Uncovering Latent Reasoning Strategies in Language Models
The paper tries to separate a model’s hidden problem-solving methods without changing its overall answers.
Altabaa and Lafferty frame a pretrained model’s reasoning output as a mix of latent strategies, with a router choosing among strategy codes and a generator producing responses conditioned on them. They argue ordinary variational training can ignore those codes because the base generator already models the response distribution. Their objective pushes the latent variable to explain high-surprisal tokens and strategy-specific variation. On a new multi-strategy algorithmic benchmark, they report codes that align with reference strategies while preserving the base model’s distribution. HF Daily Papers' note
Altabaa and Lafferty frame a pretrained model’s reasoning output as a mix of latent strategies, with a router choosing among strategy codes and a generator producing responses conditioned on them. They argue ordinary variational training can ignore those codes because the base generator already models the response distribution. Their objective pushes the latent variable to explain high-surprisal tokens and strategy-specific variation. On a new multi-strategy algorithmic benchmark, they report codes that align with reference strategies while preserving the base model’s distribution. HF Daily Papers' note
score 5