Researchers propose axiomatic framework to evaluate latent thought representations in LLMs

A new arXiv preprint proposes an axiomatic evaluation framework to assess latent thought representations in large language models (LLMs), arguing that existing evaluations conflate representation quality with model capacity. The authors—Fahd Seddik and Fatemeh Fard—introduce four functional axioms—Causality, Minimality, Separability, and Stability—each paired with a quantitative metric computed directly on the model’s internal representations, independent of downstream task performance.

The framework was applied to audit open-weight LLMs across 23 reasoning tasks, including spatial reasoning and factual question answering. The study reports that no candidate model satisfies all four axioms simultaneously, indicating a structural gap in how these models encode intermediate reasoning steps. The authors also find that while latent representations can reliably distinguish the type of task being performed, they fail to differentiate between two distinct questions within the same task category.

Additionally, the representations were found to encode little information beyond what is already present in the input embedding, suggesting that current models may not be learning meaningful intermediate representations during inference. This failure pattern persisted across multiple model families, including dense models, reasoning-distilled variants, and RL-trained systems, implying that the observed limitations are not merely a function of model size or training procedure but reflect deeper architectural or conceptual constraints.

The paper contributes a methodological tool—an axiomatic framework—that enables researchers to evaluate latent thought representations without relying on downstream benchmark scores, which often mask representational failures by rewarding superficial performance gains.