arXiv cs.CY· Harsh Kumar, Karina Vold, Louis Tay, Ashton Anderson·· 12 小时前AI 评分38
诊断并修复 LLM 建议中的“人格坍缩”
Diagnosing and Repairing Persona Collapse in LLM Advice
AI 导读
研究提出 LLM 建议存在“人格坍缩”：在 1,280 个 Reddit 情境、14 类建议场景中，人类回复会系统变化，而三个前沿 LLM 默认采用同一种安慰型人格。对比 13 种训练方法后发现，结合人类示例引导的模型自写回复与质量反馈强化学习效果最佳，在 N=159 的人类研究中其预期长期帮助评分超过基线。模型能编码何时该安慰、质疑或给实用建议，但回复未能可靠体现这些区分。
正文 · 原文
Abstract:A good advisor comforts someone in crisis, challenges someone in denial, and gives practical guidance on a logistical question. We ask whether large language models (LLMs) vary their approach in this way. We define five advisory personas based on how comforting or challenging a reply is, and whether it supports a realistic understanding of the situation and informed decision-making. Across 1,280 Reddit situations in 14 advice contexts, human replies vary systematically, while three frontier LLMs default to a standard reassuring persona. We call this narrowing persona collapse. Training models to imitate human replies broadens persona use but often turns support into enabling, and challenge into cynicism. In a blinded study (N=178), expert raters judge these replies less favorably than the baseline model's on both immediate preference and potential longer-term help. We compare 13 training methods to investigate how models can learn different approaches without sacrificing advice quality. The best combination of persona fit and advice quality comes from learning from model-written replies guided by human examples, followed by reinforcement learning with quality feedback. In a second human study (N=159), raters place this model above the baseline for anticipated longer-term help. We find that models encode information about when to comfort, challenge, or give practical guidance, but their replies do not reliably reflect these distinctions. Our training narrows this representation-action gap, while activation steering alone does not reliably improve advice quality. Better LLM advice requires learning both which persona fits a situation and how to express it constructively.
| Comments: | Working draft |
| Subjects: | Computers and Society (cs.CY) |
| Cite as: | arXiv:2607.08326 [cs.CY] |
| (or arXiv:2607.08326v2 [cs.CY] for this version) | |
| https://doi.org/10.48550/arXiv.2607.08326 arXiv-issued DOI via DataCite |
Submission history
From: Harsh Kumar [view email]
[v1]
Thu, 9 Jul 2026 10:12:18 UTC (2,987 KB)
[v2]
Wed, 14 Oct 2026 18:28:50 UTC (2,495 KB)
来源：arXiv cs.CY · arxiv.org