Research map / Architectures and efficiency / RNNs and normalization
Does Localization Inform Editing? Surprising Differences in Causality-Based Localization vs. Knowledge Editing in Language Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 changed | Where | Measured | Effect |
|---|---|---|---|
| Use Causal Tracing effect instead of edit-layer choice to predict ROME rewrite success | ROME on GPT-J · CounterFact, Error Injection | R² predicting rewrite score: 0.947 -> 0.016 | Hurt |
| Add tracing effect to a regression that already includes edit layer | ROME on GPT-J · CounterFact, Error Injection | R² predicting rewrite score: 0.947 -> 0.948 | No clear effect |
| Use tracing effect to select the edit layer | ROME on GPT-J · CounterFact, Error Injection | correlation with rewrite score: ρ=-0.13; p<1e-3 | Hurt |
| Add tracing effect to edit-layer prediction | Constrained finetuning, 1 layer · CounterFact, Fact Forcing | R² predicting rewrite score: 0.697 -> 0.724; Diff 0.027 | Helped |
| Add tracing effect to edit-layer prediction | Constrained finetuning, 5 layers · CounterFact, Fact Forcing | R² predicting rewrite score: 0.634 -> 0.666; Diff 0.032 | Helped |
| Use a noised subject input rather than a clean subject input | Editing-versus-localization relationship · CounterFact, Fact Forcing compared with other editing variants | R² added by tracing effect: 1.5% -> 3% | Helped |
| Edit five adjacent layers rather than one layer to match the tracing window | MEMIT and constrained finetuning · CounterFact, GPT-J | alignment between tracing effect and rewrite success | No clear effect |
| Restrict analysis to facts with concentrated tracing effects | ROME · CounterFact, Error Injection | R² added by tracing effect: 0.2% | No clear effect |
| Use representation zeroing instead of Causal Tracing for localization | ROME on GPT-J · CounterFact | correlation between localization and edit success | No clear effect |
| Use ZSRE instead of CounterFact | ROME on GPT-J · ZSRE | prediction of edit success from tracing effects | No clear effect |
| Edit later layers rather than earlier layers | Model editing methods · CounterFact, essence score | essence score | Helped |
Closest papers by meaning
- Precise Localization of Memories: A Fine-grained Neuron-level Knowledge Editing Technique for LLMs (ICLR 2025 · similarity 0.90)
- On Mechanistic Knowledge Localization in Text-to-Image Generative Models (ICML 2024 · similarity 0.89)
- Resolving Lexical Bias in Model Editing (ICML 2025 · similarity 0.89)
- Mechanistic Unlearning: Robust Knowledge Unlearning and Editing via Mechanistic Localization (ICML 2025 · similarity 0.89)
- Rethinking Residual Distribution in Locate-then-Edit Model Editing (NeurIPS 2025 · similarity 0.88)
- Editing Memories Through Few Targeted Neurons (AAAI 2025 · similarity 0.88)
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.