dailyai.report

Google DeepMind trained Gemini 1.5 Flash to adopt specific values by midtraining it on synthetic documents describing those traits. This process, followed by chat finetuning, ensures the behaviors remain robust even in out-of-distribution scenarios. The method provides a scalable blueprint for steering model alignment without relying solely on human-curated datasets.