Google DeepMind trains Gemini 3 Flash with synthetic docs to instill positive traits robustly
Synthetic document midtraining plus chat fine-tuning yields OOD-aligned behavior in frontier models.
Deep Dive
Google DeepMind's interpretability team used synthetic document midtraining on Gemini 3 Flash to instill positive traits. Inspired by Li et al. and Marks et al., they generated pretraining-style documents and synthetic chat data where the model demonstrated traits. The chat SFT proved effective for OOD generalization. They used a critique stage and autorater to ensure quality, and checked LMSYS and agentic coding evals to guard against significant capability regressions.
Key Points
- Method combines synthetic document midtraining with chat SFT, inspired by Li et al. (model spec midtraining) and Marks et al.
- Chat data generated by Gemini 3.1 Pro with traits in system prompt, then refined and filtered via critique stage and autorater.
- Robust trait instillation achieved without capability regression on LMSYS and agentic coding evals, and OOD safety evals passed.
Why It Matters
Enables alignment techniques that train principles on frontier models, improving safety and generalization beyond training data.