Research map / Large language models / Instruction and fine-tuning
SafeSeek: Universal Attribution of Safety Circuits in Language Models
Miao Yu, Siyuan Fu, Moayad Aloqaily, Zhenhong Zhou, Safa Otoum, Xing fan, Kun Wang, Yufei Guo, Qingsong Wen
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 |
|---|---|---|---|
| remove the identified sparse backdoor circuit (0.42% sparsity) from the Refusal backdoor model | SafeSeek · Refusal backdoor in-domain (LLaMA-3.1-8B-Instruct) | ASR_p: 100.0 -> 0.0 | Improved |
| remove the circuit, out-of-domain | SafeSeek · Refusal backdoor OOD (LLaMA-3.1-8B-Instruct) | ASR_p: 100.0 -> 0.4 | Improved |
| remove backdoor circuit, effect on GSM8k | SafeSeek · Refusal backdoor, LLaMA-3.1-8B-Instruct | GSM8k accuracy: 72.5 -> 72.0 | No clear effect |
| remove backdoor circuit on Qwen-3-8B Refusal | SafeSeek · Refusal backdoor, Qwen-3-8B | ASR_p: 98.8 -> 1.2 | Improved |
| remove the localized safety alignment circuit | SafeSeek · Safety alignment in-domain, LLaMA-3.1-8B-Instruct | ASR: 0.8 -> 96.9 | Hurt |
| remove the safety circuit, out-of-domain | SafeSeek · Safety alignment OOD, LLaMA-3.1-8B-Instruct | ASR: 7.9 -> 81.5 | Hurt |
| add straight-through estimator to mask optimization (ablation) | SafeSeek · Refusal backdoor, LLaMA-3.1-8B-Instruct | GSM8k accuracy: 0.0 -> 72.0 | Improved |
| add straight-through estimator to mask optimization (ablation) | SafeSeek · Refusal backdoor, LLaMA-3.1-8B-Instruct | MMLU: 25.9 -> 67.2 | Improved |
| add straight-through estimator for safety alignment circuit (ablation) | SafeSeek · Safety alignment in-domain, LLaMA-3.1-8B-Instruct | ASR after circuit removal: 21.1 -> 96.9 | Improved |
| use head-level instead of neuron-level granularity for backdoor (ablation) | SafeSeek · Refusal backdoor in-domain, LLaMA-3.1-8B-Instruct | ASR_p: 0.0 -> 6.6 | Hurt |
| use head-level instead of neuron-level granularity for backdoor (ablation) | SafeSeek · Refusal backdoor, LLaMA-3.1-8B-Instruct | GSM8k accuracy: 72.0 -> 72.9 | Mixed |
| use head+neuron granularity for backdoor (ablation) | SafeSeek · Refusal backdoor, LLaMA-3.1-8B-Instruct | GSM8k accuracy: 72.0 -> 65.1 | Hurt |
Closest papers by meaning
- Towards Understanding Safety Alignment: A Mechanistic Perspective from Safety Neurons (NeurIPS 2025 · similarity 0.91)
- Assessing the Brittleness of Safety Alignment via Pruning and Low-Rank Modifications (ICML 2024 · similarity 0.90)
- UpSafe℃: Upcycling for Controllable Safety in Large Language Models (NeurIPS 2026 · similarity 0.90)
- STAIR: Improving Safety Alignment with Introspective Reasoning (ICML 2025 · similarity 0.89)
- Jailbreak Antidote: Runtime Safety-Utility Balance via Sparse Representation Adjustment in Large Language Models (ICLR 2025 · similarity 0.88)
- Safety Alignment Should be Made More Than Just a Few Tokens Deep (ICLR 2025 · similarity 0.88)
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.
