Uncovering Latent Reasoning Strategies in Language Models

This paper studies how to decompose the implicit mixture of reasoning strategies learned by a language model into a structured, strategy-conditioned representation. It factorizes the response distribution into a router over latent strategies and a generator conditioned on the selected strategy. Because a generator initialized from the base model can already reproduce the original distribution without using the latent variable, standard variational inference suffers from posterior collapse. The proposed objective measures fractional information gain relative to the base model’s response loss and emphasizes high-surprisal tokens. On a benchmark of multi-strategy algorithmic tasks, the method recovers latent codes aligned with reference strategies while preserving the base model’s response distribution.