AI Paper Atlas

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

ICLR 2026 · Large language models › Instruction and fine-tuning · approach: Safety alignment after fine-tuning · 0 citations (OpenAlex)

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 changedWhereMeasuredEffect
add SafeMoE routing regularization to fine-tuningvanilla LoRA fine-tuning · DeepSeek-V2 on SQLharmfulness score: 72.0 -> 4.0Improved
add SafeMoE routing regularizationvanilla fine-tuning · OLMoE on SAMSumharmfulness score: 62.0 -> 5.0Improved
add SafeMoE (vs Lisa parameter-drift constraint)Lisa · OLMoE on SAMSumharmfulness score: 21.0 -> 5.0Improved
add SafeMoE to fine-tuningvanilla fine-tuning · Mixtral on 500 harmful samplesharmfulness score: 78.0 -> 8.0Improved
add SafeMoE to fine-tuningvanilla fine-tuning · gpt-oss on 500 harmful samplesMMLU-Redux: 77.5 -> 79.6Improved
add SafeMoE to fine-tuningvanilla fine-tuning · OLMoE on SAMSumfine-tuning accuracy: 49.3 -> 48.9No clear effect
block expert-layer gradients into embedding in SafeMoE (ablation)SafeMoE · OLMoE on SAMSumharmfulness score: 5.0 -> 18.0Hurt
apply SafeMoE only to upper layers 8-15 (ablation)SafeMoE · OLMoE on SAMSumharmfulness scoreNo clear effect
apply SafeMoE to lower layers only vs upper layers (ablation)SafeMoE · OLMoE on SAMSumharmfulness scoreHurt
apply SafeMoE under adaptive attackSafeMoE · OLMoE on SAMSumharmfulness score: 73.0 -> 32.0Mixed
add SafeMoE under strong 5k harmful samplesvanilla fine-tuning · OLMoEharmfulness score: 72.0 -> 11.0Improved
SaLoRA initializationvanilla fine-tuning · OLMoE on SAMSumextra training time: 92.41%Hurt

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