[arXiv]score: 0.12
M2R Framework Cuts Long-Form Hallucination via Dual-Level Retrieval
May 29, 2026
M2R (Micro-Macro Retrieval) addresses hallucination in long-form generation by combining coarse-grained external retrieval with a key-information repository built during reasoning, keeping critical facts proximal to outputs. The approach targets the known proximity-accuracy correlation in RALMs, where evidence distance from output degrades factual accuracy.
cs.CLcs.AI
HOW THIS AFFECTS YOU
●
builderIf the proximity-accuracy finding holds at scale, M2R offers a drop-in reasoning-loop modification for RAG pipelines prone to factual drift in long outputs.
●
researcherThe proximity-to-output hypothesis is operationalized here as a testable architectural constraint — worth examining the benchmark deltas against standard multi-turn RAG baselines.