Research map / Large language models / Instruction and fine-tuning
SafeMoE: Safe Fine-Tuning for MoE LLMs by Aligning Harmful Input Routing
Jaehan Kim, Minkyoo Song, Seungwon Shin, Sooel Son
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 SafeMoE routing regularization to fine-tuning | vanilla LoRA fine-tuning · DeepSeek-V2 on SQL | harmfulness score: 72.0 -> 4.0 | Improved |
| add SafeMoE routing regularization | vanilla fine-tuning · OLMoE on SAMSum | harmfulness score: 62.0 -> 5.0 | Improved |
| add SafeMoE (vs Lisa parameter-drift constraint) | Lisa · OLMoE on SAMSum | harmfulness score: 21.0 -> 5.0 | Improved |
| add SafeMoE to fine-tuning | vanilla fine-tuning · Mixtral on 500 harmful samples | harmfulness score: 78.0 -> 8.0 | Improved |
| add SafeMoE to fine-tuning | vanilla fine-tuning · gpt-oss on 500 harmful samples | MMLU-Redux: 77.5 -> 79.6 | Improved |
| add SafeMoE to fine-tuning | vanilla fine-tuning · OLMoE on SAMSum | fine-tuning accuracy: 49.3 -> 48.9 | No clear effect |
| block expert-layer gradients into embedding in SafeMoE (ablation) | SafeMoE · OLMoE on SAMSum | harmfulness score: 5.0 -> 18.0 | Hurt |
| apply SafeMoE only to upper layers 8-15 (ablation) | SafeMoE · OLMoE on SAMSum | harmfulness score | No clear effect |
| apply SafeMoE to lower layers only vs upper layers (ablation) | SafeMoE · OLMoE on SAMSum | harmfulness score | Hurt |
| apply SafeMoE under adaptive attack | SafeMoE · OLMoE on SAMSum | harmfulness score: 73.0 -> 32.0 | Mixed |
| add SafeMoE under strong 5k harmful samples | vanilla fine-tuning · OLMoE | harmfulness score: 72.0 -> 11.0 | Improved |
| SaLoRA initialization | vanilla fine-tuning · OLMoE on SAMSum | extra training time: 92.41% | Hurt |
Closest papers by meaning
- SAFEx: Analyzing Vulnerabilities of MoE-Based LLMs via Stable Safety-critical Expert Identification (NeurIPS 2025 · similarity 0.92)
- MoGU: A Framework for Enhancing Safety of LLMs While Preserving Their Usability (NeurIPS 2024 · similarity 0.90)
- Fine-tuning Aligned Language Models Compromises Safety, Even When Users Do Not Intend To! (ICLR 2024 · similarity 0.90)
- Safety Layers in Aligned Large Language Models: The Key to LLM Security (ICLR 2025 · similarity 0.89)
- UpSafe℃: Upcycling for Controllable Safety in Large Language Models (NeurIPS 2026 · similarity 0.89)
- 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.
