Formalizing Latent Thoughts: Axiomatic Framework for Evaluating LLM Reasoning Representations
University of British Columbia
Introduces an axiomatic framework for evaluating latent thought representations in LLMs independent of downstream benchmark scores. Defines four axioms — Causality, Minimality, Separability, and Stability — with quantitative measures. Testing across 23 reasoning tasks on open-weight models reveals no model satisfies all four axioms simultaneously, and representations encode minimal information beyond what is already in input embeddings.
Why it matters
Provides a principled, benchmark-agnostic way to audit whether a model's internal 'thoughts' are meaningful — important for interpretability and chain-of-thought evaluation. 46 upvotes on HuggingFace Daily Papers on June 29, 2026.
Importance: 2/5
46 HF Daily upvotes; benchmark-agnostic interpretability audit framework for chain-of-thought models