Research map / Large language models / Retrieval and knowledge
Structural Entropy Guided Agent for Detecting and Repairing Knowledge Deficiencies in LLMs
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes SENATOR, which uses an LLM agent doing MCTS over a biomedical knowledge graph (SPOKE) with a structural-entropy reward to find knowledge deficiencies, then synthesizes QA pairs on those paths for SFT of Llama-3-8B and Qwen2-7B.
Problem
LLMs remain weak in knowledge-intensive domains like medicine, and synthetic data generation is redundant because it does not target the model's actual knowledge gaps.
Approach
Uses per-triplet self-information from cloze-style probabilities, aggregated into a one-dimensional structural entropy over KG subgraphs, as the MCTS reward (PUCT selection) to steer exploration toward uncertain knowledge paths; the resulting paths guide QA synthesis for SFT.
Key results
- SENATOR synthetic data plus instruction tuning raises Llama-3-8B average accuracy from 42.31 to 47.38 (+11.98%) across the medical benchmarks.
- For Qwen2-7B, average accuracy rises from 48.32 to 52.74 (+9.15%), with GPQA Genetics improving from 32.5 to 40.0.
- Manual check of 501 synthetic QA pairs found 62.08% fully valid, with logical errors (19.56%) the main failure mode and hallucination errors at 1.59%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Fine-tune with instruction data only (ablation) | Llama-3-8B · medical benchmarks, average | accuracy: 42.31 → 42.39 | No clear effect |
| Fine-tune with instruction data only (ablation) | Qwen2-7B · medical benchmarks, average | accuracy: 48.32 → 47.67 | Hurt |
| Add SENATOR synthetic data to instruction tuning (ablation) | Llama-3-8B · medical benchmarks, average | accuracy: +5.07 | Helped |
| Add SENATOR synthetic data to instruction tuning (ablation) | Llama-3-8B · GPQA Genetics | accuracy: +7.5 points | Helped |
| Add SENATOR synthetic data to instruction tuning (ablation) | Llama-3-8B · GPQA Molecular Biology | accuracy: +3.71 points | Helped |
| Add SENATOR synthetic data to instruction tuning (ablation) | Qwen2-7B · GPQA Genetics | accuracy: 7.5-point increase | Helped |
| Add SENATOR synthetic data to instruction tuning (ablation) | Qwen2-7B · GPQA Molecular Biology | accuracy: 3.7-point gain | Helped |
| Increase the proportion of synthetic data while keeping instruction data fixed (ablation) | SENATOR models · medical benchmarks | weighted-average performance | Helped |
| Train on another model's SENATOR-generated synthetic data instead of own-model data (ablation) | SENATOR models · medical benchmarks | performance | Helped |
Surprising findings
- Instruction tuning on general medical instruction data alone gave no gain or a slight drop (Qwen2-7B 48.32 to 47.67).
- Swapping synthetic data between models (Llama-3 data for Qwen2 and vice versa) still corrected deficiencies, which the authors attribute to shared pretraining corpora.
Limitations
- Relies on a high-quality, human-curated domain KG (SPOKE), which may not exist in other domains.
- Only about 62% of sampled synthetic QA pairs are fully valid; multi-hop logical consistency is weak.
- No entity-type constraints in MCTS, leading to uneven domain coverage in MMLU subfields.
Experimental setup
Datasets: SPOKE knowledge graph medical instruction data (514k samples, ChatDoctor, MedQA, MedMCQA, PubMedQA, LiveQA, MedicationQA, UMLS) MedQA MedMCQA PubMedQA
Benchmarks: MedQA MedMCQA PubMedQA GPQA (genetics, molecular biology) MMLU
Metrics: accuracy average accuracy relative improvement
Baselines: Med-Alpaca PMC-LLaMA HuatuoGPT-II Baichuan2 Llama-2 Chat-Doctor
Builds on: Monte Carlo Tree Search PUCT structural entropy SPOKE knowledge graph
Closest papers by meaning
- metabench - A Sparse Benchmark of Reasoning and Knowledge in Large Language Models (ICLR 2025 · similarity 0.87)
- KoLA: Carefully Benchmarking World Knowledge of Large Language Models (ICLR 2024 · similarity 0.87)
- PertEval: Unveiling Real Knowledge Capacity of LLMs with Knowledge-Invariant Perturbations (NeurIPS 2024 · similarity 0.87)
- Understanding LLM Behaviors via Compression: Data Generation, Knowledge Acquisition and Scaling Laws (NeurIPS 2025 · similarity 0.87)
- Deliberation on Priors: Trustworthy Reasoning of Large Language Models on Knowledge Graphs (NeurIPS 2025 · similarity 0.87)
- From Static to Dynamic: Knowledge Metabolism for Large Language Models (AAAI 2024 · 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.