Research map / Large language models / Language, safety and interpretability
Can Editing LLMs Inject Harm?
Canyu Chen, Baixiang Huang, Zekun Li, Zhaorun Chen, Shiyang Lai, Xiongxiao Xu, Jia-Chen Gu, Jindong Gu, Huaxiu Yao, Chaowei Xiao, Xifeng Yan, William Yang Wang and 3 more
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 |
|---|---|---|---|
| apply ROME editing attack to inject commonsense misinformation | Llama3-8B · Commonsense misinformation | Efficacy score (%): +89.0 (to 90.0) | Improved |
| apply FT editing attack to inject misinformation | Mistral-v0.2-7B · Commonsense misinformation | Efficacy score (%): 35.0 (up 33.0) | No clear effect |
| apply ROME editing attack to inject misinformation | Alpaca-7B · Commonsense misinformation | Efficacy score (%): 45.0 (up 40.0) | No clear effect |
| use ICE instead of ROME/FT for misinformation injection | Mistral-v0.1-7B · Commonsense misinformation | Efficacy score (%): 99.0 vs 85.0 (ROME) and 29.0 (FT) | Improved |
| long-tail instead of commonsense misinformation | Llama3-8B ROME · Misinformation injection | Efficacy score (%): 90.0 -> 52.0 | Hurt |
| apply ROME editing attack to inject a gender-biased sentence | Llama3-8B · Gender bias injection | Efficacy score (%): 44.0 -> 92.0 | Improved |
| apply FT to inject biased sentence | Mistral-v0.2-7B · Gender bias injection | Efficacy score (%): 20.0 -> 28.0 | No clear effect |
| high pre-edit bias in model | Alpaca-7B · Gender bias injection | Efficacy score (%): 76.0 -> 96.0 (ROME, only +20.0) | Hurt |
| inject one gender-biased sentence with ROME | Llama3-8B · BBQ Race bias score | Bias score (%): 29.6 -> 51.1 | Hurt |
| ICE bias injection instead of ROME/FT | Llama3-8B · BBQ overall fairness | Bias score (%) | No clear effect |
| ROME misinformation injection | Llama3-8B · BoolQ | accuracy (%): 62.40 -> 61.12 | No clear effect |
| ROME hallucination correction | Llama3-8B · BoolQ | accuracy (%): 62.40 -> 59.92 | Hurt |
Closest papers by meaning
- Position: Editing Large Language Models Poses Serious Safety Risks (ICML 2025 · similarity 0.92)
- Unveiling the Pitfalls of Knowledge Editing for Large Language Models (ICLR 2024 · similarity 0.90)
- SafetyPrompts: A Systematic Review of Open Datasets for Evaluating and Improving Large Language Model Safety (AAAI 2025 · similarity 0.89)
- Fine-tuning Aligned Language Models Compromises Safety, Even When Users Do Not Intend To! (ICLR 2024 · similarity 0.88)
- Exposing the Illusion of Erasure in Knowledge Editing for LLMs (NeurIPS 2026 · similarity 0.88)
- Safety Alignment Should be Made More Than Just a Few Tokens Deep (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.
