arxivcs.LGcs.AI2026-07-20
Uncovering Latent Reasoning Strategies in Language Models
A language model $p_θ(y \mid x)$ trained on reasoning tasks learns to solve problems via multiple distinct strategies, yet these strategies are implicit and entangled within the model's response distribution. We study the problem of decomposing the response distribution of a give…