Event ·Official first-hand

STOCKTAKE:用公平预言衡量LLM智能体感知与行动差距

STOCKTAKE是一个26周供应链补货基准,基于部分可观测马尔可夫决策过程,含六个隐藏因子过程。它通过公平参考策略计算技能分数(0到1),并测量状态估计滞后和知道-做率。在50个种子压力情景下,Claude Sonnet 5、GPT-5.4、DeepSeek-V4-Pro和Grok 4.5检测84-88%隐藏失败,但技能分数从0.62到-0.23,两个模型低于症状盲基线。正确诊断的压力周中仍有34-43%缺货,且该比率与技能分数负相关。

Current read

这篇论文用STOCKTAKE基准把LLM智能体的‘没看懂’和‘看懂了却没行动’分开了,四个主流模型的表现差距比想象中大。

3 sources53° AI heatLast updated 2026/7/15 09:08:27

Evidence

3 sources
Primary sourcearXiv: DeepSeek
View source

Conflict check

Dedup data only shows that these sources discuss the same event, not that they agree. There is no structured supporting or refuting evidence yet, so no conflict is inferred automatically.