Woodpecker Distillation: New Technique to Fix Reasoning Bugs in Large Language Models

It's a weak-to-strong training framework that learns from contrastive local interventions (successful and failed patches) by weak models at the same prefix, distilling the signal into the strong model to improve reasoning.

Instead of imitating patch text directly, it focuses on reshaping the strong model's future reasoning distribution by contrasting successful and failed interventions, leading to more effective internalization.

Yes, regional companies can enhance their language models for specialized tasks like math or coding without massive resources, using available weak models as diagnostic tools.

Experiments focused on mathematical reasoning; results may need extension to other domains like legal or medical reasoning, and reliance on weak models may limit complex corrections.