Meta researchers introduced a method that compresses repeated reasoning patterns into short, named procedures—“behaviors”—and then conditions models to use them at inference or distills them via fine-tuning. The result: up to 46% fewer reasoning tokens on MATH while matching or improving accuracy, and up to 10% accuracy gains in a self-improvement setting on AIME, without […]
The post Meta AI Proposes ‘Metacognitive Reuse’: Turning LLM Chains-of-Thought into a Procedural Handbook that Cuts Tokens by 46% appeared first on MarkTechPost.