Research map / Large language models / Retrieval and knowledge
Probing the Knowledge Boundary: An Interactive Agentic Framework for Deep Knowledge Extraction
Yuheng Yang, Siqi Zhu, Tao Feng, Ge Liu, Jiaxuan You
Read the paperCheck your idea against 137,217 papers
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 |
|---|---|---|---|
| use Recursive Taxonomy L5W5 instead of Taxonomy L2W2 (ablation) | Llama-3.1-405B probing · Strategy comparison | yield ratio: 1.00 -> 5.98 | Improved |
| increase taxonomy width/depth from L2W2 to L3W3 (ablation) | Recursive Taxonomy · Strategy comparison | yield ratio: 1.00 -> 2.15 | Improved |
| replace Sequential probing with Recursive Taxonomy L5W5 | Llama-3.1-405B probing · Strategy comparison | yield ratio: 0.28 -> 5.98 | Improved |
| add self-reflection critic loop to sequential probing (ablation) | Sequential probing · Strategy comparison | yield ratio: 0.28 -> 0.32 | Improved |
| increase expert personas from N=3 to N=20 (ablation) | Multi-Perspective probing · Strategy comparison | yield ratio: 1.42 -> 4.19 | Improved |
| scale model from Llama-3.1-8B to 70B (ablation) | Taxonomy-L5W5 · Scaling, average of 3 domains | recall: 27.3% -> 36.9% | Improved |
| scale model from 70B to 405B (ablation) | Taxonomy-L5W5 · Scaling, average of 3 domains | recall: 36.9% -> 39.0% | Improved |
| scale model from 70B to 405B (ablation) | Taxonomy-L5W5 · Scaling, average of 3 domains | accuracy: 89.0% -> 87.7% | Hurt |
| scale model from 8B to 70B (ablation) | Taxonomy-L5W5 · Scaling, average of 3 domains | accuracy: 81.2% -> 89.0% | Improved |
| swap Qwen2.5-7B Instruct for RL-tuned Coder (ablation) | Qwen2.5-7B · Specialization, Probabilistic Methods | final recall: 66.7% -> 34.0% | Hurt |
| swap Qwen2.5-7B Instruct for RL-tuned Coder (ablation) | Qwen2.5-7B · Specialization, Deep Learning | final recall: 57.7% -> 42.2% | Hurt |
| swap Qwen2.5-7B Instruct for RL-tuned Coder (ablation) | Qwen2.5-7B · Specialization, Deep Learning | initial accuracy: 92.3% -> 96.4% | Improved |
Closest papers by meaning
- KARMA: Leveraging Multi-Agent LLMs for Automated Knowledge Graph Enrichment (NeurIPS 2025 · similarity 0.89)
- Structural Entropy Guided Agent for Detecting and Repairing Knowledge Deficiencies in LLMs (NeurIPS 2025 · similarity 0.87)
- Physics of Language Models: Part 3.1, Knowledge Storage and Extraction (ICML 2024 · similarity 0.87)
- Knowledge Card: Filling LLMs' Knowledge Gaps with Plug-in Specialized Language Models (ICLR 2024 · similarity 0.86)
- Can Compressed LLMs Truly Act? An Empirical Evaluation of Agentic Capabilities in LLM Compression (ICML 2025 · similarity 0.86)
- Synergistic Multi-Agent Framework with Trajectory Learning for Knowledge-Intensive Tasks (AAAI 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.
