Daily Digest 2026-07-31

ArXiv ID: 2607.26910

Authors: Qianru Li, Xuyang Chen, Erkin Türköz, Lu Liu, Xuqin Wang, Liqiu Meng, Tao Wu, Yanfeng Zhang

Abstract:

Automatically generating cinematically expressive camera trajectories through 3D scenes from natural language descriptions is a challenging task of high practical value, with applications ranging from real-estate advertising to virtual tour creation. Existing methods either lack true 3D spatial awareness by relying on 2D image priors, or treat trajectory generation as a geometric path planning problem divorced from cinematographic semantics. We present CinemaTraj, a framework that reframes camera trajectory planning as a language-grounded spatial reasoning problem. Given a set of RGB-D images and a user prompt, CinemaTraj equips an LLM agent with a structured 3D scene graph: the agent decomposes the prompt into a sequence of atomic cinematographic movements (dolly, orbit, crane, pan, tilt, zoom, arc). Each movement is instantiated via a novel parametric trajectory representation that is both cinematographically expressive and optimizable for collision avoidance. The scene graph acts as a structured spatial prior, grounding the agent's reasoning in accurate geometric and semantic knowledge of the environment. CinemaTraj further generates synchronized voiceover and subtitles aligned with camera motion, producing narrated cinematic video outputs. We evaluate CinemaTraj on real-world ScanNet++ environments, and show that it produces prompt-faithful, collision-free trajectories with high cinematographic quality, outperforming existing approaches on prompt alignment, trajectory quality, and safety metrics.

ArXiv ID: 2607.28251

Authors: Linxuan He, Yuying Tian, Lingxiang Fan, Jiaqi Pi, Yinqiao Lu, Shang Su, Mengkai Shi, Shuo Feng

Abstract:

Despite rapid advances in policy pretraining, embodied AI systems routinely plateau during task-specific finetuning. The root cause lies in how finetuning data are collected: the default pipeline gathers data randomly, treating every sample as informative. Datasets become dominated by nominal scenarios, while rare failure cases--the most valuable for improvement--are missed. We propose a self-evolving method that breaks this plateau. Our core insight is that a state-wise criticality model, learned from the policy's own execution outcomes to predict the probability of future failure, can guide importance sampling toward failure-prone scenarios. After replacing redundant nominal scenarios with diverse failure-prone ones, importance weights are used to resample the data during training. This effectively preserves an unbiased learning objective while fundamentally increasing the information density of the training pool. Across quadrupedal locomotion, multi-task manipulation, vision-language-action benchmarks, and a real-robot task, our method reduces failure rates by 51--67% relative to trained baselines and by 8-25% relative to state-of-the-art vision-language-action models.

ArXiv ID: 2607.28226

Authors: Fazhong Liu, Zhuoyan Chen, Haozhen Tan, Yan Meng, Guoxing Chen, Haojin Zhu

Abstract:

World models give embodied AI a predictive core: they compress observations into states, simulate action-conditioned futures, and enable planning beyond reactive control. This predictive layer, however, opens a new security boundary-compromise can propagate from data, sensors, prompts, or feedback into physical action. Rather than treating world models as an isolated component, this survey traces threats across their entire lifecycle-from data construction and representation learning, through state grounding and imagination, to trajectory evaluation, execution, and long-term adaptation via memory and tools. We show that familiar attack families: poisoning, backdoors, adversarial examples, sensor spoofing, prompt injection, trajectory manipulation, and supply-chain attacks take on distinct meanings when they corrupt world states, learned dynamics, affordance estimates, or safety costs. We also highlight a duality: world models can serve as runtime safety shields, yet when compromised or over-trusted they generate predictive safety illusions. The survey offers a lifecycle taxonomy, maps existing attacks to world-model security properties, outlines evaluation protocols for safety failures, and structures defenses across provenance, robust grounding, uncertainty-aware prediction, trajectory gating, feedback auditing, and deployment assurance.

ArXiv ID: 2607.28083

Authors: Qingwen Liu, Mingqing Liu

Abstract:

Artificial intelligence (AI) is rapidly evolving from a centralized computing capability into a pervasive infrastructure that interacts directly with the physical world. While recent perspectives highlight the roles of energy, chips, infrastructure, models, and applications in enabling large-scale AI systems, these frameworks primarily assume a static, cloud-centric computing paradigm. However, emerging intelligent applications, including autonomous vehicles, drones, robots, and wearable systems, require AI to operate in highly dynamic and mobile environments. This shift introduces mobility as a fundamental constraint across the entire AI ecosystem, affecting energy supply, computation, and intelligence deployment. In this article, we introduce the concept of the Mobile AI Stack, a mobility-aware architectural framework that integrates five tightly coupled layers: mobile energy networks, energy-efficient AI chips, cloud-edge-mobile infrastructure, distributed AI models, and embodied AI applications. The proposed framework provides a systematic perspective for understanding how energy delivery, computing architectures, communication networks, and AI algorithms must co-evolve to support large-scale mobile intelligence. We further discuss key research challenges and future directions toward building scalable, reliable, and energy-efficient mobile AI systems. Mobile AI Stack offers a conceptual blueprint of the next-generation infrastructure which deeply integrates the networks of computation, energy, and communications for mobile intelligence.

ArXiv ID: 2607.27967

Authors: Dawei Wang, Di Zhao, Xinyuan Liu, Marci Chi Ma, Xiaoyang Liu, Chengming Zhou, Gary Ushaw, Richard Davison

Abstract:

Credit assignment is a fundamental challenge in cooperative multi-agent reinforcement learning, particularly in embodied AI settings characterized by limited and delayed feedback as well as dynamically changing numbers of active agents. We propose MARS-RA, a framework that reformulates credit assignment as a rank aggregation problem using contribution-based pairwise comparisons among agents generated by large multimodal models. This shift from absolute to relative estimation ensures robustness against noise and dynamic agent participation, converting comparison results into contribution scores for potential-based reward shaping. We provide theoretical justification for the convergence and robustness of the proposed framework, and show that Shapley values can be used as an interpretive reference. Experimental results on challenging tasks of different types indicate that MARS-RA can guide agents toward effective cooperation.

ArXiv ID: 2607.27443

Authors: Xu Zheng, Zhuomin Chen, Chaohao Lin, Hua Wei, Haifeng Chen, Wei Cheng, Dongsheng Luo

Abstract:

Large Language Model~(LLM)-based agents have demonstrated exceptional performance across a wide range of complex interactive tasks. However, they often struggle with long-horizon interactive tasks common in domains, such as embodied AI. The complexity and vast action spaces in these settings lead to compounding errors, where a single suboptimal action can derail an entire trajectory, causing the agent to exhaust its limited step budget on inefficient or unrecoverable paths. To overcome this without costly fine-tuning, we draw inspiration from software debugging, where execution logs are analyzed to preemptively catch errors. We propose \textit{Trajectory Graph Copilot}, a novel framework that acts as a ``copilot'' for LLM agents by diagnosing potential action errors before they are executed. At its core,\textit{Graph Debugger} models historical trajectories as a probabilistic graph and uses a Graph Neural Network to identify sequential action patterns that frequently lead to failure. Functioning as a proactive diagnostic sandbox, our method provides early warnings on potentially flawed actions, prompting the agent to self-correct. This pre-action error diagnosis prevents costly mistakes, significantly enhancing the agent's ability to complete long-horizon tasks successfully. The extensive experiments on four benchmarks with three LLM agents demonstrate a $14.69\%$ pass ratio improvement on average.

ArXiv ID: 2607.27145

Authors: Piyush Jain, Kousik Dasgupta, Rajarshi Roy, Subarna Tripathi

Abstract:

As Multimodal Large Language Models (MLLMs) are increasingly deployed in decision-critical pipelines such as robotics, embodied AI, and safety monitoring, the opacity of their spatial judgments limits operator trust and auditability. MLLMs demonstrate strong reasoning but often struggle with fine-grained spatial understanding and object hallucination. Prior work, ByDeWay, introduced Layered-Depth-Based Prompting (LDP), a training-free framework that mitigates hallucinations by structuring prompts using monocular depth estimation. However, coarse depth layering falls short in resolving object-to-object spatial relationships within the same geometric plane, such as projective ("left of", "above") and topological ("inside", "touching") relations. We propose ByDeWay-V2, which integrates explicit spatial relational context alongside depth cues, expressed as human-readable predicates that serve as auditable evidence for downstream decision support. Using an open-vocabulary object detector (YOLO-World-L), our framework computes pairwise geometric relations between detected objects and injects them as structured spatial predicates into the MLLM prompt, bridging 3D scene depth and 2D spatial semantics without any training. We evaluate ByDeWay-V2 on the Visual Spatial Reasoning (VSR) and BLINK benchmarks across multiple MLLMs, with hallucination grounding assessed via POPE. On the BLINK spatial subset, ByDeWay-V2 achieves a 46 percent relative F1 improvement over LDP for Qwen2.5-VL, and recovers BLIP-Base's spatial reasoning on VSR from near-random performance to a competitive F1 of 0.53. Our lightest configuration operates under a strict 40-token context budget on CPU, showing the framework's suitability for resource-constrained, real-time decision-support settings.

ArXiv ID: 2607.26733

Authors: Wassim Swaileh, Florent Imbert, Yann Soullard, Romain Tavenard, Eric Anquetil

Abstract:

Handwriting with digital pens is a common way to facilitate human-computer interaction through the use of Online Handwriting (OH) trajectory reconstruction. In this work, we focus on a digital pen equipped with sensors from which one wants to reconstruct the OH trajectory. Such a pen allows to write on any surface and to get the digital trace, which can help learning to write, by writing on paper, and can be useful for many other applications such as collaborative meetings, etc. In this paper, we introduce a novel processing pipeline that maps the sensor signals of the pen to the corresponding OH trajectory. Notably, in order to tackle the difference of sampling rates between the pen and the tablet (which provides ground truth information), our preprocessing pipeline relies on Dynamic Time Warping to align the signals. We introduce a dedicated neural network architecture, inspired by a Temporal Convolutional Network, to reconstruct the online trajectory from the pen sensor signals. Finally, we also present a new benchmark dataset on which our method is evaluated both qualitatively and quantitatively, showing a notable improvement over its most notable competitor.

ArXiv ID: 2607.28587

Authors: Manyi Wang, Junjielong Xu, Pinjia He

Abstract:

SWE-bench-like benchmarks are widely used for evaluating LLM's issue resolution capability. They typically follow a common construction pipeline: each PR (Pull Request) is paired with its linked issue by extracting issue references from the PR description; the issue description is used as the problem statement, and the PR patch serves as the test oracle. However, due to the inherent complexity of developing and maintaining large repositories, such PR-Issue pairings are often misaligned in practice. In this work, we systematically study SWE-bench Verified instances, finding that 13.6% exhibit misalignment across five patterns in eleven fine-grained scenarios. To enable reliable and scalable construction of those benchmarks in the future, we propose PAIChecker, a multi-agent system for checking PR-Issue misalignment in SWE-bench-like benchmarks. Specifically, PAIChecker adopts a three-phase design that combines specific pattern identification, cross-agent label synthesis, and code-level validation, thereby enabling more accurate, generalizable, and progressively verified detection. Experiments on SWE-Gym and SWE-bench Multilingual show that PAIchecker achieves the best performance across all four LLM backbones, reaching up to 92.12% and 91.67% binary accuracy, respectively.

ArXiv ID: 2607.28527

Authors: Mao-xun Huang, Jerry Wang, Yi-Cheng Lai, Zhengxin Zhang, Claire Cardie, Hen-Hsen Huang

Abstract:

Large language model-based multi-agent systems improve complex problem solving through task decomposition, agent specialization, information exchange, and intermediate validation. However, existing systems typically treat communication topology as a fixed design choice or an offline optimization target. We introduce MANTA, a framework for Multi-Agent Network Topology Adaptation that enables communication structures to self-evolve at inference time. Before execution, MANTA initializes a task-conditioned topology from prior structural experience. During deployment, it monitors collaboration traces and applies bounded structural updates when the current organization becomes insufficient. These updates can modify agent roles, communication links, execution order, information visibility, and validation pathways while preserving the task interface and agent budget. We evaluate MANTA against representative single-agent and multi-agent baselines on five benchmarks spanning information seeking, tool use, planning, workflow execution, and mathematical reasoning. MANTA achieves the highest average score of 74.0, outperforming the strongest baseline by 5.8 percentage points and obtaining the best result on PlanCraft. These results show that inference-time self-improvement can extend to the architecture of collaboration itself.

ArXiv ID: 2607.28430

Authors: Xinxing Ren, Qianbo Zang, Ziyan Wang, Caelum Forder, Suman Deb, Peter Carroll, Zekun Guo

Abstract:

Understanding large codebases is a long-horizon task for Large Language Model (LLM) agents: answering a single question can require building and running the software, tracing execution across files, and synthesizing evidence over tens of minutes. On SWE-Atlas QnA, a benchmark of long-horizon questions over production repositories, a single Claude Code agent (Opus 4.6) resolves only 32.3% of tasks. Dividing the work among agents with clean contexts mitigates this limitation. However, the subtasks of code comprehension are interdependent. One agent's findings can rewrite another's task, so agents must coordinate during execution, not only at phase boundaries. Existing multi-agent systems support such exchange only between phases, through staged handoffs or synchronized rounds. Communication and work remain mutually exclusive. A discovery made mid-execution cannot be shared until the next boundary. We present AgentRadio, an asynchronous message-passing layer that equips coding-agent harnesses with three primitives: threads, messages, and waiting for mentions. The last runs as a background task, surfacing teammates' messages without interrupting foreground work, so each agent remains passively aware of its peers and folds new findings into its ongoing task. Under a five-phase protocol of division of labor and negotiation, four agents organized by AgentRadio resolve 62.1% of tasks, 29.8 points above a single agent and above Claude Code with the newer Opus 4.8 (57.2%). Rubric-level analysis shows the gain growing with task difficulty, consistent with mid-course correction as the underlying mechanism. Our code is available at https://github.com/Coral-Protocol/AgentRadio.

ArXiv ID: 2607.28147

Authors: Yu Cui, Wuli Yang, Yirui Shi, Junhao Xia, Hui Jiang, Lei Gao, Chenfu Bao

Abstract:

Autonomous multi-agent systems (AMAS) built on large language models (LLMs), such as Hermes, increasingly rely on inference-time harnesses to coordinate reasoning and action. Constructing these harnesses requires substantial engineering effort and computational resources, as they are iteratively optimized over a combinatorial search space while co-evolving with the underlying LLM. Inference-time harnesses therefore constitute valuable intellectual property (IP). Although prior work has investigated IP leakage in static multi-agent systems with pre-configured architectures, it remains unclear whether similar risks arise in AMAS, where harness behavior emerges dynamically during inference. To address this gap, we introduce Agent Harness Distillation (AHD), a framework for studying the security risks arising from inference-time harness extraction in AMAS. We formalize harness extraction as a new security problem and develop an evaluation framework for quantifying such risks. AHD extracts inference-time harness capabilities from a target agent through black-box interactions and consists of two stages. In the pre-distillation stage, AHD infers inference-time harness behaviors from the responses of the target agent and constructs an initial harness. In the post-distillation stage, AHD iteratively refines the initial harness to align with the behavioral patterns of the target agent. Experiments on real-world AMAS across multiple backbone LLMs demonstrate the effectiveness of AHD and reveal substantial IP leakage risks. We further propose a deception-based defense that reduces harness extraction effectiveness while preserving the utility of the protected agent. Our findings uncover a previously underexplored security threat to AMAS.

ArXiv ID: 2607.27958

Authors: Peilin Feng, Suorong Yang, Soujanya Poria

Abstract:

Memory is central to long-horizon LLM agents, yet existing memory systems primarily preserve interaction content rather than modeling which agents can be trusted and under what conditions. This limitation is particularly important in multi-agent systems, where a central model may be unable to directly verify plausible or correlated peer responses. We introduce $Σ$-Mem, an online reliability memory that records historical competence evidence for individual peers and peer relationship evidence across the peer set. Both forms of evidence are maintained as real symmetric states and updated from post-decision correctness feedback. By Weyl's inequality, the spectral change caused by each event-level update is bounded, enabling stable online adaptation without retraining the underlying models. $Σ$-Mem provides a general write-and-read interface: the same memory can be used for residual steering of a central model, response-free peer routing, or reliability-weighted voting. Across five Qwen-family models, $Σ$-Mem adapts to counterfactual reliability shifts and generalizes to unseen peers and task domains. Direct memory readouts also outperform majority voting and the best fixed peer over the full OOD evaluation set. Moreover, performance improves consistently as more correctness feedback becomes available, indicating that $Σ$-Mem progressively accumulates actionable reliability information. These results establish reliability memory as a reusable foundation for adaptive coordination in LLM-based multi-agent systems.

ArXiv ID: 2607.27942

Authors: Linus Sander, Fengjunjie Pan, Vahid Zolfaghari, Andre Schamschurko, Nenad Petrovic, Alois Knoll

Abstract:

LLM-based multi-agent systems have the potential to enable collective intelligence and scale toward solving highly complex tasks through coordinated ensembles of specialized agents. However, despite their theoretical potential, the architectural design space remains largely non-systematized and lacks broadly established design principles. Furthermore, the scalability characteristics of such systems are only partially understood so far. This paper makes two contributions. We first distill four design principles for scalable MAS architectures from a structured analysis of prior work: simplicity, elastic feedback, sequential workflows with optional loops, and summary-based communication. We operationalize these principles in a reference architecture whose topology is formalized as a constrained directed workflow graph, and we evaluate four configurations of increasing complexity on a standardized benchmark of terminal-based system engineering tasks using two LLMs of differing capability. Our findings show that scaling yields measurable accuracy improvements with approximately linear cost growth, but only when the underlying LLM exceeds a minimum capability threshold. Performance peaks at intermediate complexity, then degrades due to timeouts and evaluation limitations. In addition, persistent consistency issues emerge as a central challenge across all scaling levels. These results provide concrete design guidance for practitioners and highlight consistency and evaluation standardization as key targets for future research.

ArXiv ID: 2607.27852

Authors: Lanhao Zhao

Abstract:

This paper investigates the strong structural controllability of multi-agent networks. Based on the definition of equitable partitions, an upper bound for the strong structural controllable subspace (SSCS) is established. To reflect the physical significance of matrix weights where the state dimension is greater than one, the multi-agent system is modeled using higher-order dynamics. Furthermore, to address matrix singularity and asymmetric couplings, a matrix space basis decomposition method is proposed to transform the matrix-weighted network into layered scalar networks. Additionally, by extending this basis decomposition to the lower bound estimation, a layer-specific distance partition (LDP) is introduced. This formulation establishes a tighter Squeeze Theorem, narrowing the mathematical boundaries for the controllable subspace by capturing layer-specific structural delays. To systematically identify the optimal basis that minimizes the bounds gap, an algebraic algorithm based on null-space projection is formulated. Furthermore, by introducing pattern matrices and generic rank, the almost-everywhere existence of this optimal basis in the parameter space is rigorously proved, perfectly aligning with the definition of strong structural controllability. To break the NP-hard combinatorial bottleneck of manually pre-defining the targets, a polynomial-time automated discovery algorithm based on the multi-layer Weisfeiler-Lehman (WL) color refinement is proposed. Finally, the strong structural observability and invariant attributes of the network are evaluated. Numerical examples with asymmetric matrix weights and directed multi-layer topologies are provided to verify the derived theorems.

ArXiv ID: 2607.27822

Authors: Qi Wang, Long-Gang Pang, Shi Pu, Xin-Nian Wang

Abstract:

We enable large language model (LLM) agents to autonomously perform end-to-end hydrodynamic simulations of the quark-gluon plasma evolution and calculation of final hadron spectra in relativistic heavy-ion collisions. We design a meta skill that allows an agent to explore a project's source code, craft a specialized skill, and iteratively refine it. Applying this meta skill to the (3+1)D viscous hydrodynamic code CLVisc, the agent builds a CLVisc skill encoding its operational knowledge and then independently executes full scientific workflows: designing parameter scans, running simulations, comparing ensemble results, and producing publication-ready figures. Crucially, the agent draws on literature-informed heavy-ion physics to select physically meaningful observables and interpret outcomes without explicit instruction. We demonstrate the pipeline in two scenarios: temperature-dependent shear viscosity over entropy density $η/s$, and nuclear-structure effects in O+O collisions at $\sqrt{s_{\mathrm{NN}}} = 5.36$~TeV using four \textit{ab initio} descriptions of $^{16}$O. In both, the agent plans, executes, and analyzes autonomously, devising new initial-state observables to explain final observations and extract qualitative knowledge. The meta skill is agnostic to code versions and Monte Carlo generators, promising future multi-agent systems in high-energy nuclear physics.

ArXiv ID: 2607.27528

Authors: Cristian Leo, Anton Dykyi, Danny Cortegaca, Daniel Begimher, Prakash Jha

Abstract:

Threat modeling is essential for secure software development, yet manual analysis of cloud-native architectures is slow and demands scarce security expertise. We present ThreatForest, a multi-agent system that generates structured attack trees from source code repositories, maps attack steps to adversary tactics, techniques, and procedures (TTPs) from a pluggable set of frameworks (MITRE ATT&CK, CAPEC, and cloud-specific threat matrices), and synthesizes actionable mitigations. ThreatForest decomposes threat modeling into a multi-stage agent pipeline -- repository analysis, context refinement, threat generation, parallel attack-tree construction with TTP mapping and mitigation synthesis, and report generation -- orchestrated as a directed graph with deterministic verification gates, bounded retries, and three human-in-the-loop validation points. A domain-specific sentence-transformer maps each attack step to candidate techniques by cosine similarity; we show empirically that this embedding stage, not the surrounding pipeline, is the dominant accuracy bottleneck. We evaluate ThreatForest across seven application domains on a sixteen-dimension rubric, scored by a panel of independent LLM raters with an adversarial verification pass and expert review. Panel-measured quality reaches 0.63-0.68 (on a 0-1 scale) for threat statements, attack trees, and mitigations, but only 0.29 for embedding-only TTP mapping -- a gap stable across all seven domains that isolates the binding constraint. A controlled single-call baseline on the same model more than doubles mapping defensibility, pinning the limitation on the embedding encoder rather than the multi-agent design. To our knowledge, ThreatForest is the first end-to-end system that turns a code repository into TTP-mapped attack trees with evidence-based mitigations across adversary frameworks, with a reusable framework for benchmarking such systems.

ArXiv ID: 2607.27484

Authors: Jinwei Hu, Yi Qi, Xinmiao Huang, Youcheng Sun, Yi Dong, Xiaowei Huang

Abstract:

Reusable skills are becoming a standard interface for extending language agents with task procedures. Yet evaluators usually infer skill use from visible reasoning or the agent's own attribution. These signals show what the agent appears to use, not whether the skill changed its decision. We ask whether skill-augmented agents exhibit a \textbf{Reasoning Backroom}, a systematic gap between stated skill use and intervention-measured influence. We introduce BACKTRACE, an evaluation framework that pairs each skill-conditioned answer with a matched no-skill counterfactual, intervenes on skill meaning, wording, identity, content, and assignment, and elicits attribution only after the answer is committed. We instantiate the framework as BACKROOMBench, a verified testbed spanning controlled logic and competition mathematics, multiple skill conditions, single-agent and multi-agent settings, and diverse model families. Our evaluation reveals a pervasive provenance failure. Across models and domains, stated skill use often remains stable while causal reliance and signed utility vary, producing both silent uptake and performative use. Behavioral effects follow procedural content more reliably than displayed skill identity, whereas stated attributions respond strongly to artifact availability. Observational detectors based on direct skill-use claims, text mentions, trace similarity, and an LLM judge do not identify which decisions actually depend on the skill. In multi-agent systems, skill influence can survive communication even after its source is lost, while no-skill teams still name skills and sources that were never supplied. These findings establish the Reasoning Backroom as a general AI provenance problem whose audit requires intervention.

ArXiv ID: 2607.28609

Authors: Qiushi Sun, Kanzhi Cheng, Yian Wang, Bowen Yang, Hang Yan, Liheng Chen, Fangzhi Xu, Zichen Ding, Nuo Chen, Jialin Cao, Xingdong Gong, Zehao Li, Kaiming Jin, Xinfeng Yuan, Zhoumianze Liu, Jingyang Gong, Zhangyue Yin, Jiahui Gao, Zhiyong Wu, Tianbao Xie, Jianbing Zhang, Ben Kao, Lingpeng Kong

Abstract:

Computer-using agents (CUAs) are advancing rapidly across the digital world. A CUA trajectory records the agent's actions, states, and reasoning. Verifying whether it fulfilled the task instruction is central to CUA evaluation, data curation, and reinforcement learning. Neither human-written verifiers nor human annotators can provide such verification at scale, so the field increasingly turns to vision-language models (VLMs) as judges of CUA trajectories. But a fundamental question has long gone unexamined: are these VLM judges reliable enough? To study it systematically, we introduce OSReward, a realistic, high-quality benchmark that evaluates VLM judges on CUA trajectories. The trajectories come from diverse agent backbones executing human-verified instructions across platforms, then rigorously labeled with ground-truth verdicts through multi-stage human annotation. Building on it, we derive OSReward-Hard, a challenge set concentrating genuinely hard cases, and OSReward-Multi for fine-grained efficiency and alignment scoring. The most comprehensive evaluation of VLM judges to date finds even state-of-the-art models fall short of an ideal judge, sharing a systematic leniency bias that mislabels failed runs as successes. The few reliable enough to trust are too expensive to run at scale, while affordable open models trail far behind. To close this gap, we construct and release OS-Shepherd-100K, an open corpus of reasoning-annotated trajectory judgments for the CUA community. On it, we train OS-Shepherd (9B and 35B), open reward models that supply low-cost, stable, and reliable reward signals, matching commercial judges at 30-60% lower cost than the frontier. Extensive analyses further inform the design of reliable CUA reward at scale. Our code, benchmark, dataset, and model checkpoints are available at https://os-copilot.github.io/OSReward-Home/.

ArXiv ID: 2607.28466

Authors: Jia Yu, Yan Zhu, Yili He, Zilong Wang, Xinyang Jiang, Peiyao Fu, Ruijie Yang, Tianyi Chen, Siyuan Li, Zhihua Wang, Fei Wu, Quanlin Li, Xian Yang, Pinghong Zhou, Shuo Wang

Abstract:

Vision-language models remain underused in colonoscopy despite the rich expert descriptions recorded in routine reports. These reports document lesion appearance, size and location but summarise entire procedures rather than caption individual frames, leaving clinical findings only weakly linked to the corresponding images. Here we develop EndoCLIP, a colonoscopy vision-language foundation model trained on 125,756 lesion-level image-text pairs progressively recovered from 280,476 routine colonoscopy records. Across lesion-level image-text retrieval, structured report generation and six multi-centre clinical classification tasks, EndoCLIP outperforms general-purpose and biomedical vision-language encoders in both zero-shot and linear-probe settings. On benign-versus-malignant classification, its linear probe approaches the performance of expert readers in a blinded study involving 12 endoscopists. These results suggest that recovering finding-to-frame correspondence can transform routine documentation into scalable supervision, enabling clinical targets to be specified in language rather than separately annotated for each task.