Google DeepMind pilots cryptographic double-blind evaluation for Gemini model

Google DeepMind said it is piloting a double-blind external evaluation of a Gemini Flash Lite model using a cryptographically secured environment. The company is working with the Singapore AI Safety Institute, OpenMined, AVERI and MLCommons to test the model against confidential benchmarks while preventing evaluation material from being used later to optimize model performance. Google said the pilot adds technical safeguards to existing contractual and zero-logging protections intended to limit benchmark contamination.