Skip to content
WeChat · Mo'anju·Indexed by this site · Originally published· Defenses →SelectedAttention42

MoE inference acceleration also weakens the safety capabilities of large models

AI summary

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.

Why it was selected

The paper quantified the degree of refusal-capability degradation caused by MoE expert reduction, and presents an inference modification that largely eliminates this degradation by protecting only the beginning of generation.

Related papers
Read originalmp.weixin.qq.com