dailyai.report

Google DeepMind trained Gemini 3 Flash to adopt specific values by midtraining it on synthetic documents describing those traits. The team followed this with chat finetuning to ensure the behaviors remained robust during out-of-distribution tasks. This two-step pipeline improves how developers bake specific alignment properties into LLMs before final deployment.