跳到正文
arXiv cs.AI· Nils A. Herrmann, Leander Girrbach, Kirill Bykov, Zeynep Akata·· 4 小时前AI 评分40

LLM 是否存在内部多重对话?从 persona 视角研究推理

Do LLMs Experience an Internal Polylogue? Investigating Reasoning through the Lens of Personas

AI 导读

一项 arXiv 研究提出“polylogue”概念,将 LLM 的 persona 向量视为动态信号,追踪生成过程中 persona 方向与隐藏激活的对齐时序。在四个开源权重模型上的实验显示,polylogue 特征预测 MMLU-Pro 正确率的表现不亚于低维激活基线,同时保持可解释性。基于此的段落条件干预在四个模型中的三个上提升准确率,但降低了第四个,说明分阶段潜空间引导可行但尚不稳健。

正文 · 原文

View PDF HTML (experimental)

Abstract:Recent work shows that large language models (LLMs) encode behavioural traits ("personas") as linear directions in activation space, often called "persona vectors". Prior work has used such directions as static handles for behavioural steering. We instead treat them as dynamic signals: probes we can monitor and intervene on as reasoning unfolds. We use the term polylogue to denote the time series of alignments between persona vectors and hidden activations over the course of generation. Experiments across four open-weight models show that polylogue features predict correctness on MMLU-Pro on par with low-dimensional activation baselines, while remaining interpretable through their associated persona directions. They also suggest concrete steering targets, namely which latent directions to modulate at different stages of a response. We instantiate this as a simple paragraph-conditioned intervention that improves accuracy on three of four models but degrades the fourth, suggesting that stage-aware latent steering is possible but not yet robust. Together, this positions the polylogue as an interpretable tool for reasoning-time monitoring and intervention.
Subjects: Artificial Intelligence (cs.AI)
Cite as: arXiv:2605.09159 [cs.AI]
  (or arXiv:2605.09159v3 [cs.AI] for this version)
  https://doi.org/10.48550/arXiv.2605.09159

arXiv-issued DOI via DataCite

Submission history

From: Nils A. Herrmann [view email]
[v1] Sat, 9 May 2026 20:42:13 UTC (98 KB)
[v2] Mon, 27 Jul 2026 07:56:22 UTC (120 KB)
[v3] Fri, 13 Nov 2026 11:14:28 UTC (122 KB)

来源:arXiv cs.AI · arxiv.org