three long-video reasoning benchmarks, compared with 58.3% for Gemini-2.5-Pro, while improving macro-averaged overlap-based Grounded accuracy by 6.7%, with gains on all three benchmarks. Ablations further show that, with the same post-trained reader and inference budget, structured skill inference improves Grounded accuracy over free-form reasoning. Together, these results show that explicit evidence grounding can be integrated directly into long-video reasoning without sacrificing answer accuracy.

cs.MM↗

arXiv2026

Causeway: Restoring Task Accessibility for Instruction Switching in VLA Policies

Vision-language-action (VLA) policies can execute many tasks from standard initial states, yet a new instruction may fail after another task has altered the robot's physical state. We study instruction switching, where a new task is issued during or after the execution of a different one. We observe that a target task that is reliably completed from its standard initial states can become inaccessible from states produced by a preceding task. We call such states task islands. We propose Causeway, a training-free inference-time intervention. Given the current state and a re-entry pose for the target task, Causeway back-propagates through the frozen decoding computation and applies a state-directed write within the action-stream representation. The VLA decodes the return motion itself, without parameter updates, a new action head, or external action generation. Across 71 cross-object pairs, three switch timings, and three VLA architectures on LIBERO-Goal, Causeway raises bare-switch success from 3-26% to 47-65% and increases the rate of reaching the handoff neighborhood by 42-72 percentage points across models. Additional experiments on LIBERO-Object and a real xArm platform show that the recovery extends beyond the main LIBERO-Goal setting, both in simulation and on a robot.

cs.RO↗

arXiv2026

Graph-of-Skills: Dependency-Aware Structural Retrieval for Massive Agent Skills

As LLM agents act across personal applications, web browsers, and other interfaces, their reusable skill libraries can scale to thousands of skills. This scale introduces two challenges. First, loading the full library saturates the context window, driving up token costs, hallucination, and latency. Second, semantic retrieval surfaces topically relevant skills but can miss upstream and downstream prerequisite skills, creating a prerequisite gap that leaves the retrieved bundle insufficient for execution. We present Graph-of-Skills (GoS), an inference-time structural retrieval layer for large skill libraries. GoS constructs an executable skill graph offline from skill packages, then retrieves a bounded, dependency-aware bundle through hybrid semantic-lexical seeding, reverse-aware Personalized PageRank, and context-budgeted hydration. Across SkillsBench and ALFWorld, with three model families (Claude Sonnet 4.5, MiniMax M2.7, and GPT-5.2 Codex), GoS attains the highest average reward in all six model-benchmark blocks, at a fraction of the token cost of loading the full library. On SkillsBench with GPT-5.2 Codex it raises average reward by 7.0 absolute points over full skill loading, a 25.6% relative gain, while cutting total tokens by 56.7%. Ablations isolate the mechanism: replacing reverse traversal with forward propagation costs 9.1 reward points, a larger loss than removing the graph altogether. The gain thus comes from traversing dependencies backwards, not from graph diffusion as such. A budget-matched retrieval study holding seeding, reranking, hydration, and context budget fixed reproduces the same ordering, with dependency-pair co-recovery falling from 0.654 to 0.362. Code is available at https://github.com/davidliuk/graph-of-skills

cs.AI↗

arXiv2026

Recursive Agentic Reasoning

Test-time reasoning methods such as iterative refinement, decomposition, and repeated sampling are often evaluated in isolation, making their gains difficult to compare across models, benchmarks, and evaluation pipelines. We introduce a unified view of these methods as recursion operators over an agent's reasoning trace: GROW, which deepens a single reasoning path; PRUNE, which decomposes and recomposes the problem; and BRANCH, which samples alternative reasoning paths and selects among them. We evaluate all three operators against a single-pass chain-of-thought baseline under a shared harness with identical prompts, token budgets, and grading code. Across five benchmarks and three frontier models, comprising 14 model-benchmark settings, 49,327 graded items, and 151,876 model calls, BRANCH improves accuracy in all 14 settings by an average of 5.98 percentage points and is the best-performing operator in 12. In contrast, GROW yields a mean gain of 2.18 points and degrades performance in two settings, while PRUNE improves accuracy by 0.94 points on average. Analysis shows that BRANCH's advantage arises not only from exploring multiple reasoning paths, but also from recovering from truncation: its gains strongly correlate with the baseline rate of empty, budget-exhausted outputs (r = 0.72). These results weaken the hypothesis that different problems require routing among test-time reasoning operators; at this level of abstraction, repeated branching is consistently dominant. Finally, we show that unpaired evaluation and treating scoring-pipeline failures as model errors can materially change, and even reverse, comparative conclusions, motivating paired scoring as a standard protocol for test-time-compute evaluation.

cs.AI↗

arXiv2026

Noise Floor Audit for Agent Benchmarks

We audit measurement variability for 3 native tool-calling endpoints across 2 providers on the official BFCL multiple and parallel categories, using matched AST grading. At temperature 0, reruns are nearly deterministic across Groq endpoints and a thinking-enabled Gemini setting: ever-flip fractions are 0.7%, 2.0%, and 2.7%, with mean run correlations of 0.997, 0.966, and 0.961. Semantics-preserving prompt perturbations create the larger floor on all endpoints, with median perturbation paired SDs 11x to 58x larger than rerun paired SDs. The failure character also shifts: malformed-output failures account for 30%, 7%, and <1% of task failures, so marginal accuracy hides not only stability but also failure mode.

cs.CL↗

arXiv2026

A Survey of State of the Art Large Vision Language Models: Alignment, Benchmark, Evaluations and Challenges

Large vision-language models (VLMs) have evolved rapidly from contrastive image-text encoders and adapter-based assistants into natively multimodal foundation models that support long-context reasoning, agentic workflows, and, increasingly, unified perception-generation-action loops. Since 2025, the frontier has consolidated around a small number of large pretraining families, including GPT, Gemini, Claude, Grok, Qwen, Gemma, DeepSeek, Kimi, and MiniMax, while evaluation has shifted from short-form visual question answering toward spatial, temporal, embodied, and calibration-sensitive benchmarks. In this survey, we provide an updated overview of VLMs through 2026 with an emphasis on three developments. First, we trace the architectural evolution from contrastive or bridged two-tower models to LLM-backbone adapter models, native multimodal-input models, omni-modal unified input/output models, and emerging world-action models. Second, we summarize how benchmarks expanded from classical OCR, VQA, and chart understanding toward long-video reasoning, embodied evaluation, reward-model judging, and domain-specific decision making. Third, we review the growing role of post-training, including supervised fine-tuning, preference optimization, and reinforcement learning methods such as GRPO-style multimodal alignment. Rather than exhaustively listing every release, we focus on representative frontier families, influential benchmarks, and the major open challenges that remain in hallucination, safety, efficiency, data quality, and multimodal alignment.

cs.CV↗

arXiv2026

COMFYCLAW: Self-Evolving Skill Harnesses for Image Generation Workflows

Agents are increasingly used to construct workflows and assist humans in completing recurring tasks more efficiently. As these workflows become repeated and domain-specific, agent memory and reusable skills become increasingly important: agents should be able to recall workflow patterns, execution constraints, and user preferences from previous runs. We study this problem in workflow-based image generation and introduce COMFYCLAW, an agentic skill evolution harness for controlling ComfyUI workflows. COMFYCLAW formulates workflow construction as typed graph editing, exposes tools organized by construction stage, automatically reverts invalid edits, and uses a region-level vision-language model (VLM) verifier to translate visual failures into actionable repair suggestions. The framework further evolves a progressively disclosed skill library, where trajectories, execution errors, and verifier feedback from previous runs are distilled into reusable Agent Skills. Across four benchmark splits, three agent models, and two image backbones, COMFYCLAW achieves the best average image-generation evaluation score across all six agent configurations, outperforming a verifier-only baseline without skill evolution. Human annotations further show that annotators prefer COMFYCLAW over variants without skill evolution. Our results suggest that skill evolution is an effective mechanism for improving agent reliability and performance in recurring visual workflow construction.

cs.AI↗

arXiv2026

Act on What You See: Unlocking Safe Social Navigation in Vision-Language-Action Models

Safe social navigation requires robots to distinguish people from ordinary obstacles and to react before danger becomes imminent. We show that pretrained Vision-Language-Action (VLA) models already encode pedestrian-object distinctions and future collision signals in their internal representations, but behavior cloning fails to translate these signals into socially appropriate actions. To address this mismatch, we propose SALSA, a two-stage annotation-free post-training framework: (1) social behavioral alignment bridges intermediate-layer social features to the action head and trains on counterfactual human-object scene pairs to break visual saliency shortcuts; (2) temporal safety alignment provides automatically generated future-risk supervision to enable anticipatory collision avoidance. On SCAND and real-world deployment, SALSA reduces near-collisions by 86.4% and improves social counterfactual accuracy from 53% to 93%, demonstrating that safer social navigation can be achieved by teaching VLA policies to act on representations they already possess. These results show that pretrained VLA policies can be adapted for safer social navigation by better aligning their latent representations with action generation.

cs.RO↗

arXiv2026

Co-Evolving LLM Decision and Skill Bank Agents for Long-Horizon Tasks

Long horizon interactive environments are a testbed for evaluating agents skill usage abilities. These environments demand multi step reasoning, the chaining of multiple skills over many timesteps, and robust decision making under delayed rewards and partial observability. Games are a good testbed for evaluating agent skill usage in environments. Large Language Models (LLMs) offer a promising alternative as game playing agents, but they often struggle with consistent long horizon decision making because they lack a mechanism to discover, retain, and reuse structured skills across episodes. We present COSPLAY, a co evolution framework in which an LLM decision agent retrieves skills from a learnable skill bank to guide action taking, while an agent managed skill pipeline discovers reusable skills from the agents unlabeled rollouts to form a skill bank. Our framework improves both the decision agent to learn better skill retrieval and action generation, while the skill bank agent continually extracts, refines, and updates skills together with their contracts. Experiments across six game environments show that COSPLAY with an 8B base model achieves over 25.1 percent average reward improvement against four frontier LLM baselines on single player game benchmarks while remaining competitive on multi player social reasoning games.

cs.AI↗

arXiv2026

MASS: Motion-Aware Spatial-Temporal Grounding for Physics Reasoning and Comprehension in Vision-Language Models

Vision Language Models (VLMs) perform well on standard video tasks but struggle with physics-related reasoning involving motion dynamics and spatial interactions. We present a novel approach to address this gap by translating physical-world context cues into interpretable representations aligned with VLM perception, comprehension, and reasoning. We introduce MASS, a model-agnostic approach that injects spatiotemporal signals into the VLM language space via depth-based 3D encoding and visual grounding, coupled with a motion tracker for object dynamics. We also contribute a comprehensive benchmark, MASS-Bench, consisting of 4,350 real-world and AIGC videos and 8,361 free-form video question-answering pairs focused on physics-related comprehension tasks, with detailed annotations including visual detections and grounding over sub-segments, as well as full-sequence 3D motion tracking of entities. To strengthen cross-modal alignment and reasoning, we apply reinforcement fine-tuning to MASS. Experiments and ablations show that our refined VLMs outperform comparable baselines, larger models, and prior state-of-the-art models, achieving performance comparable to closed-source state-of-the-art VLMs, with only a 2\% gap to Gemini-2.5-Flash on physics reasoning and comprehension.

cs.CV↗

arXiv2026

SABER: A Stealthy Agentic Black-Box Attack Framework for Vision-Language-Action Models

Vision-language-action (VLA) models enable robots to follow natural-language instructions grounded in visual observations, but the instruction channel also introduces a critical vulnerability: small textual perturbations can alter downstream robot behavior. Systematic robustness evaluation therefore requires a black-box attacker that can generate minimal yet effective instruction edits across diverse VLA models. To this end, we present SABER, an agent-centric approach for automatically generating instruction-based adversarial attacks on VLA models under bounded edit budgets. SABER uses a GRPO-trained ReAct attacker to generate small, plausible adversarial instruction edits using character-, token-, and prompt-level tools under a bounded edit budget that induces targeted behavioral degradation, including task failure, unnecessarily long execution, and increased constraint violations. On the LIBERO benchmark across six state-of-the-art VLA models, SABER reduces task success by 20.6%, increases action-sequence length by 55%, and raises constraint violations by 33%, while requiring 21.1% fewer tool calls and 54.7% fewer character edits than strong GPT-based baselines. These results show that small, plausible instruction edits are sufficient to substantially degrade robot execution, and that an agentic black-box pipeline offers a practical, scalable, and adaptive approach for red-teaming robotic foundation models. The codebase is publicly available at https://github.com/wuxiyang1996/SABER.

cs.RO↗

arXiv2026

First Frame Is the Place to Go for Video Content Customization

What role does the first frame play in video generation models? Traditionally, it's viewed as the spatial-temporal starting point of a video, merely a seed for subsequent animation. In this work, we reveal a fundamentally different perspective: video models implicitly treat the first frame as a conceptual memory buffer that stores visual entities for later reuse during generation. Leveraging this insight, we show that it's possible to achieve robust and generalized video content customization in diverse scenarios, using only 20-50 training examples without architectural changes or large-scale finetuning. This unveils a powerful, overlooked capability of video generation models for reference-based video customization.

cs.CV↗

arXiv2026

MM-Zero: Self-Evolving Multi-Model Vision Language Models From Zero Data

Self-evolving has emerged as a key paradigm for improving foundational models such as Large Language Models (LLMs) and Vision Language Models (VLMs) with minimal human intervention. While recent approaches have demonstrated that LLM agents can self-evolve from scratch with little to no data, VLMs introduce an additional visual modality that typically requires at least some seed data, such as images, to bootstrap the self-evolution process. In this work, we present Multi-model Multimodal Zero (MM-Zero), the first RL-based framework to achieve zero-data self-evolution for VLM reasoning. Moving beyond prior dual-role (Proposer and Solver) setups, MM-Zero introduces a multi-role self-evolving training framework comprising three specialized roles: a Proposer that generates abstract visual concepts and formulates questions; a Coder that translates these concepts into executable code (e.g., Python, SVG) to render visual images; and a Solver that performs multimodal reasoning over the generated visual content. All three roles are initialized from the same base model and trained using Group Relative Policy Optimization (GRPO), with carefully designed reward mechanisms that integrate execution feedback, visual verification, and difficulty balancing. Our experiments show that MM-Zero improves VLM reasoning performance across a wide range of multimodal benchmarks. MM-Zero establishes a scalable path toward self-evolving multi-model systems for multimodal models, extending the frontier of self-improvement beyond the conventional two-model paradigm.

cs.CV↗

arXiv2026

FALCON: Future-Aware Learning with Contextual Object-Centric Pretraining for UAV Action Recognition

We introduce FALCON, a unified self-supervised video pretraining approach for UAV action recognition from raw RGB aerial footage, requiring no additional preprocessing at inference. UAV videos exhibit severe spatial imbalance: large, cluttered backgrounds dominate the field of view, causing reconstruction-based pretraining to waste capacity on uninformative regions and under-learn action-relevant human/object cues. FALCON addresses this by integrating object-aware masked autoencoding with object-centric dual-horizon future reconstruction. Using detections only during pretraining, we construct objectness priors that (i) enforce balanced token visibility during masking and (ii) concentrate reconstruction supervision on action-relevant regions, preventing learning from being dominated by background appearance. To promote temporal dynamics learning, we further reconstruct short- and long-horizon future content within an object-centric supervision region, injecting anticipatory temporal supervision that is robust to noisy aerial context. Across UAV benchmarks, FALCON improves top-1 accuracy by 2.9\% on NEC-Drone and 5.8\% on UAV-Human with a ViT-B backbone, while achieving 2$\times$--5$\times$ faster inference than supervised approaches that rely on heavy test-time augmentation.

cs.CV↗

arXiv2025

VideoHallu: Evaluating and Mitigating Multi-modal Hallucinations on Synthetic Video Understanding

Vision-Language Models (VLMs) have achieved strong results in video understanding, yet a key question remains: do they truly comprehend visual content or only learn shallow correlations between vision and language? Real visual understanding, especially of physics and common sense, is essential for AI systems that interact with the physical world. Current evaluations mostly use real-world videos similar to training data, so high benchmark scores may not reflect real reasoning ability. To address this, we propose negative-control tests using videos that depict physically impossible or logically inconsistent events. We introduce VideoHallu, a synthetic dataset of physics- and commonsense-violating scenes generated with Veo2, Sora, and Kling. It includes expert-annotated question-answer pairs across four categories of violations. Tests of leading VLMs (Qwen-2.5-VL, Video-R1, VideoChat-R1) show that, despite strong results on benchmarks such as MVBench and MMVU, they often miss these violations, exposing gaps in visual reasoning. Reinforcement learning fine-tuning on VideoHallu improves recognition of such violations without reducing standard benchmark performance. Our data is available at https://github.com/zli12321/VideoHallu.git.

cs.CV↗