arXiv：cs.AI· Celia Cintas, Miriam Rateike, Erik Miehling, Elizabeth Daly, Skyler Speakman·· 17 小时前
LLM 中人格表征的定位研究：人格特征编码在解码器层最后三分之一
Localizing Persona Representations in LLMs
AI 导读
研究揭示，多个预训练 decoder-only LLM 仅在解码器层最后三分之一才对不同人格（由人类特征、价值观和信念定义）表现出显著的表征空间差异。伦理视角如道德虚无主义与功利主义的激活存在重叠，呈现多义性，而保守主义与自由主义等政治意识形态则分布在更独立的区域。该发现有助于理解 LLM 内部信息表征，并为调控模型输出中特定人类特质提供参考。
正文
Abstract:We present a study on how and where personas -- defined by distinct sets of human characteristics, values, and beliefs -- are encoded in the representation space of large language models (LLMs). Using a range of dimension reduction and pattern recognition methods, we first identify the model layers that show the greatest divergence in encoding these representations. We then analyze the activations within a selected layer to examine how specific personas are encoded relative to others, including their shared and distinct embedding spaces. We find that, across multiple pre-trained decoder-only LLMs, the analyzed personas show large differences in representation space only within the final third of the decoder layers. We observe overlapping activations for specific ethical perspectives -- such as moral nihilism and utilitarianism -- suggesting a degree of polysemy. In contrast, political ideologies like conservatism and liberalism appear to be represented in more distinct regions. These findings help to improve our understanding of how LLMs internally represent information and can inform future efforts in refining the modulation of specific human traits in LLM outputs. Warning: This paper includes potentially offensive sample statements.
| Comments: | Corrected small naming inconsistencies for Level 0, 1, and 2 analysis |
| Subjects: | Computation and Language (cs.CL); Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2505.24539 [cs.CL] |
| (or arXiv:2505.24539v5 [cs.CL] for this version) | |
| https://doi.org/10.48550/arXiv.2505.24539 arXiv-issued DOI via DataCite |
|
| Journal reference: | Proceedings of the AAAI ACM Conference on AI, Ethics, and Society, 8(1), 630-642, 2025 |
| Related DOI: | https://doi.org/10.1609/aies.v8i1.36577
DOI(s) linking to related resources |
Submission history
From: Miriam Rateike [view email]
[v1]
Fri, 30 May 2025 12:46:44 UTC (5,521 KB)
[v2]
Tue, 3 Jun 2025 08:45:28 UTC (5,523 KB)
[v3]
Mon, 8 Sep 2025 18:14:07 UTC (5,523 KB)
[v4]
Tue, 28 Jul 2026 16:02:36 UTC (5,997 KB)
[v5]
Sat, 10 Oct 2026 15:44:47 UTC (5,997 KB)
来源：arXiv：cs.AI · arxiv.org