OSReward: Instituting Standardized Evaluation for Cross-Platform Computer-Use Reward Models
2026-07-30 • Artificial Intelligence
Artificial IntelligenceComputation and LanguageComputer Vision and Pattern Recognition
AI summaryⓘ
The authors study how well vision-language models (VLMs) can judge if computer-using agents (CUAs) successfully complete their tasks by analyzing their recorded actions and reasoning. They created OSReward, a benchmark with human-verified task results, to test these models fairly and found current leading models often mistakenly mark failures as successes. Because the most accurate models are costly and cheaper ones are less reliable, the authors developed OS-Shepherd, an open-source reward model that offers reliable judgments at a lower cost. They also provide datasets and tools to help improve large-scale evaluation of CUAs in the future.
Computer-using agentsTrajectory verificationVision-language modelsBenchmarkingReward modelsTask evaluationReinforcement learningHuman annotationModel reliabilityCost-efficiency
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/.