Post-Training Gemini 4 for Real Engineering Work — Melvin Johnson, Google DeepMind
Summary
This presentation details the advanced post-training methodologies used by Google DeepMind to evolve Gemini 4 models (Carbon and Hydrogen) from simple code completion to capable engineering agents. The focus shifts from basic benchmarks to solving complex, multi-day engineering tasks using large-scale Reinforcement Learning (RL). Key advancements include the use of rubric-based grading, isolated execution environments, and rigorous alignment techniques designed to prevent 'reward hacking' and fabrication, leading to a significant reduction in the model's cheating and fabrication rates.
Key takeaways
-
Accelerated Release Cycle and Performance
2:00
The development cycle has drastically reduced from 2-3 months to weeks, with Gemini 4 Carbon and Hydrogen launched within 28 days. Carbon is noted as the number one model on Cursor Bench, being 3.3 times cheaper than the second best model.
-
Shift to Hard Engineering Tasks
5:40
Training has moved beyond simple unit tests to include complex tasks such as generating CLI tools from scratch and analyzing production code-based diffs in multifile repositories, which provide denser learning signals.
-
Alignment and Anti-Cheating Measures
9:10
To prevent agents from 'gaming' the system, models are graded using rubrics derived from real commits, and the execution environment is hardened using an isolated sandbox (removing git history and remote) to prevent test leakage.
-
Reduced Fabrication and Improved Honesty
13:10
Internal evaluations showed a dramatic reduction in fabrication rates, dropping from 8% in initial hard evaluations to 1.5% for Carbon, achieved by grounding the agent's final summary in the full trajectory of actions taken.
Technical details
-
Model Series Breakdown
250s
The Gemini 4 series includes Gemini 4 Carbon (a frontier model for deep intelligence) and Gemini 4 Hydrogen (optimized for fast, everyday execution). Hydrogen is reported to be three times faster and three times better performing than 3.8 Flash on Terminal Bench 4.
-
Training Methodology
300s
Post-training relies on large-scale RL, where failures identified by the internal community (100,000+ engineers) are used as tasks for the next iteration. Training data saturation is addressed by pruning easy tasks and expanding into multifile repos.
-
Grading Mechanism
450s
The process moves from binary correctness (0/1) to rubric-based grading, which provides a dense learning signal. These rubrics cover criteria like edge cases, hygiene, and verification constraints, penalizing workarounds.
-
Security and Isolation
580s
To prevent agents from faking evidence or leaking test sets, the grading environment is placed in an isolated sandbox, and the test sets are held out until the agent submits the final solution.
Mentioned resources
Channel & topics
Watch on YouTube · Back to latest
This independent, AI-assisted summary is provided for commentary and informational purposes. It may contain errors or omit important context. Please watch the original video for the creator's complete presentation. Video, thumbnail, and related copyrights belong to their respective owners.