Yale Researchers Propose Framework to Decompose LLM Latent Reasoning Strategies

YaleUniversity · hf · 2026-07-30

Language models often employ multiple implicit strategies for reasoning tasks, but these strategies are typically entangled within the model's response distribution. A research team from Yale University proposes a novel method to decompose the response distribution of a pretrained model into a structured, strategy-conditioned representation.

The approach introduces a latent-variable factorization consisting of a router (mapping inputs to a distribution over latent strategies) and a generator (producing responses conditioned on the strategy). To overcome the posterior collapse issue common with standard variational inference, the researchers introduce a new variational objective. This objective measures fractional information gain relative to the base model's loss, concentrating reconstruction pressure on tokens with high base model surprisal to force the latent variables to encode strategy-relevant information.

The team also introduced a benchmark of multi-strategy algorithmic tasks. Experiments demonstrate that this objective successfully recovers latent codes aligned with distinct reference strategies while preserving the base model's original response distribution.

Original post →

More from Research

Research channel →