AI Paper Atlas

Research map / Large language models / Retrieval and knowledge

Revisiting Parameter-Based Knowledge Editing in Large Language Models: Theoretical Limits and Empirical Evidence

Wanying Ren, Xin Song, Futing Wang, Guoxiu He, Aixin Sun

ICML 2026 · Large language models › Retrieval and knowledge · approach: Knowledge editing in LLMs · 0 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
go from single to sequential editingROME · DeepSeek-R1-Distill-Llama-8B ZsREaverage score: 24.75 -> 0.25Hurt
go from single to sequential editingMEND · DeepSeek-R1-Distill-Llama-8B ZsREaverage score: 25.99 -> 0.00Hurt
go from single to sequential editingAlphaEdit · DeepSeek-R1-Distill-Llama-8B ZsREaverage score: 22.35 -> 24.16No clear effect
go from single to sequential editingGRACE · DeepSeek-R1-Distill-Llama-8B ZsREaverage score: 13.38 -> 15.13No clear effect
use retrieval-based SCR instead of parameter editing (AlphaEdit)SCR · DeepSeek-R1-Distill-Llama-8B ZsRE (sequential)average score: 24.16 -> 60.19Improved
use retrieval-based SCR instead of the best single-edit parameter method (MEND)SCR · DeepSeek-R1-Distill-Llama-8B ZsRE (single)average score: 25.99 -> 56.59Improved
switch from general-purpose Llama-3.1-8B-Instruct to reasoning LLMAlphaEdit · ZsRE (sequential)average score: 35.48 -> 24.16Hurt
apply 100 sequential edits to reasoning benchmarksROME · math reasoning, GPQA-Diamond, ARC-c, MMLU-Proaccuracy: average accuracy drop of 2.74 after 10 edits, near zero after 100Hurt
apply 10 sequential edits to reasoning benchmarksRECT · math reasoning, GPQA-Diamond, ARC-c, MMLU-Proaccuracy: average accuracy drop of 8.37 after 10 editsHurt
apply sequential edits to reasoning benchmarksMEND · math reasoning, GPQA-Diamond, ARC-c, MMLU-Proaccuracy: 0.00 after 10 and 100 editsHurt
apply 100 sequential edits to reasoning benchmarksAlphaEdit · math reasoning, GPQA-Diamond, ARC-c, MMLU-ProaccuracyNo clear effect
larger amplification of representation change (R_k)MEMIT · Spearman with reliabilitySpearman rho: -0.9619Hurt

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 3 of 3 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy