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.
1 stories, newest first · one card per story
Router-gradient expert selection suppresses MoE LLM refusals more effectively than activation frequency, dropping OLMoE refusals 73.5% without breaking outputs.
Today 1.6K readers · 2.5K views · 779 reads · all time 189K views