Frequency Is Not Sensitivity Identifying Safety-Sensitive Experts in Sparse MoE LLM
Router-gradient expert selection suppresses MoE LLM refusals more effectively than activation frequency, dropping OLMoE refusals 73.5% without breaking outputs.
The paper shows that suppressing a small set of routed experts can weaken sparse MoE LLM safety without retraining, and proposes router-gradient sensitivity to identify which experts matter. Across five MoE architectures tested on benign and malicious prompts, gradient-based selection reduced refusals more than activation frequency in 24 of 25 conditions. In OLMoE, refusals fell from 34 to 9 of 100 malicious prompts (73.53% relative) with no degraded outputs. A cross-model analysis links malicious-versus-benign routing concentration gaps to stronger gradient effects (rho = 0.90, p = 0.083, n = 5).
- Suppressing few routed experts can disable MoE safety behavior without retraining
- Gradient-based expert selection beat activation frequency in 24 of 25 conditions
- OLMoE refusals dropped from 34 to 9 of 100 prompts with intact generation quality
- Effect persisted after matching expert counts per layer in 23 of 25 conditions
From the original article
1 min read at arxiv.orgSuppressing a small set of routed experts can weaken the safety behavior of a sparse Mixture-of-Experts (MoE) language model without retraining. Which experts to suppress is therefore a security question, and the usual answer is activation frequency, but frequency measures use, not influence.
By Md Nurul Absar Siddiky · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.