Gemini 2.5
Google DeepMind — Gemini
rich detail 11 method fields published
Evaluation
Automated red teaming for safety
Multi-agent automated red teaming game: populations of attacker Gemini models elicit responses violating objectives (e.g., safety policy violations, unhelpfulness), scored by judges (prompted and finetuned Gemini models) as reward signal
How it was run
Open a release to see the setup, scoring, and other details that the lab published. Only fields the lab actually disclosed are shown.