From Associations to Activations: Comparing Behavioral and Hidden-State Semantic Geometry in LLMs
Not on the map. Why?
Abstract (arXiv)
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.
Key facts. Cited by 2 papers in the map (data through 7 Oct 2026). Builds on 26 of its 57 references that are in the map. First posted on arXiv 31 Jan 2026.
Builds on
26 of 57 references.
References · 26
- 2018BERT: Pre-training of Deep Bidirectional Transformers for Language Understandingcited by 40,204
- 2019Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformercited by 14,491
- 2019Similarity of Neural Network Representations Revisitedcited by 1,695
- 2016Enriching Word Vectors with Subword Informationcited by 1,578
- 2019BERT Rediscovers the Classical NLP Pipelinecited by 1,046
- 2019How Contextual are Contextualized Word Representations? Comparing the Geometry of BERT, ELMo, and GPT-2 Embeddingscited by 770
- 2024The Platonic Representation Hypothesiscited by 362
- 2023Language Models Represent Space and Timecited by 306
- 2023Exploring Collaboration Mechanisms for LLM Agents: A Social Psychology Viewcited by 221
- 2024Stealing Part of a Production Language Modelcited by 152
- 2023Getting aligned on representational alignmentcited by 142
- 2021Memorization vs. Generalization: Quantifying Data Leakage in NLP Performance Evaluationcited by 69
14 more
- 2021WhiteningBERT: An Easy Unsupervised Sentence Embedding Approachcited by 68
- 2024Fantastic Semantics and Where to Find Them: Investigating Which Layers of Generative LLMs Reflect Lexical Semanticscited by 43
- 2023Instruction-tuning Aligns LLMs to the Human Braincited by 39
- 2019Blackbox meets blackbox: Representational Similarity and Stability Analysis of Neural Language Models and Brainscited by 37
- 2020Enriching ImageNet with Human Similarity Judgments and Psychological Embeddingscited by 23
- 2021A comparative evaluation and analysis of three generations of Distributional Semantic Modelscited by 19
- 2024ReSi: A Comprehensive Benchmark for Representational Similarity Measurescited by 18
- 2024Objective drives the consistency of representational similarity across datasetscited by 16
- 2023Conceptual structure coheres in human cognition but not in large language modelscited by 14
- 2025Clone What You Can't Steal: Black-Box LLM Replication via Logit Leakage and Distillationcited by 3
- 2024Measuring individual semantic networks: A simulation studycited by 2
- 2025A word association network methodology for evaluating implicit biases in LLMs compared to humanscited by 2
- 2026Where meaning lives: Layer-wise accessibility of psycholinguistic features in encoder and decoder language modelscited by 2
- 2025The truth is no diaper: Human and AI-generated associations to emotional wordscited by 1
Built on by
Citing papers per year: 2026: 2.
2 papers.