Metacognitive Reuse: Turning Recurring LLM Reasoning Into Concise Behaviors

Aniket Didolkar, Nicolas Ballas, Sanjeev Arora, Anirudh Goyal

Introduction

While Chain-of-Thought (CoT) prompting has significantly advanced the reasoning capabilities of Large Language Models (LLMs), it remains structurally inefficient. Current models often re-derive common sub-procedures—such as geometric formulas or unit conversions—from first principles every time they encounter a new problem. This "reinventing the wheel" leads to inflated token counts, increased latency, and unnecessary consumption of the context window.

To address this, researchers from Meta, Mila, and Princeton have proposed a "metacognitive pathway." This approach allows LLMs to extract reusable reasoning patterns, termed behaviors, from their own problem-solving traces. By shifting from purely declarative knowledge retrieval (like standard RAG) to procedural knowledge reuse, models can bridge the gap between slow, deliberative "System 2" thinking and fast, intuitive "System 1" execution.

Methodology: The Behavior Curation Pipeline

The core of the framework is a three-stage pipeline designed to transform raw reasoning traces into a structured Behavior Handbook.

  1. Solve: A "Metacognitive Strategist" model solves a complex problem using standard reasoning.
  2. Reflect: The model critiques its own solution to identify generalizable steps or missing procedural shortcuts that could have streamlined the derivation.
  3. Extract: These reflections are converted into named behaviors (e.g., behavior_inclusion_exclusion_principle) paired with concise, actionable instructions.
The complete metacognitive framework, illustrating the curation pipeline and the three primary modes of behavior utilization.

Application Strategies

The resulting Behavior Handbook is utilized through three distinct mechanisms:

  • Behavior-Conditioned Inference (BCI): Relevant behaviors are retrieved via embedding similarity and provided in-context, allowing the model to solve new problems more concisely by referencing the handbook.
  • Behavior-Guided Self-Improvement: The model uses behaviors extracted from its own past failures to generate more accurate second attempts, effectively learning from its own mistakes.
  • Behavior-Conditioned SFT (BC-SFT): A student model is fine-tuned on reasoning traces generated by a teacher model that was guided by these behaviors. This process internalizes procedural knowledge directly into the model's weights.

Key Insights and Results

The researchers evaluated this approach on challenging benchmarks including MATH and AIME (2022-2025). The results demonstrate that metacognitive reuse provides significant gains in both efficiency and performance.

  • Efficiency Gains: In BCI settings, models achieved comparable or superior accuracy while using up to 46% fewer reasoning tokens. By providing a named behavior as a "procedural hint," the model naturally skips long-winded derivations, making conciseness an emergent property of the system.
  • Accuracy Improvements: In self-improvement tasks, behavior-based hints led to a 10% accuracy boost over standard critique-and-revise baselines.
  • Distillation Quality: BC-SFT proved more effective than vanilla SFT at transforming base models into capable reasoners. The behaviors act as a scaffolding that helps the model learn the underlying structure of logical arguments rather than just mimicking surface-level text.

Discussion and Future Work

Metacognitive reuse represents a shift toward closed-loop learning, where models improve by reflecting on their own experiences. However, several challenges remain. Currently, behaviors are retrieved statically at the start of a prompt. Future iterations could implement dynamic retrieval, allowing a model to call upon the handbook as a tool mid-reasoning. Additionally, while highly effective in mathematics, the framework's applicability to more open-ended domains like creative writing or complex coding requires further exploration. As these handbooks scale to thousands of behaviors, managing retrieval latency and context clutter will be critical for production-level deployment.

Summary updated

BibTeX
@misc{didolkar2025metacognitive,
  author = {Aniket Didolkar and Nicolas Ballas and Sanjeev Arora and Anirudh Goyal},
  title = {{Metacognitive Reuse: Turning Recurring LLM Reasoning Into Concise Behaviors}},
  year = {2025},
  eprint = {2509.13237},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2509.13237},
}
Figure