Research map / Graph learning / Knowledge and temporal graphs
LG-Bench: A Graph-Structured Evaluation Benchmark for Life Science
Lu Sun, Xiangyi Zhang, Xiangyang Zhu, Zijian Chen, Qiang Hu, Yuan Tian, Guangtao Zhai
Read the paperCheck your idea against 137,217 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.
Introduces LG-Bench, a graph-structured benchmark of over 10,000 multiple-choice questions across medicine, biology, and chemistry that evaluates models over the topology of scientific knowledge.
Problem
Existing life-science evaluation benchmarks flatten interconnected scientific knowledge into independent question lists, ignoring its underlying topological structure.
Approach
Organizes evaluation questions as a knowledge graph and builds a weighted evaluation over that graph, rather than scoring isolated flat questions.
Key results
- LG-Bench contains over 10,000 high-quality multiple-choice questions across medicine, biology, and chemistry.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| increase model scale within Qwen2.5 (0.5B to 72B) | Qwen2.5 · LG-Bench Overall | KBS (lower is more balanced): 16.11 -> 10.26 | Improved |
| increase model scale within Qwen2.5 (14B to 72B) | Qwen2.5 · LG-Bench Overall | accuracy and GCS: 83.69/70.42 -> 86.05/74.52 | Improved |
| domain-specific medical fine-tuning (Llama3-Med42-8B vs Llama-3-8b) | Llama-3-8b · LG-Bench Medicine | GCS (%): 54.11 -> 61.60 | Improved |
| domain-specific medical fine-tuning | Llama-3-8b · LG-Bench Medicine | KBS: 35.65 -> 25.61 | Improved |
| domain-specific medical fine-tuning | Llama-3-8b · LG-Bench Chemistry | KBS: 27.97 -> 30.78 | Hurt |
| raise similarity threshold theta from 0.4 to 0.45, 0.5, 0.6 (sparser graph) (ablation) | LG-Bench graph · LG-Bench GCS | GCS and model rank order | No clear effect |
| vary hybrid similarity weight gamma between 0.4 and 0.6 (ablation) | LG-Bench graph · LG-Bench | model ranking: Qwen2.5-7B overall GCS 62.20 vs 62.68 | No clear effect |
| use graph-based GCS vs human expert ratings (450 questions) | LG-Bench metrics · 450-question human study | Spearman correlation: 0.94 for GCS, 0.90 for KBS | Improved |
Experimental setup
Benchmarks: LG-Bench
Metrics: weighted evaluation score
Closest papers by meaning
- Semantic-KG: Using Knowledge Graphs to Construct Benchmarks for Measuring Semantic Similarity (NeurIPS 2025 · similarity 0.86)
- SciEval: A Multi-Level Large Language Model Evaluation Benchmark for Scientific Research (AAAI 2024 · similarity 0.86)
- Reasoning-Enhanced Healthcare Predictions with Knowledge Graph Community Retrieval (ICLR 2025 · similarity 0.85)
- GraphVis: Boosting LLMs with Visual Knowledge Graph Integration (NeurIPS 2024 · similarity 0.85)
- Medmarks: A Comprehensive Open-Source LLM Benchmark Suite for Medical Tasks (NeurIPS 2026 · similarity 0.85)
- CGBench: Benchmarking Language Model Scientific Reasoning for Clinical Genetics Research (NeurIPS 2025 · similarity 0.84)
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.
