MLCommons

DeepMind pilots a double-blind way to test its own models against secret benchmarks

A cryptographic "box" built on confidential computing lets outside evaluators test a proprietary model without seeing its weights, and without Google seeing the test questions.