Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Learn Like Humans: Use Meta-cognitive Reflection for Efficient Self-Improvement

Xinmeng Hou, Peiliang Gong, Bohao Qu, Wuqi Wang, Qing Guo, Yang Liu

TL;DR

The paper addresses the inefficiency of traditional self-improving agents that rely on fixed prompts and costly multi-turn recurrences. It introduces Metacognitive Agent Reflective Self-improvement (MARS), a three-phase, single-cycle framework that blends principle-based (error avoidance) and procedural (step-by-step strategy) reflection, producing optimized prompts via structured failure analysis and weighted synthesis. The approach yields targeted prompt enhancements—concise, reasoning, and their combination—applied through a category-aware hybrid strategy, achieving superior results across six benchmarks with significantly reduced computational overhead compared to recursive baselines. This work demonstrates that human-inspired metacognition can enable efficient, scalable self-improvement for LLM-driven agents, with strong practical implications for deploying capable autonomous systems at lower cost, while leaving room for extending the taxonomy and iterative use in future work.

Abstract

While Large Language Models (LLMs) enable complex autonomous behavior, current agents remain constrained by static, human-designed prompts that limit adaptability. Existing self-improving frameworks attempt to bridge this gap but typically rely on inefficient, multi-turn recursive loops that incur high computational costs. To address this, we propose Metacognitive Agent Reflective Self-improvement (MARS), a framework that achieves efficient self-evolution within a single recurrence cycle. Inspired by educational psychology, MARS mimics human learning by integrating principle-based reflection (abstracting normative rules to avoid errors) and procedural reflection (deriving step-by-step strategies for success). By synthesizing these insights into optimized instructions, MARS allows agents to systematically refine their reasoning logic without continuous online feedback. Extensive experiments on six benchmarks demonstrate that MARS outperforms state-of-the-art self-evolving systems while significantly reducing computational overhead.

Learn Like Humans: Use Meta-cognitive Reflection for Efficient Self-Improvement

TL;DR

The paper addresses the inefficiency of traditional self-improving agents that rely on fixed prompts and costly multi-turn recurrences. It introduces Metacognitive Agent Reflective Self-improvement (MARS), a three-phase, single-cycle framework that blends principle-based (error avoidance) and procedural (step-by-step strategy) reflection, producing optimized prompts via structured failure analysis and weighted synthesis. The approach yields targeted prompt enhancements—concise, reasoning, and their combination—applied through a category-aware hybrid strategy, achieving superior results across six benchmarks with significantly reduced computational overhead compared to recursive baselines. This work demonstrates that human-inspired metacognition can enable efficient, scalable self-improvement for LLM-driven agents, with strong practical implications for deploying capable autonomous systems at lower cost, while leaving room for extending the taxonomy and iterative use in future work.

Abstract

While Large Language Models (LLMs) enable complex autonomous behavior, current agents remain constrained by static, human-designed prompts that limit adaptability. Existing self-improving frameworks attempt to bridge this gap but typically rely on inefficient, multi-turn recursive loops that incur high computational costs. To address this, we propose Metacognitive Agent Reflective Self-improvement (MARS), a framework that achieves efficient self-evolution within a single recurrence cycle. Inspired by educational psychology, MARS mimics human learning by integrating principle-based reflection (abstracting normative rules to avoid errors) and procedural reflection (deriving step-by-step strategies for success). By synthesizing these insights into optimized instructions, MARS allows agents to systematically refine their reasoning logic without continuous online feedback. Extensive experiments on six benchmarks demonstrate that MARS outperforms state-of-the-art self-evolving systems while significantly reducing computational overhead.
Paper Structure (45 sections, 6 equations, 13 figures, 14 tables, 1 algorithm)

This paper contains 45 sections, 6 equations, 13 figures, 14 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methodology
  4. Evaluation
  5. Failure Allocation
  6. Enhancement Generation
  7. Experiments
  8. Datasets.
  9. Implementation.
  10. Results and Analysis
  11. Main Results
  12. Generalization to Challenging Benchmarks
  13. Performance Gain Analysis
  14. Conclusion
  15. Algorithm
  16. ...and 30 more sections

Figures (13)

  • Figure 1: The Cognitive Inspiration behind MARS. This framework parallels human reflection with the MARS (Metacognitive Agent with Reflective Self-improvement) agent, converting baseline agent failures into principled-based and procedural instructions to synthesize enhanced prompts.
  • Figure 2: Overview of the proposed framework: (1) diagnose failed questions into structured analyses $\mathcal{A}_i$, (2) group by type-topic keys and aggregate error profiles $\Psi_j$, and (3) synthesize enhancements via weighted aggregation to produce $P'^{(c)}$ and $P'^{(r)}$.
  • Figure 3: Performance gain analysis. (a) Inverse relationship between task difficulty and enhancement effectiveness. (b) Performance landscape comparison between Knowledge Coverage (blue) and Reasoning Capacity (red).
  • Figure 4: Zero-shot prompt template.
  • Figure 5: Zero-shot chain-of-thought prompt template.
  • ...and 8 more figures