dailyai.report

Google DeepMind trained Gemini 1.5 Flash to adopt specific values by midtraining it on synthetic documents describing those traits. The team then applied chat finetuning to ensure these properties remained robust during out-of-distribution tasks. This multi-step approach proves more effective for behavioral alignment than standard finetuning. Practitioners can now better steer model personality via synthetic data.