跳到正文
原文
机器之心 JIQIZHIXIN· @jiqizhixin · X·本站收录· 原文发表· AI 动态 →· 扫描记录里的位置 →

Emory 与 USC 等机构推出 CLASHarness 基准,评测计算机操作智能体如何处理多智能体冲突

AI 导读

Emory 大学、USC 及合作机构的研究者推出 CLASHarness 基准,在可执行计算机环境中评测智能体如何处理多智能体冲突,论文见 https://arxiv.org/abs/2610.19128v1,代码开源于 https://github.com/google-research/CLASHarness。基准包含 600 个多轮任务,针对协作中的过时报告、矛盾指令与共享资源竞争,以分阶段生成流水线构造场景,先固定决定性的视觉事实再渲染,执行参考解法并从轨迹导出代码与语义检查;任务中的视觉证据需要智能体亲自查看,而非只凭同伴消息推断。对 15 个前沿模型的评测暴露了智能体在证据核查、主动修复与协调上的失败;在步数预算消融中,Gemini 3.8 Flash 的通过率从 20 步的 12.8% 升至 60 步的 30.7%。

相关论文
正文 · 原文

When a teammate's message contradicts the dashboard, which evidence should a computer-use agent trust?

Researchers from Emory University, USC and collaborating institutions introduce CLASHarness, an arXiv benchmark for resolving multi-agent conflicts in executable computer environments.

The bottleneck: Collaboration can involve stale reports, contradictory instructions and shared-resource contention. Detecting a contradiction is only the start; agents must repair state and complete the handoff.

The mechanism: A staged pipeline constructs scenarios, fixes decisive visual facts before rendering, executes reference solutions and derives code-based and semantic checks from those trajectories. Its benchmark contains 600 multi-round tasks, with visual evidence that agents must inspect rather than infer from peer messages alone.

The payoff: Evaluation of 15 frontier models exposes failures in evidence checking, active repair and coordination. In a step-budget ablation, Gemini 3.8 Flash rises from 12.8% pass rate at 20 steps to 30.7% at 60. These are results on generated sandbox scenarios; stronger checkpoint scores do not guarantee successful end-to-end coordination.

CLASHarness: Benchmarking Multi-Agent Conflict in Computer Use via a Controllable and Verifiable Generation Pipeline

Paper: https://arxiv.org/abs/2610.19128v1
Code: https://github.com/google-research/CLASHarness
Our Report: https://www.jiqizhixin.com/articles/2026-11-01

📬 #PapersAccepted by Jiqizhixin

来源:机器之心 JIQIZHIXIN · x.com