dailyai.report

Google DeepMind trained Gemini 3 Flash to adopt specific values by midtraining it on synthetic documents. The team followed this with chat finetuning to ensure these traits remained robust during out-of-distribution tasks. This two-step process proves more effective than standard SFT. Practitioners can now better control model behavior through targeted synthetic document generation.