Reproduce “Meta-R1: Empowering Large Reasoning Models with Metacognition” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Haonan Dong, Haoran Ye, Wenhao Zhu, Kehan Jiang, Guojie Song

Meta-R1 integrates a metacognitive system, inspired by human "thinking about thinking," into Large Reasoning Models (LRMs) by using a smaller LLM as a meta-level monitor and controller. This framework orchestrates proactive planning, online regulation for error correction, and adaptive termination, yielding average accuracy improvements of 0.9% to 3.6% and token consumption reductions of 24.2% to 32.7% across mathematical reasoning benchmarks.

Key insights

  • Explicitly engineering a meta-level cognitive system, drawing from human metacognition, can effectively enhance the controllability, reliability, and flexibility of Large Reasoning Models.
  • Relatively small instruct models (e.g., 1.5B or 3B parameters) are sufficient and efficient to serve as the meta-level, effectively assessing problem difficulty and generating corrective advice without significant computational overhead.
  • The Online Metacognitive Regulation stage, which enables real-time monitoring and error correction between the meta and object levels, is critical for achieving performance gains, highlighting the importance of dynamic intervention during reasoning.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.