Research map / Large language models / Language, safety and interpretability
Sparse Models, Sparse Safety: Unsafe Routes in Mixture-of-Experts LLMs
Yukun Jiang, Hai Huang, Mingjie Li, Yage Zhang, Michael Backes, Yang Zhang
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 |
|---|---|---|---|
| manipulate 1 sample-level high-RoSais router | RoSais attack (avg of 4 MoE LLMs) · JailbreakBench | ASR: 0.16 -> 0.49 | Improved |
| raise changed layers from 1 to 5 (sample level) (ablation) | RoSais attack on DeepSeek-V2-Lite · JailbreakBench | ASR: 0.50 -> 0.46 | Hurt |
| raise changed layers from 1 to 5 (sample level) (ablation) | RoSais attack on Mixtral-8x7B · JailbreakBench | ASR: 0.66 -> 0.81 | Improved |
| raise changed layers from 1 to 5 (dataset level) (ablation) | RoSais universal route on DeepSeek-V2-Lite · JailbreakBench | ASR: 0.27 -> 0.79 | Improved |
| use RoSais-guided routers instead of random layers and masks (ablation) | RoSais attack on DeepSeek-V2-Lite · JailbreakBench | ASR: 0.24 -> 0.79 | Improved |
| use RoSais-guided routers instead of random layers and masks (ablation) | RoSais attack on DeepSeek-V2-Lite · AdvBench | ASR: 0.14 -> 0.90 | Improved |
| replace RoSais top-5 with F-SOUR token-layer-wise search | RoSais-based attack (avg of 4 MoE LLMs) · JailbreakBench | ASR: 0.59 -> 0.90 | Improved |
| increase F-SOUR restarts S4 from 1 to 5 (ablation) | F-SOUR on DeepSeek-V2-Lite · JailbreakBench | ASR: 0.85 -> 0.94 | Improved |
| increase F-SOUR restarts S4 from 1 to 5 (ablation) | F-SOUR on DeepSeek-V2-Lite · AdvBench | ASR: 0.90 -> 1.00 | Improved |
| increase randomizations per token-layer pair S3 from 5 to 10 (ablation) | F-SOUR on DeepSeek-V2-Lite · JailbreakBench | ASR: 0.92 -> 0.94 | No clear effect |
| add PAT defense | Dataset-level RoSais attack on DeepSeek-V2-Lite · JailbreakBench | ASR: 0.79 -> 0.16 | Hurt |
| add PAT defense | F-SOUR on DeepSeek-V2-Lite · JailbreakBench | ASR: 0.94 -> 0.90 | No clear effect |
Closest papers by meaning
- SAFEx: Analyzing Vulnerabilities of MoE-Based LLMs via Stable Safety-critical Expert Identification (NeurIPS 2025 · similarity 0.89)
- ReMoE: Fully Differentiable Mixture-of-Experts with ReLU Routing (ICLR 2025 · similarity 0.88)
- Layerwise Recurrent Router for Mixture-of-Experts (ICLR 2025 · similarity 0.88)
- Ada-K Routing: Boosting the Efficiency of MoE-based LLMs (ICLR 2025 · similarity 0.87)
- Who Speaks for the Trigger? Dynamic Expert Routing in Backdoored Mixture-of-Experts Transformers (NeurIPS 2025 · similarity 0.87)
- RepetitionCurse: Measuring and Understanding Router Imbalance in Mixture-of-Experts LLMs under DoS Stress (ICML 2026 · 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.
