LeAct: Learning to Reason from Expert Actions
A new method called LeAct (Learning to reason from Actions) recovers chain-of-thought reasoning from expert systems that produce near-optimal actions without explanation. By treating the reasoning as a latent variable, LeAct optimizes candidate CoTs that improve the student's probability of reproducing the expert action. The approach is tested on imperfect-information games.
Key facts
- Modern reasoning models rely on reasoning data from human annotations or distilled from stronger LLMs.
- Expert systems like game engines and theorem provers produce near-optimal actions but do not write down chain-of-thought reasoning.
- LeAct treats the chain-of-thought as a latent variable and recovers it from the action alone.
- The student samples candidate CoTs for each expert action and retains those that improve its own probability of recovering the action.
- The method is evaluated on imperfect-information games.
Entities
—