Model
18d agoGoogle Tests Double-Blind AI Benchmarking
Google DeepMind partnered with the Singapore AI Safety Institute to pilot double-blind evaluations using Gemini Flash Lite. Cryptographic protection via Confidential Space hides test questions from Google and model weights from evaluators. This prevents data contamination during testing.
The Signal
Practitioners gain a more reliable, tamper-proof method to verify frontier model performance without compromising proprietary IP.