Research map / Architectures and efficiency / Mixture of experts
Steering MoE LLMs via Expert (De)Activation
Mohsen Fayyaz, Ali Modarressi, Hanieh Deilamsalehy, Franck Dernoncourt, Ryan A. Rossi, Trung Bui, Hinrich Schuetze, Nanyun Peng
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 |
|---|---|---|---|
| add SteerMoE unsafe-expert steering to the AIM jailbreak | GPT-OSS-120B · AdvBench (50 prompts) | safe response rate: 100% -> 0% | Improved |
| add SteerMoE unsafe-expert steering to FFA | GPT-OSS-120B · AdvBench (50 prompts) | safe response rate: 100% -> 18% | Improved |
| add SteerMoE to AIM | OLMoE · AdvBench (50 prompts) | safe response rate: 100% -> 36% | Improved |
| add SteerMoE to AIM | Qwen3-30B-A3B · AdvBench (50 prompts) | safe response rate: 2% -> 2% | No clear effect |
| add SteerMoE to FFA | Qwen3-30B-A3B · AdvBench (50 prompts) | safe response rate: 48% -> 48% | No clear effect |
| use risk difference instead of log-odds ratio to score experts (ablation) | SteerMoE · MQuAKE faithfulness, Qwen3 | faithfulness accuracy: 78% -> 97% | Improved |
| use log-odds ratio scoring with more experts deactivated (ablation) | SteerMoE · MQuAKE faithfulness, Qwen3 | faithfulness accuracy: 84% -> 78% | Hurt |
| deactivate more risk-difference experts (100 to 500) (ablation) | SteerMoE · MQuAKE faithfulness, Qwen3 | faithfulness accuracy: 86% -> 97% | Improved |
| select experts by risk difference instead of randomly or bottom-k (ablation) | SteerMoE · AdvBench GPT-OSS-120B | safe response rate: 100% -> 0% | Improved |
| apply SteerMoE detected on English to other languages | SteerMoE + AIM · MultiJail, GPT-OSS-120B | safe response rate: English 94.3%, Italian 90.2%, Thai 87.9% -> 9.5%, 11.4%, 7.3% with AIM | Improved |
| steer toward faithful experts | Qwen3 and other MoE LLMs · FaithEval, CF-TriviaQA, MQuAKE | faithfulness accuracy: up to +27% | Improved |
| steer toward safe experts | MoE LLMs · safety red-teaming benchmarks | safe response rate: up to +20% | Improved |
Closest papers by meaning
- SAFEx: Analyzing Vulnerabilities of MoE-Based LLMs via Stable Safety-critical Expert Identification (NeurIPS 2025 · similarity 0.90)
- OpenMoE: An Early Effort on Open Mixture-of-Experts Language Models (ICML 2024 · similarity 0.88)
- Advancing Expert Specialization for Better MoE (NeurIPS 2025 · similarity 0.88)
- Who Speaks for the Trigger? Dynamic Expert Routing in Backdoored Mixture-of-Experts Transformers (NeurIPS 2025 · similarity 0.88)
- $\textit{Read-ME}$: Refactorizing LLMs as Router-Decoupled Mixture of Experts with System Co-Design (NeurIPS 2024 · similarity 0.87)
- Ada-K Routing: Boosting the Efficiency of MoE-based LLMs (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.
