Post-Training Gemini 4 for Real Engineering Work — Melvin Johnson, Google DeepMind
Passing the tests does not mean a coding agent solved the problem. At CODE Summit 2026, Google DeepMind's Melvin Johnson explains Gemini 4 post training for real engineering work, from harder tasks and rubric based grading to isolated evaluation environments. He connects reinforcement learning, reward hacking, honest model behavior and feedback from Google's engineers to faster releases, alongside Gemini 4 Carbon and Hydrogen. Speaker info: - https://melvin.app/ - https://research.google/people/105147/ - https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-hydrogen/ Timestamps: 0:00 - Gemini 4 Carbon and Hydrogen 5:32 - From code completion to real engineering work 7:23 - Harder tasks and richer learning signals 10:00 - Grading agents without rewarding cheating 13:10 - Training agents to report honestly 15:30 - Shipping faster through engineering feedback




Join the discussion
Have something to add?
Sign in to reply, follow this thread, and earn XP when your answer helps someone.