OSReward：为跨平台计算机使用奖励模型建立标准化评估

AI 摘要

OSReward 是一个用于评估跨平台计算机使用智能体（CUA）轨迹的视觉语言模型（VLM）评判器的新基准。研究发现，即使是先进的 VLM 评判器也存在系统性宽松偏差，将失败运行误标为成功，且可靠模型成本过高。为此，研究团队构建了 OS-Shepherd-100K 语料库并训练了 OS-Shepherd 9B 和 35B 开放奖励模型，以低成本提供稳定可靠的奖励信号，性能可与商业评判器相媲美，成本降低 30-60 倍。

正文节选

OSReward: Instituting Standardized Evaluation for Cross-Platform Computer-Use Reward Models 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

发布时间：—

抓取时间：2026-08-07 15:19

来源机构：Hugging Face