Research map / Large language models / Retrieval and knowledge
KnowGPT: Knowledge Graph based Prompting for Large Language Models
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 KnowGPT, a black-box KG prompting framework that uses deep RL to extract reasoning paths from knowledge graphs and a multi-armed bandit to pick the knowledge extraction strategy and prompt template per question, applied to GPT-3.5.
Problem
LLMs hallucinate on domain-specific questions, and existing KG prompting methods for closed-source LLMs suffer from huge KG search space, high API cost, and manually engineered hard prompts.
Approach
Uses an RL policy over KG paths with reachability, context-relatedness, and conciseness rewards to extract concise knowledge, then an OLS/UCB-style multi-armed bandit that selects among extraction strategies and prompt formats (triples, sentences, graph description) per question, using only LLM API access.
Key results
- KnowGPT reaches 82.7/81.8 (IHdev/IHtest) on CommonsenseQA, 92.4 on OpenBookQA test and 78.1 on MedQA test, versus 71.0 test for GPT-3.5 and 76.3 for GPT-4 on MedQA.
- On the OpenBookQA official leaderboard KnowGPT scores 0.926, above the reported human performance of 0.917 and above GenMC (0.898) without ensembling.
- The paper reports an average 23.7% accuracy gain over GPT-3.5 across the three datasets.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add KG-based prompting with KnowGPT to GPT-3.5 | GPT-3.5 · CommonsenseQA | accuracy: + 10.8% | Helped |
| add KG-based prompting with KnowGPT to GPT-3.5 | GPT-3.5 · OpenBookQA | accuracy: + 32.4% | Helped |
| add KG-based prompting with KnowGPT to GPT-3.5 | GPT-3.5 · MedQA | accuracy: + 29.4% | Helped |
| add RL-based knowledge extraction instead of heuristic subgraph extraction (ablation) | GPT-3.5 with Sentence prompt · CommonsenseQA IHtest | accuracy: 0.739 -> 0.800 | Helped |
| add RL-based knowledge extraction instead of heuristic subgraph extraction (ablation) | GPT-3.5 with Sentence prompt · OpenBookQA Test | accuracy: 0.865 -> 0.889 | Helped |
| add RL-based knowledge extraction instead of heuristic subgraph extraction (ablation) | GPT-3.5 with Sentence prompt · MedQA Test | accuracy: 0.695 -> 0.755 | Helped |
| use Sentence prompt instead of Triples with P_sub (ablation) | GPT-3.5 with P_sub extraction · MedQA Test | accuracy: 0.589 -> 0.695 | Helped |
| use Graph Description instead of Sentence prompt with P_sub (ablation) | GPT-3.5 with P_sub extraction · OpenBookQA Test | accuracy: 0.865 -> 0.871 | Helped |
| use Graph Description instead of Sentence prompt with P_RL (ablation) | GPT-3.5 with P_RL extraction · OpenBookQA Test | accuracy: 0.889 -> 0.906 | Helped |
| use Sentence prompt instead of Triples with P_RL (ablation) | GPT-3.5 with P_RL extraction · CommonsenseQA IHtest | accuracy: 0.769 -> 0.800 | Helped |
| use Graph Description instead of Sentence prompt with P_RL (ablation) | GPT-3.5 with P_RL extraction · MedQA Test | accuracy: 0.755 -> 0.762 | Helped |
| select prompt format by question type rather than use one format for all (ablation) | KnowGPT on CommonsenseQA · CommonsenseQA simple questions | accuracy: 94.12% | Mixed |
Surprising findings
- Llama3 (8B) zero-shot scores 0.697 on MedQA test, while GPT-3.5 scores 0.487 in the paper's table.
- Ablation shows the triple format performs worst overall, while sentence and graph-description formats perform better on different question types.
Limitations
- Noisy triples in automatically constructed KGs may mislead the LLM.
- RL-based extraction fails on long-tail entities with sparse neighborhoods, requiring a heuristic fallback.
Experimental setup
Datasets: CommonsenseQA OpenBookQA MedQA-USMLE ConceptNet USMLE knowledge graph
Benchmarks: CommonsenseQA OpenBookQA MedQA-USMLE
Metrics: accuracy
Baselines: Bert-base Bert-large RoBERTa-large MHGRN QA-GNN HamQA JointLK GreaseLM GrapeQA ChatGLM ChatGLM2 Baichuan-7B
Builds on: policy gradient upper confidence bound MINERVA-style KG path reasoning
Closest papers by meaning
- Graph Neural Prompting with Large Language Models (AAAI 2024 · similarity 0.90)
- A Systematic Exploration of Knowledge Graph Alignment with Large Language Models in Retrieval Augmented Generation (AAAI 2025 · similarity 0.89)
- Simple is Effective: The Roles of Graphs and Large Language Models in Knowledge-Graph-Based Retrieval-Augmented Generation (ICLR 2025 · similarity 0.89)
- Think-on-Graph 2.0: Deep and Faithful Large Language Model Reasoning with Knowledge-guided Retrieval Augmented Generation (ICLR 2025 · similarity 0.88)
- Think-on-Graph: Deep and Responsible Reasoning of Large Language Model on Knowledge Graph (ICLR 2024 · similarity 0.88)
- LightPROF: A Lightweight Reasoning Framework for Large Language Model on Knowledge Graph (AAAI 2025 · similarity 0.88)
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.