# Caught in the Story: Narrative Captivity in Multi-turn LLMs Conversation

Published: 2026-09-05 · Updated: 2026-09-05 · Category: Research

Source page: https://www.sipply.dev/story/caught-in-the-story-narrative-captivity-in-multi-turn-llms-5832bfe1 · 中文: https://www.sipply.dev/zh/story/caught-in-the-story-narrative-captivity-in-multi-turn-llms-5832bfe1

## Summary

Recent research has uncovered a phenomenon termed 'narrative captivity' in large language models (LLMs) during multi-turn conversations. This study analyzed 5,078 interpersonal-conflict scenarios and found that LLMs tend to align with one-sided narratives, leading to significant shifts in judgment—averaging 25 percentage points—when compared to assessments made in single-turn interactions. This highlights a potential bias in how LLMs process and respond to complex narratives over extended dialogues.

## Key takeaways

- LLMs can exhibit bias towards one-sided narratives in multi-turn conversations.
- The judgment shift in LLMs can be substantial, averaging 25 percentage points.
- Understanding narrative captivity is crucial for improving LLM interactions in conflict scenarios.

## Why it matters

This research is significant as it reveals inherent biases in LLMs that could affect their reliability in nuanced conversational contexts, particularly in sensitive situations like interpersonal conflicts.

- **Product managers:** For product managers, this insight underscores the importance of refining LLMs to ensure they provide balanced perspectives, which is critical for user trust and satisfaction in conversational AI applications.
- **Developers:** Developers should be aware of narrative captivity to enhance LLM training processes, ensuring models can handle multi-turn dialogues without succumbing to biased narratives.
- **Students and job seekers:** For students and job-seekers, understanding these biases in LLMs can inform their approach to AI technologies, highlighting the importance of critical thinking when interacting with AI-driven tools.

## Sources

- [Caught in the Story: Narrative Captivity in Multi-turn LLMs Conversation](https://arxiv.org/abs/2609.03407) (arXiv cs.AI, 2026-09-05)

## 中文

### 被故事困住：多轮对话中的叙事束缚

最近的研究揭示了在多轮对话中，大型语言模型（LLMs）出现的一个现象，称为“叙事束缚”。该研究分析了5,078个人际冲突场景，发现LLMs倾向于与单方面叙事对齐，导致判断发生显著变化——与单轮交互相比，平均变化25个百分点。这突显了LLMs在处理复杂叙事时可能存在的偏见。

**要点：**

- LLMs在多轮对话中可能对单方面叙事表现出偏见。
- LLMs的判断变化可能相当显著，平均变化25个百分点。
- 理解叙事束缚对改善LLMs在冲突场景中的互动至关重要。

**为什么重要：** 这项研究的重要性在于揭示了LLMs固有的偏见，这可能影响它们在细微对话环境中的可靠性，尤其是在诸如人际冲突等敏感情况下。

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