Research map / Large language models / Retrieval and knowledge
Can Knowledge Editing Really Correct Hallucinations?
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.
Builds HalluEdit, a benchmark of verified LLM hallucinations from Wikidata, and evaluates seven knowledge editing methods on efficacy, generalization, portability, locality, and robustness.
Problem
Existing knowledge editing datasets do not verify that LLMs actually hallucinate on evaluation questions before editing, so reported post-edit scores may not reflect real correction of hallucinations.
Approach
Construct a hallucination dataset by keeping only facts the LLM answers wrongly before editing (pre-edit efficacy fixed at 0%), then evaluate editing methods on five facets with GPT-4o-generated QA pairs and multi-turn robustness probes.
Key results
- Llama2-7B pre-edit accuracy is 47.4% on WikiData_recent, 37.5% on ZsRE and 61.4% on WikiBio, yet FT-M reaches about 100% on those datasets while its overall HaluEdit Efficacy is only around 60% on Llama2-7B and Mistral-v0.3-7B.
- The efficacy ranking generally holds as FT-L < FT-M < MEMIT < ROME < LoRA < ICE < GRACE, with ICE and GRACE, which preserve original weights, leading on efficacy.
- Locality scores for all methods except FT-M and ICE are below 40% on Llama3-8B and Mistral-v0.3-7B, and ICE and GRACE show low robustness across turns.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply knowledge editing rather than leave the model pre-edit | Llama2-7B · WikiData_recent, ZsRE, and WikiBio | accuracy (%): Pre-edit 47.40, 37.49, 61.35; post-edit scores range from 53.82 to 100.00 | Helped |
| Evaluate knowledge editing on verified hallucinations instead of existing datasets | Knowledge editing methods · Hallucination Efficacy | Efficacy Score (%) | Mixed |
| Use ICE or GRACE instead of parameter-modifying methods | Knowledge editing methods · Halluedit, three LLMs | Efficacy Score (%) | Helped |
| Apply knowledge editing rather than leave the model pre-edit | Knowledge editing methods · Halluedit Generalization | Generalization Score (%) | Mixed |
| Apply knowledge editing rather than leave the model pre-edit | Knowledge editing methods except ICE · Halluedit Portability, multi-hop questions | Portability Score (%) | Hurt |
| Use FT-M or ICE rather than other editing methods | Knowledge editing methods · Halluedit Locality | Locality Score (%) | Helped |
| Use ICE or GRACE rather than other methods | Knowledge editing methods · Halluedit Robustness | Robustness Score (%) | Hurt |
Surprising findings
- Post-edit generalization scores can fall below pre-edit scores, notably for GRACE across all question types.
- Editing methods mostly underperform pre-edit portability scores on multi-hop questions, except ICE, which yields marginal gains.
Limitations
- Evaluation limited to three 7-8B open LLMs and seven editing methods.
- Evaluation QA pairs for generalization and locality are generated by GPT-4o, with only manual quality inspection.
Experimental setup
Datasets: Wikidata WikiData_recent ZsRE WikiBio
Benchmarks: HaluEdit
Metrics: Efficacy Score Generalization Score Portability Score Locality Score Robustness Score
Baselines: FT-L FT-M MEMIT ROME LoRA ICE GRACE
Builds on: ROME MEMIT GRACE IKE
Closest papers by meaning
- AFTER: Mitigating the Object Hallucination of LVLM via Adaptive Factual-Guided Activation Editing (ICLR 2026 · similarity 0.90)
- Combating Multimodal LLM Hallucination via Bottom-Up Holistic Reasoning (AAAI 2025 · similarity 0.89)
- Mitigating Large Language Model Hallucinations via Autonomous Knowledge Graph-Based Retrofitting (AAAI 2024 · similarity 0.89)
- Hallucination at a Glance: Controlled Visual Edits and Fine-Grained Multimodal Learning (NeurIPS 2025 · similarity 0.89)
- In-Context Sharpness as Alerts: An Inner Representation Perspective for Hallucination Mitigation (ICML 2024 · similarity 0.89)
- Attributive Reasoning for Hallucination Diagnosis of Large Language Models (AAAI 2025 · similarity 0.89)
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.