CLASHarness：可控可验证的多智能体计算机使用冲突基准
原标题：CLASHarness: Benchmarking Multi-Agent Conflict in Computer Use via a Controllable and Verifiable Generation Pipeline
AI 摘要
该论文提出 CLASHarness，一个可控且可验证的多智能体冲突生成框架，用于评测视觉语言模型在计算机使用场景中的冲突解决能力。它通过分层场景规划、预提交多模态事实和执行编译验证，构建了包含 600 个多轮任务、20 个操作域、7970 个视觉资产和 3760 个混合检查器的基准。对 15 个前沿模型的评测揭示三大瓶颈：弱模型盲从误导信息而强模型会核查视觉证据；发现冲突远易于解决冲突；协调协议是普遍短板，即使前沿模型在收尾动作上失败率也超过 58%。
正文节选
0001 CLASHarness: Benchmarking Multi-Agent Conflict in Computer Use via a Controllable and Verifiable Generation Pipeline Abstract Vision-language models (VLMs) are increasingly deployed in multi-agent systems to automate complex computer-use workflows, prompting numerous benchmarks to evaluate and advance agent capabilities. However, existing benchmarks mostly assume consistent, cooperative environments. In contrast, real-world collaboration rarely runs on consistent information: peer or sub-ag