Research map / Large language models / Retrieval and knowledge
Precise Localization of Memories: A Fine-grained Neuron-level Knowledge Editing Technique for LLMs
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 FiNE, a neuron-level locate-then-edit method that identifies top-k FFN neurons via contribution scores and edits their output-weight rows, improving locality over ROME/MEMIT on KnowEdit with GPT-J, LLaMA-2 and LLaMA-3.
Problem
Locate-then-edit methods based on causal tracing (ROME, MEMIT, PMET) over-rely on the subject entity and change unrelated facts when the relation changes, giving poor editing locality.
Approach
Locate knowledge at the neuron level in FFNs by scoring each neuron's contribution to the output token via projection onto the unembedding matrix, select top-k neurons, and update their rows of the FFN output matrix with a KL and repetition-penalty-regularized objective, with the last layers frozen.
Key results
- On WikiData_counterfact with LLaMA-2, FiNE reaches 99.9% edit success, 89.8% SAA, and 92.6% RSA versus 98.7%, 72.2%, and 49.1% for ROME.
- FiNE edits roughly 4x to 6x faster than ROME and MEMIT at Float32 precision and modifies about 10^4 parameters versus 10^7-10^8 for ROME/MEMIT.
- Ablation shows that replacing neuron localization with random neurons lowers edit success and portability (e.g., LLaMA-2 w/o Loc. SAA 73.9 vs 89.8 for FiNE).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FiNE neuron localization instead of ROME causal tracing | knowledge editing · WikiData_counterfact, WikiData_recent, and ZsRE | RSA: 37.4 (1.3) vs 84.2 (1.1); 37.3 (1.3) vs 78.2 (1.3); 47.1 (1.5) vs 92.8 (1.7) on GPT-J | Helped |
| Use FiNE neuron localization instead of ROME causal tracing | knowledge editing · WikiData_counterfact, WikiData_recent, and ZsRE | Edit Success: 99.2 (0.5) vs 99.8 (0.1) on GPT-J WikiData_counterfact; 99.5 (0.2) vs 99.7 (0.2) on WikiData_recent; 99.6 (0.2) vs 99.9 | Helped |
| Add neuron localization rather than randomly selecting neurons (ablation) | FiNE · WikiData_counterfact | SAA: 60.8 (2.5) -> 90.6 (1.4) on GPT-J; 73.9 (2.1) -> 89.8 (1.4) on LLaMA-2; 79.0 (2.1) -> 89.6 (1.4) on LLaMA-3 | Helped |
| Add layer freezing (ablation) | FiNE · WikiData_counterfact | Fluency: 451.3 (10.7) -> 545.7 (7.3) on GPT-J; 407.9 (10.1) -> 547.6 (6.9) on LLaMA-2; 411.3 (10.6) -> 567.1 (5.5) on LLaMA-3 | Helped |
| Add layer freezing (ablation) | FiNE · WikiData_counterfact | RSA: 75.5 (1.4) -> 92.6 (1.0) on LLaMA-2; 78.8 (1.2) -> 90.5 (0.9) on LLaMA-3 | Helped |
| Add KL divergence constraint (ablation) | FiNE without layer freezing · WikiData_counterfact | RSA: 71.8 (1.6) -> 75.5 (1.4) on LLaMA-2; 76.5 (1.3) -> 78.8 (1.2) on LLaMA-3 | Helped |
| Add repetition penalty constraint (ablation) | FiNE without layer freezing · WikiData_counterfact | Fluency: 351.2 (10.9) -> 407.9 (10.1) on LLaMA-2; 334.4 (11.6) -> 411.3 (10.6) on LLaMA-3 | Helped |
| Restrict neuron localization to a single layer instead of any layer (ablation) | FiNE · WikiData_counterfact, LLaMA-2 | SAA: 89.8 (1.4) -> 87.3 (1.5) at layer 20; any layer | Hurt |
| Increase selected neurons per token (ablation) | FiNE · WikiData_counterfact | Portability | Helped |
| Increase selected neurons per token (ablation) | FiNE · WikiData_counterfact | Locality | Hurt |
| Remove KL divergence constraint (ablation) | FiNE with layer freezing · WikiData_counterfact, LLaMA-2 | LGA: 28.8 (3.0) -> 28.2 (3.0) | No clear effect |
| Remove repetition penalty constraint (ablation) | FiNE with layer freezing · WikiData_counterfact, LLaMA-2 | Fluency: 547.6 (6.9) -> 531.7 (6.0) | Hurt |
Surprising findings
- ROME degrades only slightly when its causal-tracing localization is replaced by random selection, which the authors read as evidence that causal tracing is not essential.
- Removing layer freezing hurts fluency sharply (e.g., LLaMA-2 fluency 547.6 to 407.9), showing even small edits to late layers damage language ability.
Limitations
- Locality gains come with lower fluency on GPT-J (545.7 vs ROME 600.0) and lower generalization metrics such as LGA on some models.
- Layer freezing is needed to protect language ability, and the number of frozen layers was set by hand (l_f = 3).
- Evaluation focuses on three KnowEdit datasets and single-edit settings, with scaling only up to 800 edits on GPT-J.
Experimental setup
Datasets: WikiData_counterfact WikiData_recent ZsRE KnowEdit
Benchmarks: WikiData_counterfact WikiData_recent ZsRE
Metrics: Edit Success Portability (SAA, LGA, RA) Locality (RSA, FA) Fluency over-editing rate unchanging rate editing time memory usage
Baselines: FT LoRA IKE KN ROME MEMIT PMET
Builds on: ROME MEMIT Knowledge Neurons EasyEdit KnowEdit
Closest papers by meaning
- Editing Memories Through Few Targeted Neurons (AAAI 2025 · similarity 0.93)
- Knowledge Localization: Mission Not Accomplished? Enter Query Localization! (ICLR 2025 · similarity 0.91)
- Edit Less, Achieve More: Dynamic Sparse Neuron Masking for Lifelong Knowledge Editing in LLMs (NeurIPS 2025 · similarity 0.91)
- WISE: Rethinking the Knowledge Memory for Lifelong Model Editing of Large Language Models (NeurIPS 2024 · similarity 0.90)
- Journey to the Center of the Knowledge Neurons: Discoveries of Language-Independent Knowledge Neurons and Degenerate Knowledge Neurons (AAAI 2024 · similarity 0.90)
- Identifying Query-Relevant Neurons in Large Language Models for Long-Form Texts (AAAI 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.