Agentic data science with the Data Agent Kit
cloud.google.com·6h ago
AI Research
arXiv CS.AI·7d ago·5 min read
Frequency Is Not Sensitivity Identifying Safety-Sensitive Experts in Sparse MoE LLM
TL;DR
Router-gradient sensitivity outperforms activation frequency in identifying safety-sensitive experts in sparse Mixture-of-Experts (MoE) language models, significantly reducing malicious prompt refusals without degrading output quality.
✦ Why It Matters
Implement router-gradient sensitivity for expert selection to improve safety in your MoE models today.
Key Takeaways
- Use router-gradient sensitivity for expert suppression in MoE models.
- Expect significant reductions in malicious prompt refusals.
- Maintain output quality while enhancing safety measures.
- Evaluate expert influence rather than just usage frequency.
How It Works
Router-gradient sensitivity assesses how changes in gate weights affect the model's loss, providing a more accurate measure of an expert's influence on safety. This contrasts with activation frequency, which only indicates how often an expert is used, not its impact on model behavior.