dailyai.report

Google DeepMind trained Gemini 3 Flash to adopt specific values by midtraining on synthetic documents and finetuning on simulated chat data. This two-step process ensures traits remain robust during out-of-distribution tasks. The research provides a blueprint for steering model behavior. Practitioners can now improve trait instillation by combining document-based midtraining with targeted supervised finetuning.