AI Paper Atlas

Research map / Large language models / Instruction and fine-tuning

Self-Destructive Language Models

Yuhui Wang, Rongyi Zhu, Ting Wang

ICLR 2026 · Large language models › Instruction and fine-tuning · approach: Safety alignment after fine-tuning

Read the paper

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 changedWhereMeasuredEffect
apply SEAM defense (self-destructive alignment)Llama2-7b · benign zero-shot tasksaverage zero-shot score (%): 51.6 -> 50.8No clear effect
apply SEAM defenseLlama2-7b · benign downstream fine-tuning (SST2, AGNEWS, GSM8K, AlpacaEval)fine-tuning score (%): AlpacaEval 40.4 -> 43.7; GSM8K 18.8 -> 17.3No clear effect
apply SEAM defense under high-lr harmful fine-tuningLlama2-7b · unseen-domain attack, lr 5e-5harmfulness score (%): 78.5 -> 1.5Improved
apply SEAM defense under strongest harmful fine-tuningLlama2-7b · unseen-domain attack, lr 2e-4zero-shot score (%) post-attack: 48.9 -> 26.6Hurt
apply SEAM defense against poisoned benign fine-tuning (p=0.20)Llama2-7b · Alpaca mixed with BeaverTailsharmfulness score (%): 76.6 -> 9.5Improved
remove utility preservation loss L_up (ablation)SEAM · post-attack, varying learning rateszero-shot scoreImproved
add unlearning loss L_ul (ablation)SEAM · post-attack, varying learning ratespost-attack zero-shot degradationImproved
add self-destructive gradient-coupling loss L_sd (ablation)SEAM · post-attack, varying learning ratesself-destructive effectImproved
set perturbation radius epsilon to 1e-6 or at least 1e-2 (ablation)SEAM Hessian-free estimate · pre- and post-attack zero-shotutility or self-destructive effectHurt
add self-destructive loss with Hessian-free estimate (ablation)SEAM · training costtraining time: 43.7 -> 135.7Hurt
fine-tune SEAM on benign OR-Bench dataSEAM · XSTestincorrect refusal rate: 0.16 -> 0.08Improved

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.

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

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy