# Study reveals LLM agents often repeat failed moves in games despite feedback

Published: 2026-10-23 · Updated: 2026-10-23 · Category: Research

Source page: https://www.sipply.dev/story/study-reveals-llm-agents-often-repeat-failed-moves-in-games-6253d4ba · 中文: https://www.sipply.dev/zh/story/study-reveals-llm-agents-often-repeat-failed-moves-in-games-6253d4ba

## Summary

A recent study published on arXiv explores the decision-making processes of large language model (LLM) agents in gaming contexts, particularly focusing on how feedback influences their subsequent actions. The research found that LLM agents are prone to repeating moves they previously executed unsuccessfully, especially when those moves are illegal rather than simply losing moves. This backfire effect suggests that the way feedback is structured can significantly impact learning outcomes in AI systems, leading to unintended consequences in gameplay strategies.

## Key takeaways

- LLM agents are more likely to repeat failed moves after receiving feedback.
- The backfire effect is more pronounced with illegal moves than losing moves.
- Feedback mechanisms in AI learning can lead to counterproductive behaviors.

## Why it matters

Understanding how feedback affects AI decision-making can inform better training protocols and improve AI performance in complex environments.

- **Product managers:** Product managers should consider the implications of feedback design in AI products to enhance user experience and effectiveness.
- **Developers:** Developers need to be aware of the potential for feedback to inadvertently reinforce poor decision-making in AI systems, guiding them to refine learning algorithms.
- **Students and job seekers:** Students and job-seekers should recognize the importance of feedback mechanisms in AI, as this knowledge can be crucial for careers in AI development and research.

## Sources

- [Feedback That Backfires in Games: Do LLM Players Repeat the Move They Just Watched Fail?](https://arxiv.org/abs/2610.30205) (arXiv cs.AI, 2026-10-23)

## 中文

### 研究显示，LLM代理在游戏中经常重复失败的动作，尽管有反馈

最近在arXiv上发表的一项研究探讨了大型语言模型（LLM）代理在游戏环境中的决策过程，特别关注反馈如何影响其后续行动。研究发现，LLM代理倾向于重复之前执行失败的动作，尤其是当这些动作是非法的，而不仅仅是失败的动作。这种反效果表明，反馈的结构方式可以显著影响AI系统的学习结果，从而在游戏策略中导致意想不到的后果。

**要点：**

- LLM代理在收到反馈后更可能重复失败的动作。
- 反效果在非法动作中比在失败动作中更为明显。
- AI学习中的反馈机制可能导致适得其反的行为。

**为什么重要：** 理解反馈如何影响AI决策可以为更好的训练方案提供依据，并提高AI在复杂环境中的表现。

---
Sipply: Sip AI News Simply. https://www.sipply.dev