Browse Teach Curate

Program-Driven Self-Verification and Refinement for Large Language Model Self-Correction

A two-stage self-correction scheme in which a language model writes and self-executes a verification pseudo-program to check its own answer with structured logic (rather than free-form reflection), then jointly refines both the answer and the verification program so a faulty check cannot silently mislead the fix. Improves reliability of self-correction on instruction-following and mathematical reasoning without external feedback.

Sonny Text

ProgCo: Program Helps Self-Correction of Large Language Models Xiaoshuai Song*, Yanan Wu†, Weixun

Self-correction asks a large language model to check and fix its own answer without external feedback, but models often verify poorly and their wrong feedback misleads the revision, especially on com…