trajectory-judge: What Outcome-Only LLM Judges Miss on Agent Trajectories

AI Digest

本文揭示Outcome-Only LLM法官在评估代理轨迹时无法检测中间步骤故障，仅依赖结果导致高估检测能力。

This paper shows Outcome-Only LLM judges overestimate their ability to detect faults in agent trajectories by only considering final outcomes.

Key points

• 测试通过注入故障评估LLM法官的召回率 Test evaluates LLM judges' recall via fault injection
• 仅看结果的法官高估检测能力 Outcome-only judges overestimate detection capabilities
• 配对判别率能暴露真实检测能力 Paired discrimination reveals true detection ability
• 模型在检查每一步时仍存在漏检 Models miss faults even when checking each step
• 建议使用配对判别率和分类型评估 Recommend paired discrimination and type-based evaluation

Takeaway: 评估代理轨迹需结合过程与结果，避免仅依赖最终输出。 / Agent trajectory evaluation requires combining process and outcome analysis.

Why it matters 揭示LLM法官评估局限，为可靠代理开发提供改进方向。