---
id: 20260505-T0-08
title: "LLM不善策略博弈的病根找到了：观察、信念与行动脱节"
title_en: "Why LLMs Struggle in Strategic Play: Disconnect Between Beliefs and Actions"
url: https://ai.daily.yangsir.net/daily/20260505-T0-08
issue_date: 2026-05-05
publish_date: 2026-05-04T04:00:00.000Z
category: research
source_name: "arXiv cs.CL (NLP)"
source_url: https://arxiv.org/abs/2605.00226
---

# LLM不善策略博弈的病根找到了：观察、信念与行动脱节

大语言模型（LLM）在处理谈判和政策制定等不完全信息下的策略决策任务时，表现往往不够稳定。研究揭示了其失败的根源：模型在“观察、信念与行动”之间出现了断裂。具体来说，模型虽然能获取正确的观察信息，却无法建立准确的内部信念，或者无法将信念转化为合理的行动。这一发现为开发者指明了优化方向，即在训练策略型智能体时，不能仅靠增加参数量，而应针对信念更新机制进行专项强化。

## English Version

**Why LLMs Struggle in Strategic Play: Disconnect Between Beliefs and Actions**

LLMs often fail inconsistently in strategic decision-making under incomplete information. This research identifies the root cause: a broken link between observations, beliefs, and actions. Models either fail to form accurate internal beliefs from observations or struggle to translate those beliefs into logical actions.

---

**来源**：[arXiv cs.CL (NLP)](https://arxiv.org/abs/2605.00226)

**详情页**：https://ai.daily.yangsir.net/daily/20260505-T0-08

---

*智语观潮 · Daily — https://ai.daily.yangsir.net/llms.txt*