Medical AI learns better from weaker, not stronger fact-checkers

Original: What Makes a Medical Checker Trainable? Diagnosing Signal Collapse and Reward Hacking in Checker-Guided RAG for Biomedical QA

Source: arxiv.org ↗