Research map / Large language models / Retrieval and knowledge
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
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add differentiated region-specific knowledge augmentation (CGKE) versus vanilla SFT | Qwen2.5-7B-Instruct · 11 QA benchmarks average | accuracy: 27.98 -> 29.15 | Improved |
| add Knowledge Confused data in CGKE (ablation) | CGKE · 11 QA average, Qwen2.5-7B-Instruct | accuracy: 28.23 -> 29.15 | Improved |
| add Knowledge Missing data in CGKE (ablation) | CGKE · 11 QA average, Qwen2.5-7B-Instruct | accuracy: 28.38 -> 29.15 | Improved |
| add Knowledge Mastered data in CGKE (ablation) | CGKE · 11 QA average, Qwen2.5-7B-Instruct | accuracy: 29.04 -> 29.15 | Improved |
| add cognitive consistency alignment (entropy calibration loss) to GRPO (ablation) | CDKC · 11 QA average, Qwen2.5-7B-Instruct | accuracy: 43.91 -> 44.97 | Improved |
| add CGKE knowledge expansion before calibration (ablation) | CDKC · 11 QA average, Qwen2.5-7B-Instruct | accuracy: 42.86 -> 44.97 | Improved |
| run second expansion-calibration round | CDKC · 11 QA average, Qwen2.5-7B-Instruct | accuracy: 44.97 -> 48.29 | Improved |
| run second expansion-calibration round | CDKC · 11 QA average, Llama-3.1-8B-Instruct | accuracy: 51.44 -> 55.50 | Improved |
| apply CDKC calibration instead of GRPO | CGKE model · 11 QA average, Llama-3.1-8B-Instruct | accuracy: 49.71 -> 51.44 | Improved |
| use CDKC instead of retrieval (RAG) on PopQA | Qwen2.5-7B-Instruct · PopQA | accuracy: 43.44 -> 28.27 | Hurt |
| use Know What SFT abstention method | Qwen2.5-7B-Instruct · self-knowledge benchmark answerable | FN rate: 97.95% FN | Hurt |
| apply CDKC calibration | Qwen2.5-7B-Instruct · self-knowledge benchmark | boundary recognition (TN): 12.21% increase over vanilla, 63.37% TP | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- BARREL — BARREL: Boundary-Aware Reasoning for Factual and Reliable LRMs (ICLR 2026) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Enhancing Elusive Clues in Knowledge Learning by Contrasting Attention of Language Models (AAAI 2025 · similarity 0.87)
- Adaptive Chameleon or Stubborn Sloth: Revealing the Behavior of Large Language Models in Knowledge Conflicts (ICLR 2024 · similarity 0.87)
- metabench - A Sparse Benchmark of Reasoning and Knowledge in Large Language Models (ICLR 2025 · similarity 0.87)
- When Models Know More Than They Can Explain: Quantifying Knowledge Transfer in Human-AI Collaboration (NeurIPS 2025 · similarity 0.87)
- Seeking Neural Nuggets: Knowledge Transfer in Large Language Models from a Parametric Perspective (ICLR 2024 · similarity 0.87)
- Two Experts Are All You Need for Steering Thinking: Reinforcing Cognitive Effort in MoE Reasoning Models Without Additional Training (NeurIPS 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
