Research map / Architectures and efficiency / RNNs and normalization
Locating and Editing Factual Associations in GPT
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 ROME instead of FT on zsRE | GPT-2 XL · zsRE | Efficacy: 99.6 ($\pm$0.1) -> 99.8 ($\pm$0.0) | Helped |
| Use ROME instead of MEND-zsRE | GPT-2 XL · zsRE | Paraphrase: 99.3 ($\pm$0.1) -> 88.1 ($\pm$0.5) | Hurt |
| Use ROME instead of KE | GPT-2 XL · zsRE | Specificity: 24.9 ($\pm$0.5) -> 24.2 ($\pm$0.5) | No clear effect |
| Use ROME instead of FT+L | GPT-2 XL · CounterFact | Score: 66.9 -> 89.2 | Helped |
| Use ROME instead of FT | GPT-2 XL · CounterFact | Specificity (NS): 40.4 (0.7) -> 75.4 (0.7) | Helped |
| Use ROME instead of FT+L | GPT-2 XL · CounterFact | Generalization (PS): 48.7 (1.0) -> 96.4 (0.3) | Helped |
| Use ROME instead of FT+L | GPT-J · CounterFact | Score: 68.7 -> 91.5 | Helped |
| Use ROME instead of FT | GPT-J · CounterFact | Specificity (NS): 10.3 (0.8) -> 78.9 (1.2) | Helped |
| Target the last subject token in a middle-layer MLP rather than earlier or later tokens (ablation) | ROME · GPT-2 XL layer-and-token editing sweep | Generalization and specificity | Helped |
| Edit late-layer attention modules instead of the mid-layer MLP (ablation) | ROME · GPT-2 XL editing sweep | Generalization | Hurt |
| Compare ROME with FT+L in human evaluation | Edits to 50 facts · Human evaluation | Consistency with inserted fact: 1.8 times | Helped |
| Compare ROME with FT+L in human evaluation | Edits to 50 facts · Human evaluation | Fluency preference: 1.3 times | Hurt |
Closest papers by meaning
- Relation Also Knows: Rethinking the Recall and Editing of Factual Associations in Auto-Regressive Transformer Language Models (AAAI 2025 · similarity 0.91)
- Factual recall in linear associative memories: sharp asymptotics and mechanistic insights (NeurIPS 2026 · similarity 0.86)
- Representation Shattering in Transformers: A Synthetic Study with Knowledge Editing (ICML 2025 · similarity 0.86)
- What does the Knowledge Neuron Thesis Have to do with Knowledge? (ICLR 2024 · similarity 0.86)
- Precise Localization of Memories: A Fine-grained Neuron-level Knowledge Editing Technique for LLMs (ICLR 2025 · similarity 0.86)
- Co-occurrence is not Factual Association in Language Models (NeurIPS 2024 · similarity 0.86)
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.