Label-Free Evaluation of LLMs via Decision Theory Axioms

A recent paper published on arXiv (2608.05015) introduces a method that leverages representation theorems from decision theory to assess and regularize large language models (LLMs) without relying on external labels or human input. The author contends that the 'if and only if' nature of these theorems allows for verification of axiom adherence based on the model's responses to synthetic choice scenarios. It elaborates on three specific applications: probabilistic coherence through de Finetti's theorem, preference rationality via Afriat's theorem, and subjective expected utility based on a theorem by Echenique and Saez. This method is both computationally efficient and comprehensive, as successful checks indicate that the model cannot be dismissed on rationality grounds by any subsequent tests using the same data. The paper is classified as a cross-type announcement and is accessible on arXiv.