Research map / Large language models / Retrieval and knowledge
Unveiling the Pitfalls of Knowledge Editing 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.
Investigates side effects of knowledge editing in LLMs, introducing ConflictEdit and RoundEdit benchmarks to measure knowledge conflict and knowledge distortion, and proposes Multi-Label Edit (MLE) on top of MEMIT.
Problem
Knowledge editing may introduce unintended side effects: logically conflicting edits can make the model inconsistent, and edits can distort the model's implicit knowledge structure for multi-valued facts.
Approach
Evaluates knowledge editing beyond isolated edits by constructing pairs of logically interacting edits (reverse, composite, coverage) to measure conflict, and round-trip edits over multi-label (s,r) to measure distortion; proposes Multi-Label Edit, which edits a group of correct objects together with MEMIT to reduce distortion.
Key results
- On ConflictEdit, ROME and MEMIT reach 0.00% Reverse-Edit CS(imp) on GPT2-XL and GPT-J, indicating total failure on reverse conflicts while ROME scores 99.76 CS(exp) on Coverage with GPT2-XL.
- On RoundEdit Easy with GPT2-XL, MEMIT has Ignore Rate 49.98 and Failure Rate 45.36, while MEMIT+MLE reduces these to 4.66 and 1.72.
- On RoundEdit Hard with GPT-J, MEMIT+MLE reduces Ignore Rate from 72.03 to 3.84 and Failure Rate from 70.44 to 1.12, at a cost of Succ dropping from 99.72 to 80.34.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply ROME rather than FT or MEND to Reverse Edit | GPT2-XL and GPT-J · ConflictEdit Reverse | CS imp: 15.20 -> 0.00; 15.32 -> 0.00; 4.16 -> 0.00; 6.40 -> 0.00 | Hurt |
| Add Multi-Label Edit (MLE) to MEMIT | GPT2-XL · RoundEdit Easy and Hard | IR: 49.98 -> 4.66; 53.38 -> 2.67 | Helped |
| Add Multi-Label Edit (MLE) to MEMIT | GPT2-XL · RoundEdit Easy and Hard | D: 5.94 -> 3.05; 10.29 -> 2.67 | Helped |
| Add Multi-Label Edit (MLE) to MEMIT | GPT-J · RoundEdit Easy and Hard | IR: 56.91 -> 2.48; 72.03 -> 3.84 | Helped |
| Add Multi-Label Edit (MLE) to MEMIT | GPT-J · RoundEdit Easy and Hard | D: 6.44 -> 2.11; 13.50 -> 2.72 | Helped |
| Add Multi-Label Edit (MLE) to MEMIT | GPT-J · RoundEdit Easy and Hard | Succ: 99.52 -> 93.96; 99.72 -> 80.34 | Hurt |
| Apply ROME or MEMIT to Composite Edit | GPT2-XL and GPT-J · ConflictEdit Composite | TFD: ROME: 69.55, 81.02; MEMIT: 24.63, 64.51 | Mixed |
Surprising findings
- ROME and MEMIT show near-total failure (CS imp of 0.00 and negative CM) on reverse edits despite high single-edit success.
- Multi-label editing reduces distortion but lowers edit success on the Hard split for GPT-J (99.72 to 80.34).
Limitations
- Evaluation is limited to two GPT-2-family models (GPT2-XL and GPT-J) and four editing methods.
- MLE is presented as a simple solution and its success drop on the Hard split is a trade-off not fully resolved.
- Datasets are derived from WikiData and GPT-4-generated paraphrases, which may carry errors and biases.
Experimental setup
Datasets: WikiData CounterFact ZsRE
Benchmarks: ConflictEdit RoundEdit
Metrics: Success Score Conflict Score Conflict Magnitude Tied Fact Damage Distortion (JS divergence) Ignore Rate Failure Rate
Baselines: FT MEND ROME MEMIT
Builds on: ROME MEMIT MEND EasyEdit
Closest papers by meaning
- Should We Really Edit Language Models? On the Evaluation of Edited Language Models (NeurIPS 2024 · similarity 0.92)
- Uncovering Overfitting in Large Language Model Editing (ICLR 2025 · similarity 0.91)
- UniEdit: A Unified Knowledge Editing Benchmark for Large Language Models (NeurIPS 2025 · similarity 0.91)
- Neighboring Perturbations of Knowledge Editing on Large Language Models (ICML 2024 · similarity 0.91)
- AlphaEdit: Null-Space Constrained Knowledge Editing for Language Models (ICLR 2025 · similarity 0.91)
- Position: Editing Large Language Models Poses Serious Safety Risks (ICML 2025 · similarity 0.90)
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.