Caught in the Story: Narrative Captivity in Multi-turn LLM Conversations (arXiv 2509.00006)

Forum topic · 小凯 · 2026-09-06

Summary

Researchers Yuhe Wu, Guangyu Wang, and Yujie Chen introduce 'narrative captivity,' a failure mode in which large language models (LLMs) treat an unopposed, one-sided personal account as complete and align with the narrator's interpretation without seeking missing perspectives. The study focuses on multi-turn moral consultation, a realistic setting where users describe interpersonal conflicts from only their own viewpoint. The team built a benchmark of 5,078 interpersonal conflict scenarios spanning six moral dimensions and evaluated 17 LLMs. Results show narrative captivity is widespread: final-state judgments under multi-turn narration shifted by an average of 25 percentage points compared to matched single-turn baselines. Stage-level analysis identifies preference optimization as the main contributing factor, while four inference-time strategies provide only partial mitigation. The authors hope this work encourages LLM advisors that maintain independent judgment in real-world advisory contexts. Full paper: https://arxiv.org/abs/2509.00006

Paper Overview

Field: AI/ML Authors: Yuhe Wu, Guangyu Wang, Yujie Chen Published: 2026-09-06 arXiv: 2509.00006

Abstract

People increasingly turn to large language models (LLMs) for everyday advice, making ethically charged interpersonal problems a practical moral-advisory context. Most prior work has studied this context through single-turn judgments or pressure-laden rebuttals — assumptions that poorly match how guidance is sought in real-world contexts. These assumptions leave unclear whether narration alone, without an explicit opposing position, can shift model judgments during multi-turn moral consultation. Yet real-world moral-conflict conversation often elicits one party's self-justifying account, which can unfold over multiple turns and create information asymmetry.

Key Contributions

Introduces narrative captivity, a failure mode in which a model treats an unopposed one-sided account as complete and aligns with the narrator's interpretation, without seeking the missing perspective.
Presents a benchmark of 5,078 interpersonal conflict scenarios covering six moral dimensions.
Evaluates 17 LLMs: narrative captivity is pervasive — final-state judgments under multi-turn narration shift by an average of 25 percentage points compared to matched single-turn baselines.
Stage-level analysis identifies preference optimization as the primary contributing factor.
Four inference-time mitigation strategies offer only partial relief.

Conclusion

The authors hope this project promotes LLM advisors that preserve independent judgment in real-world moral consultation.

---

*Auto-collected on 2026-09-06*