Generalization of LRM chain-of-thought explanations

Determine whether chain-of-thought explanations generated by large reasoning models generalize in the sense of capturing problem-level patterns rather than model-specific idiosyncrasies, so that these explanations reflect generalizable reasoning independent of the particular large reasoning model that produced them.

Background

Large reasoning models (LRMs) produce chain-of-thought (CoT) text as they solve tasks, and these CoTs are often treated as human-readable explanations of the model’s reasoning. A central concern is whether such explanations capture general, task-level structure rather than artifacts unique to the specific model that generated them.

This paper proposes to operationalize generalization by testing whether a CoT produced by one LRM induces the same behavior when provided to other LRMs (cross-model consistency), and explores strategies like transfer and sentence-level ensemble CoTs. The explicit uncertainty highlighted in the abstract motivates this evaluation framework.