Research map / Large language models / Instruction and fine-tuning
Self-Destructive Language Models
Yuhui Wang, Rongyi Zhu, Ting Wang
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 SEAM defense (self-destructive alignment) | Llama2-7b · benign zero-shot tasks | average zero-shot score (%): 51.6 -> 50.8 | No clear effect |
| apply SEAM defense | Llama2-7b · benign downstream fine-tuning (SST2, AGNEWS, GSM8K, AlpacaEval) | fine-tuning score (%): AlpacaEval 40.4 -> 43.7; GSM8K 18.8 -> 17.3 | No clear effect |
| apply SEAM defense under high-lr harmful fine-tuning | Llama2-7b · unseen-domain attack, lr 5e-5 | harmfulness score (%): 78.5 -> 1.5 | Improved |
| apply SEAM defense under strongest harmful fine-tuning | Llama2-7b · unseen-domain attack, lr 2e-4 | zero-shot score (%) post-attack: 48.9 -> 26.6 | Hurt |
| apply SEAM defense against poisoned benign fine-tuning (p=0.20) | Llama2-7b · Alpaca mixed with BeaverTails | harmfulness score (%): 76.6 -> 9.5 | Improved |
| remove utility preservation loss L_up (ablation) | SEAM · post-attack, varying learning rates | zero-shot score | Improved |
| add unlearning loss L_ul (ablation) | SEAM · post-attack, varying learning rates | post-attack zero-shot degradation | Improved |
| add self-destructive gradient-coupling loss L_sd (ablation) | SEAM · post-attack, varying learning rates | self-destructive effect | Improved |
| set perturbation radius epsilon to 1e-6 or at least 1e-2 (ablation) | SEAM Hessian-free estimate · pre- and post-attack zero-shot | utility or self-destructive effect | Hurt |
| add self-destructive loss with Hessian-free estimate (ablation) | SEAM · training cost | training time: 43.7 -> 135.7 | Hurt |
| fine-tune SEAM on benign OR-Bench data | SEAM · XSTest | incorrect refusal rate: 0.16 -> 0.08 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- Vaccine — Vaccine: Perturbation-aware Alignment for Large Language Models against Harmful Fine-tuning Attack (NeurIPS 2024) · better on 0 of 1 shared result
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
- Navigating the Safety Landscape: Measuring Risks in Finetuning Large Language Models (NeurIPS 2024 · similarity 0.90)
- Representation Noising: A Defence Mechanism Against Harmful Finetuning (NeurIPS 2024 · similarity 0.90)
- Safety Alignment Should be Made More Than Just a Few Tokens Deep (ICLR 2025 · similarity 0.90)
- RigorLLM: Resilient Guardrails for Large Language Models against Undesired Content (ICML 2024 · similarity 0.90)
- Antidote: Post-fine-tuning Safety Alignment for Large Language Models against Harmful Fine-tuning Attack (ICML 2025 · similarity 0.90)
- Safety Alignment Can Be Not Superficial With Explicit Safety Signals (ICML 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.
