SafeStart: Mitigating the Overlooked Safety Cost of Activated-Expert Reduction in MoE Large Language Model Inference Serving
SafeStart: Mitigating the Overlooked Safety Cost of Active Expert Reduction in MoE LLM Serving
Translation in progress — untranslated text is omitted.
Researchers found that reducing the number of experts activated per token to speed up MoE large language model inference serving weakens the model's refusal of harmful requests, and proposed SafeStart, which largely restores refusal while retaining most of the speedup. On a single RTX A6000 with vLLM and batch size 8, retaining about two-thirds of the selected experts speeds up offline decoding by 15% to 28%; under a jailbreak attack instructing the model not to refuse, the proportion of harmful requests answered rose significantly only on Qwen1.5-MoE, OLMoE and DeepSeek-V2-Lite among the nine tested models, by 18.4, 9.5 and 6.6 percentage points respectively, while MMLU dropped by less than 2 points.
- MoE serving commonly accelerates by reducing the number of experts per token, and such methods are validated only with accuracy benchmarks. The authors state that they may weaken the model's refusal of harmful requests while accuracy is nearly unchanged, and that no prior work has examined this at the recommended operating point.
- The authors conduct a paired, noise-calibrated refusal audit of nine MoE models in vLLM, localize the refusal change to prompt processing and the first few generated tokens, and propose SafeStart: run all experts only at these starting positions and use the reduced budget for the remaining tokens, implemented via a fused vLLM kernel.
- Under the refusal suppression jailbreak, truncation retaining about two-thirds of experts raises ASR by 18.4 pp for Qwen1.5-MoE, 9.5 pp for OLMoE, and 6.6 pp for DeepSeek-V2-Lite, while MMLU drops by no more than 1.8 points; SafeStart (N=8) compresses the change of the three reduction methods to −2.1 to +1.3 pp and retains at least 85% of the offline decoding speedup.
- The authors note: templates that prescribe an opening or place the harmful task later can bypass SafeStart, and adaptive attacks were not tested; human annotation covers only 587 responses from two models; the largest model is 57B, and timing is limited to a single RTX A6000 and vLLM default kernels; the patch is single-turn and does not cover quantized MoE kernels or expert parallelism.
Related events and news
1 events · 1 news mentions
News that mentions this paper
研究:MoE推理减专家削弱拒答,SafeStart缓解
- 公众号 模安局cites the paper's id or linkMoE 推理加速也会削弱大模型的安全能力
2026 年 11 月,南加州大学和 Intel Labs 的研究人员发表论文《SafeStart: Mitigating the Overlooked Safety Cost of Active Expert Reduction in MoE LLM Serving》,揭示了一个此前容易被忽视的风险: