Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Know More, Know Clearer: A Meta-Cognitive Framework for Knowledge Augmentation in Large Language Models

Hao Chen, Ye He, Yuchun Fan, Yukun Yan, Zhenghao Liu, Qingfu Zhu, Maosong Sun, Wanxiang Che

TL;DR

This work addresses reliability gaps in knowledge augmentation for large language systems by leveraging meta-cognitive signals to differentiate knowledge states. It introduces Cognition-Guided Knowledge Expansion (CGKE) and Cognition-Driven Knowledge Calibration (CDKC) within a Group Relative Policy Optimization (GRPO) framework, enabling targeted knowledge expansion and confidence calibration. Empirical results across diverse QA benchmarks show substantial gains in accuracy and improved meta-cognitive alignment, with iterative expansion-calibration cycles delivering additional improvements and lower calibration error. The approach demonstrates robust gains across grounding states and model scales, highlighting practical implications for safer, more reliable knowledge elicitation in real-world deployments.

Abstract

Knowledge augmentation has significantly enhanced the performance of Large Language Models (LLMs) in knowledge-intensive tasks. However, existing methods typically operate on the simplistic premise that model performance equates with internal knowledge, overlooking the knowledge-confidence gaps that lead to overconfident errors or uncertain truths. To bridge this gap, we propose a novel meta-cognitive framework for reliable knowledge augmentation via differentiated intervention and alignment. Our approach leverages internal cognitive signals to partition the knowledge space into mastered, confused, and missing regions, guiding targeted knowledge expansion. Furthermore, we introduce a cognitive consistency mechanism to synchronize subjective certainty with objective accuracy, ensuring calibrated knowledge boundaries. Extensive experiments demonstrate the our framework consistently outperforms strong baselines, validating its rationality in not only enhancing knowledge capabilities but also fostering cognitive behaviors that better distinguish knowns from unknowns.

Know More, Know Clearer: A Meta-Cognitive Framework for Knowledge Augmentation in Large Language Models

TL;DR

This work addresses reliability gaps in knowledge augmentation for large language systems by leveraging meta-cognitive signals to differentiate knowledge states. It introduces Cognition-Guided Knowledge Expansion (CGKE) and Cognition-Driven Knowledge Calibration (CDKC) within a Group Relative Policy Optimization (GRPO) framework, enabling targeted knowledge expansion and confidence calibration. Empirical results across diverse QA benchmarks show substantial gains in accuracy and improved meta-cognitive alignment, with iterative expansion-calibration cycles delivering additional improvements and lower calibration error. The approach demonstrates robust gains across grounding states and model scales, highlighting practical implications for safer, more reliable knowledge elicitation in real-world deployments.

Abstract

Knowledge augmentation has significantly enhanced the performance of Large Language Models (LLMs) in knowledge-intensive tasks. However, existing methods typically operate on the simplistic premise that model performance equates with internal knowledge, overlooking the knowledge-confidence gaps that lead to overconfident errors or uncertain truths. To bridge this gap, we propose a novel meta-cognitive framework for reliable knowledge augmentation via differentiated intervention and alignment. Our approach leverages internal cognitive signals to partition the knowledge space into mastered, confused, and missing regions, guiding targeted knowledge expansion. Furthermore, we introduce a cognitive consistency mechanism to synchronize subjective certainty with objective accuracy, ensuring calibrated knowledge boundaries. Extensive experiments demonstrate the our framework consistently outperforms strong baselines, validating its rationality in not only enhancing knowledge capabilities but also fostering cognitive behaviors that better distinguish knowns from unknowns.
Paper Structure (36 sections, 2 theorems, 32 equations, 12 figures, 9 tables)

This paper contains 36 sections, 2 theorems, 32 equations, 12 figures, 9 tables.

Table of Contents

  1. Preliminaries
  2. Meta-Cognition: Foundations and Empirical Laws
  3. Group Relative Policy Optimization (GRPO)
  4. Method
  5. Cognition-Guided Knowledge Expansion (CGKE)
  6. Cognition-Driven Knowledge Calibration (CDKC)
  7. Experimental Methodology
  8. Results and Analysis
  9. Overall Performance
  10. Ablation Study
  11. Meta-cognition Assessment in Self-Knowledge Task
  12. Analysis of the Optimized Structural Decay Law
  13. Related work
  14. Conclusion
  15. Theoretical Analysis of Meta-cognitive Calibration
  16. ...and 21 more sections

Key Result

Proposition 1.1

The policy $\pi_\theta$ successfully breaks free from the conservative “refusal trap” if the magnitude of the calibration gradient overrides the gravitational pull of the reference prior, satisfying:

Figures (12)

  • Figure 1: The Structural Decay Law between Accuracy and Uncertainty in Qwen2.5-7B-Instruct across Various QA Tasks.
  • Figure 2: Overview of the Meta-Cognitive Knowledge Augmentation Framework. The Cognition-Guided Knowledge Expansion (CGKE) module enables differentiated knowledge augmentation guided by internal cognitive signals. The Cognition-Driven Knowledge Calibration (CDKC) module calibrates subjective confidence with objective correctness, promoting clearer cognitive knowledge boundaries.
  • Figure 3: Distribution of Cognitive Decision States. Comparison on (a) answerable and (b) unanswerable questions across methods.
  • Figure 4: Performance Comparison across Different Metrics. Left: Answerable accuracy. Right: Multi-dimensional evaluation on AR, KEI, CBS, CAE, and NPV metrics.
  • Figure 5: Optimization of the Structural Decay Law under CDKC.
  • ...and 7 more figures

Theorems & Definitions (2)

  • Proposition 1.1: Escape Velocity Condition
  • Proposition 1.2: Manifold Flattening Condition