LLM 各向同性悬崖：选择题决策行为的几何特征研究
Isotropy Cliffs: The Geometric Signature of Decision-Making in Large Language Models
AI 导读
一项 arXiv 研究分析五个开源权重模型在选择题问答中的决策几何特征，发现决策关键的过渡层存在各向同性突变，并伴随主要的表征变化和任务相关聚类出现。该同步几何行为与下游准确率强相关（相关系数约 0.84），且该过渡对提示词变化稳健，或反映模型行为的一般机制。
正文 · 原文
Abstract:We investigate the geometry of decision-making in Multiple Choice Question Answering (MCQA) through the lens of isotropy. Analyzing five open-weight models across diverse datasets, we identify decision-critical transition layers characterized by a shift in isotropy, coinciding with a major representational change and the emergence of task-relevant clusters. We demonstrate that this synchronized geometric behavior is strongly correlated with downstream accuracy ($r\approx0.84$), displaying its relevance for successful decision-making. Furthermore, we show that this transition is robust to prompt variations, suggesting that it reflects a general mechanism of model behavior.
| Subjects: | Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2608.00828 [cs.AI] |
| (or arXiv:2608.00828v2 [cs.AI] for this version) | |
| https://doi.org/10.48550/arXiv.2608.00828 arXiv-issued DOI via DataCite |
Submission history
From: Okan Selim Coskun [view email]
[v1]
Sat, 1 Aug 2026 19:11:10 UTC (2,927 KB)
[v2]
Wed, 11 Nov 2026 15:45:01 UTC (2,989 KB)
来源：arXiv cs.AI · arxiv.org