arXiv：cs.AI· Louis Schiekiera, Max Zimmer, Christophe Roux, Sebastian Pokutta, Fritz G\"unther·· 18 小时前
从关联到激活：比较 LLM 中行为与隐藏状态的语义几何
From Associations to Activations: Comparing Behavioral and Hidden-State Semantic Geometry in LLMs
AI 导读
研究者对八个指令微调 Transformer 模型开展相似度强制选择与自由联想两类心理语言学实验，覆盖共享的 5000 词词表，收集超 1750 万次试验以构建基于行为的相似度矩阵。
正文
Abstract:We investigate the extent to which an LLM's hidden-state geometry can be recovered from its behavior in psycholinguistic experiments. Across eight instruction-tuned transformer models, we run two experimental paradigms -- similarity-based forced choice and free association -- over a shared 5,000-word vocabulary, collecting 17.5M+ trials to build behavior-based similarity matrices. Using representational similarity analysis, we compare behavioral geometries to layerwise hidden-state similarity and benchmark against FastText, BERT, and cross-model consensus. We find that forced-choice behavior aligns substantially more with hidden-state geometry than free association. In a held-out-words regression, behavioral similarity (especially forced choice) predicts unseen hidden-state similarities beyond lexical baselines and cross-model consensus, indicating that behavior-only measurements retain recoverable information about internal semantic geometry. Finally, we discuss implications for the ability of behavioral tasks to uncover hidden cognitive states.
| Comments: | Published at the 43rd International Conference on Machine Learning (ICML 2026), Seoul, South Korea. 27 pages, 16 figures, 6 tables. Code: this https URL. Data: this https URL |
| Subjects: | Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Computation and Language (cs.CL) |
| Cite as: | arXiv:2602.00628 [cs.LG] |
| (or arXiv:2602.00628v3 [cs.LG] for this version) | |
| https://doi.org/10.48550/arXiv.2602.00628 arXiv-issued DOI via DataCite |
|
| Journal reference: | Proceedings of the 43rd International Conference on Machine Learning (ICML 2026), PMLR 306, 2026 |
Submission history
From: Louis Schiekiera [view email]
[v1]
Sat, 31 Jan 2026 09:50:55 UTC (2,387 KB)
[v2]
Mon, 16 Feb 2026 12:30:19 UTC (2,387 KB)
[v3]
Mon, 12 Oct 2026 22:47:25 UTC (2,409 KB)
来源：arXiv：cs.AI · arxiv.org