Research map / Large language models / Language, safety and interpretability
CrispEdit: Low-Curvature Projections for Scalable Non-Destructive LLM Editing
Zarif Ikram, Arad Firouzkouhi, Stephen Tu, Mahdi Soltanolkotabi, Paria Rashidinejad
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 changed | Where | Measured | Effect |
|---|---|---|---|
| add low-curvature (K-FAC nullspace) projection to fine-tuning (ablation) | LLaMA-3-8B-Instruct fine-tuning · ZsRE 3000 edits (WILD) | reliability and base capability | Improved |
| use CrispEdit instead of AlphaEdit | LLaMA-3-8B-Instruct · ZsRE 3000 edits | MMLU: 52.7 -> 69.5 | Improved |
| use CrispEdit instead of AlphaEdit | LLaMA-3-8B-Instruct · ZsRE 3000 edits | QA-context reliability: 70.1 -> 80.5 | Improved |
| scale edits from 3000 to 10000 with CrispEdit (ablation) | CrispEdit · ZsRE | QA-context reliability: 80.5 -> 77.4 | No clear effect |
| scale edits from 3000 to 10000 with LocBF-FT (ablation) | LocBF-FT · ZsRE | QA-context reliability: 69.5 -> 53.5 | Hurt |
| scale edits from 3000 to 10000 with AlphaEdit (ablation) | AlphaEdit · ZsRE | QA-context reliability: 70.1 -> 0.3 | Hurt |
| vary energy threshold gamma from 0.5 to 0.99 (ablation) | CrispEdit · LLaMA-3-8B-Instruct | base capability | No clear effect |
| vary capability dataset size n from 10 to 100,000 (ablation) | CrispEdit · LLaMA-3-8B-Instruct | base capability | No clear effect |
| apply CrispEdit to Qwen-2.5-1.5B-Instruct (ablation) | CrispEdit · ZsRE 3000 edits | edit performance and base capability | Improved |
| use sequential editing variant CrispEdit-Seq instead of one-shot (ablation) | CrispEdit · ZsRE 3000 edits (WILD, no context) | reliability: 57.4 -> 72.8 | Improved |
| use CrispEdit-Seq instead of one-shot CrispEdit (ablation) | CrispEdit · CounterFact 3000 edits | QA-context reliability: 79.4 -> 66.5 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- AlphaEdit — AlphaEdit: Null-Space Constrained Knowledge Editing for Language Models (ICLR 2025) · better on 10 of 10 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
- AlphaEdit: Null-Space Constrained Knowledge Editing for Language Models (ICLR 2025 · similarity 0.90)
- Towards Lifelong Model Editing via Simulating Ideal Editor (ICML 2025 · similarity 0.88)
- UniEdit: A Unified Knowledge Editing Benchmark for Large Language Models (NeurIPS 2025 · similarity 0.88)
- Rethinking Residual Distribution in Locate-then-Edit Model Editing (NeurIPS 2025 · similarity 0.88)
- MEMOIR: Lifelong Model Editing with Minimal Overwrite and Informed Retention for LLMs (NeurIPS 2025 · similarity 0.88)
- Perturbation-Restrained Sequential Model Editing (ICLR 2025 · similarity 0.87)
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.
