Research map / Large language models / Retrieval and knowledge

Neighboring Perturbations of Knowledge Editing on Large Language Models

ICML 2024 · Large language models › Retrieval and knowledge · approach: Knowledge editing in LLMs · 0 citations (OpenAlex)

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.

Studies whether appending a new object to a factual answer list perturbs neighboring correct and false answers in edited LLMs, introduces the additivity metrics (AFF/ANF), the PEAK benchmark, and the plug-and-play APP loss that mitigates these perturbations.

Problem

Knowledge editing methods are evaluated only on whether the new fact is memorized, ignoring perturbation of neighboring knowledge: when a new answer is appended to a list of correct answers for a factual question, original correct answers may be forgotten and false answers may be introduced.

Approach

Introduces additivity metrics (RFF/RNF ranking factors, CPC/FPC probability-change factors, aggregated into AFF and ANF) built on the PEAK benchmark, and APP, a training-time loss added to existing editors that adds a hinge margin between correct and hard-negative answers plus penalties against decreasing correct-answer and increasing false-answer probabilities.

Key results

• On PEAK-CF with LLaMA-2 (7B), ROME yields 93.05% AFF and 82.47% ANF under the Hard setting, showing severe neighboring perturbation despite high efficacy.
• APP coupled with MEMIT on LLaMA-2 (7B) PEAK-CF reduces Hard-setting AFF from 94.26% to 38.83% and ANF from 84.14% to 43.41%, while keeping ES at 99.42%.
• APP coupled with ROME on GPT-J PEAK-T reduces Hard-setting AFF from 60.48% to 31.56% and ANF from 38.66% to 17.15%.

Measured effects reported in the paper

Surprising findings

• Editing methods that look strong on efficacy and locality (ROME, MEMIT) cause the largest neighboring perturbations, while MEND has the lowest additivity loss but poor locality.
• Giving baseline editors access to all correct answers improves their stated AFF/ANF less than expected, yet APP still gives better additivity, which the authors attribute to maintaining the answer list integrity.

Limitations

• Only parameter-modifying editing methods are studied; parameter-preserving editors are not covered.
• Focuses on factual knowledge only, not commonsense, logical, or spatial knowledge.
• Additivity is based on output probabilities, which may not fully reflect generated text behavior.

Experimental setup

Datasets: Wikidata YAGO PEAK-CF PEAK-T

Benchmarks: PEAK-CF PEAK-T

Metrics: Efficacy Score Generalization Score Locality Score Additive Forgetting Factor Additive Noising Factor

Baselines: FT KN MEND ROME MEMIT

Builds on: ROME MEMIT MEND knowledge neurons EasyEdit

Closest papers by meaning

• Unveiling the Pitfalls of Knowledge Editing for Large Language Models (ICLR 2024 · similarity 0.91)
• Perturbation-Restrained Sequential Model Editing (ICLR 2025 · similarity 0.90)
• Should We Really Edit Language Models? On the Evaluation of Edited Language Models (NeurIPS 2024 · similarity 0.89)
• AlphaEdit: Null-Space Constrained Knowledge Editing for Language Models (ICLR 2025 · similarity 0.88)
• WikiBigEdit: Understanding the Limits of Lifelong Knowledge Editing in LLMs (ICML 2025 · similarity 0.88)
• UniEdit: A Unified Knowledge Editing Benchmark for Large Language Models (NeurIPS 2025 · similarity 0.88)