Micro-Macro Retrieval: Reducing Long-Form Hallucination in Large Language Models

In three linesM2R (Micro-Macro Retrieval) is a retrieve-while-generate framework reducing hallucinations in long-form LLM generation. It combines macro retrieval (external evidence) and micro retrieval (key information from reasoning) to maintain proximity between factual data and outputs. Trained via reinforcement learning with rule-based rewards.

Summary generated by Claude — human-verified