Woodpecker Distillation: Weak Models Diagnose Reasoning Bugs in Strong Models

A recent paper on arXiv (2608.05168) presents a novel approach called Woodpecker Distillation, designed as a weak-to-strong training framework that employs contrastive local interventions to address reasoning errors in large language models. The authors contend that many reasoning issues stem from localized bugs during intermediate steps rather than from overarching incompetence, and these issues can be remedied by integrating brief patches produced by a weak probe model. However, simply fine-tuning on these patches does not consistently embed the corrections; rather, the key insight lies in how the intervention alters the model's subsequent reasoning distribution. The method contrasts effective and ineffective weak-model patches at identical prefixes to create a corrective training signal. This paper was recently submitted to arXiv.