Skip to content
EventPast event

Study: Reducing experts in MoE inference weakens refusal; SafeStart mitigates it

2 reports2 reporting sourcesLatest report

Start here

AI overview

A paper from the University of Southern California and Intel Labs (arXiv 2611.06147) found that MoE LLMs, at inference time, only reduce the number of experts activated per token without changing weights, and this also significantly weakens refusal capability, while accuracy on general capabilities such as MMLU remains almost unchanged. On Qwen1.5-MoE, reducing the number of experts per token from 4 to 3 raised the compliance rate on harmful requests under a refusal-suppression template from 46.6% to 65.0%; OLMoE and DeepSeek-V2-Lite rose by 9.5 and 6.6 percentage points respectively; Qwen3-30B and Moonlight-16B instead showed a decrease in compliance rate, and other models showed no statistically significant change. The researchers found that on Qwen1.5-MoE, of 101 requests that shifted from safe to harmful, 92 diverged from the original model at the first generated token, and observed that the hidden state of the last prompt token shifted along the refusal direction. Prior research proposed SafeStart, which, on a single RTX A6000 with vLLM and batch size 8, retaining about two-thirds of the selected experts can speed up offline decoding by 15% to 28% and largely restore refusal.

Generated by AI from reports · Updated 7 days ago

Attention and developments

Attention, hourly

Now 42, 0% in 24 hours·Peak 44 (Nov 13 03:00)

Nov 14Nov 15Nov 16Nov 17Nov 18Nov 19Nov 20Now 42

Progress timeline

Nov 14
  1. XiaohongshuXiaohongshu · Greatest Computer Scientist· Discussion
    arxiv 2026 | LLM safety refusal and expert pruning(opens the original post in a new tab)
Nov 12
  1. WeChat · Mo'anjuSelected
    MoE inference acceleration also weakens the safety capabilities of large models

    A paper from the University of Southern California and Intel Labs (arXiv 2611.06147) finds that MoE models, when only the number of activated experts is reduced at inference time without changing weights, also significantly weaken refusal capability, while accuracy on general capabilities such as MMLU remains almost unchanged. On Qwen1.5-MoE, reducing the number of experts used per Token from 4 to 3 raised the compliance rate with harmful requests under a refusal-suppression template from 46.6% to 65.0%, while OLMoE and DeepSeek-V2-Lite rose by 9.5 and 6.6 percentage points respectively; Qwen3-30B and Moonlight-16B instead saw their compliance rates decrease, and other models showed no statistically significant change. The researchers found that of 101 requests on Qwen1.5-MoE that changed from safe to harmful, 92 diverged from the original model at the first generated Token, and observed that the hidden state of the last Prompt Token shifted along the refusal direction.

Nov 4
  1. arXiv: Jailbreaks, Prompt Injection, Poisoning & Defenses
    SafeStart: Mitigating the Overlooked Safety Cost of Activated-Expert Reduction in MoE Large Language Model Inference Serving

    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.

Related discussion

Total 1 items

Attention trend

New sources per day

2 in 9 days·most 1 (Nov 12)·today 0

  • Nov 12: 1 new sources (1 outlets, 0 accounts)
  • Nov 13: 0 new sources (0 outlets, 0 accounts)
  • Nov 14: 1 new sources (0 outlets, 1 accounts)
  • Nov 15: 0 new sources (0 outlets, 0 accounts)
  • Nov 16: 0 new sources (0 outlets, 0 accounts)
  • Nov 17: 0 new sources (0 outlets, 0 accounts)
  • Nov 18: 0 new sources (0 outlets, 0 accounts)
  • Nov 19: 0 new sources (0 outlets, 0 accounts)
  • Nov 20: 0 new sources (0 outlets, 0 accounts)

Hourly attention

Current attention 42·Peak in the comparable period 44 (Nov 13 03:00)·Change over the comparable past 24 hours 0%

0204060Nov 1303:00Nov 1511:00Nov 1718:00Nov 2002:00