Research map / Large language models / Retrieval and knowledge
Mitigating Large Language Model Hallucinations via Autonomous Knowledge Graph-Based Retrofitting
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 KGR, a framework where an LLM itself extracts claims from a draft response, detects entities, retrieves Wikidata triples, selects facts, verifies claims, and retrofits the response iteratively to reduce factual hallucination.
Problem
LLMs produce factual hallucinations during multi-step reasoning, and existing KG-augmented methods only retrieve facts for entities in the user query, missing facts used in intermediate reasoning steps.
Approach
Instead of querying the KG with the user question, KGR uses the LLM to decompose the draft response into atomic claims, retrieve local KG subgraphs for claim entities, select relevant triples via chunked LLM fact selection, verify claims, and retrofit the draft, repeating this loop over multiple rounds.
Key results
- On Mintaka with ChatGPT, KGR reaches 53.1 EM / 60.8 F1 versus 42.9/56.1 for vanilla and 48.0/54.6 for QKR.
- On SimpleQuestions with text-davinci-003, QKR (56.0/62.0) slightly outperforms KGR (54.0/57.6), which the authors attribute to single-hop questions favoring question-relevant retrieval.
- On Vicuna 13B, KGR achieves 46.0 EM / 46.9 F1 on SimpleQuestions versus 14.0/21.9 for CoT, and 26.5/34.0 F1 on Mintaka versus 28.3 for CoT.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use response-relevant KGR instead of question-relevant QKR | ChatGPT · Mintaka | F1: 59.3 -> 60.8 | Helped |
| Use response-relevant KGR instead of question-relevant QKR | text-davinci-003 · Mintaka | F1: 51.7 -> 60.2 | Helped |
| Use response-relevant KGR instead of question-relevant QKR | ChatGPT · HotpotQA | F1: 38.1 -> 39.2 | Helped |
| Use response-relevant KGR instead of question-relevant QKR | text-davinci-003 · HotpotQA | F1: 31.9 -> 47.2 | Helped |
| Use response-relevant KGR instead of question-relevant QKR | ChatGPT · Simple Question | F1: 60.2 -> 60.7 | Helped |
| Use response-relevant KGR instead of question-relevant QKR | text-davinci-003 · Simple Question | F1: 62.0 -> 57.6 | Hurt |
| Use response-relevant KGR instead of question-relevant QKR | Vicuna 13B · HotpotQA | EM: 12.2 -> 10.2 | Hurt |
| Reduce fact-selection chunk size (ablation) | KGR · Simple Question (ChatGPT) | precision | Hurt |
| Reduce fact-selection chunk size (ablation) | KGR · Simple Question (ChatGPT) | recall | Helped |
| Use chunk size 100 (ablation) | KGR · Simple Question (ChatGPT) | triple-selection capability | Hurt |
| Randomize the order of retrieved triples (ablation) | KGR · Simple Question (ChatGPT) | triple-selection capability | No clear effect |
| Increase the number of retrieved triples (ablation) | KGR · Simple Question (ChatGPT) | recall | Helped |
Surprising findings
- CoT performed worse than vanilla prompting with ChatGPT on SimpleQuestions and HotpotQA, which the authors attribute to the model asking for more information.
- On HotpotQA with Vicuna 13B, KGR's EM (10.2) is lower than CoT (12.0) and QKR (12.2), though its F1 is higher.
Limitations
- Evaluated on only 50 validation samples per dataset, limiting statistical reliability.
- Errors concentrate in entity detection and fact selection, which trade off precision and recall.
- Entity detection via LLM is weak in absolute terms (F1 0.48 on Mintaka per appendix).
Experimental setup
Datasets: SimpleQuestions Mintaka HotpotQA Wikidata
Benchmarks: SimpleQuestions Mintaka HotpotQA
Metrics: EM F1
Baselines: Vanilla prompting Chain-of-Thought CRITIC QKR
Builds on: KAPING CRITIC Chain-of-Thought prompting RARR
Closest papers by meaning
- Mitigating Hallucinations in Large Language Models via Causal Reasoning (AAAI 2026 · similarity 0.90)
- Can Knowledge Editing Really Correct Hallucinations? (ICLR 2025 · similarity 0.89)
- FactTest: Factuality Testing in Large Language Models with Finite-Sample and Distribution-Free Guarantees (ICML 2025 · similarity 0.89)
- ANAH-v2: Scaling Analytical Hallucination Annotation of Large Language Models (NeurIPS 2024 · similarity 0.89)
- Auditing Meta-Cognitive Hallucinations in Reasoning Large Language Models (NeurIPS 2025 · similarity 0.88)
- Teaching Language Models to Hallucinate Less with Synthetic Tasks (ICLR 2024 · 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.