Safety
107d agoSmall Specialist Models Outperform Sonnet In Audits
A Gemma 2B specialist judge outperformed Sonnet 4.5 in detecting misalignment within out-of-domain safety prompts. The model distinguished between secure and insecure fine-tuning where the larger model failed. This suggests narrow, specialized judges offer better transparency and lower costs for audits.
The Signal
Practitioners can now use smaller models for targeted alignment verification.