Synthetic document finetuning for instilling positive traits

Short summary

Google DeepMind trained Gemini 3 Flash to exhibit specific alignment traits through synthetic document midtraining and chat-based SFT, then evaluated generalization across OOD adversarial scenarios. Four safety evals tested robustness against agentic misalignment, delusion validation, constraint violation, and auditor escalation. The methodology successfully instilled traits robustly, advancing practical alignment techniques for frontier models.

• •Synthetic document midtraining + chat SFT pipeline trained Gemini 3 Flash to exhibit specific alignment traits
• •Evaluated generalization via OOD safety evals: agentic misalignment, delusion validation, constraint violation, auditor escalation
• •Traits generalized robustly across distribution shifts, supporting principles-based alignment approaches

Generated with AI, which can make mistakes.