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 model maintains these properties even in out-of-distribution scenarios. The approach provides a scalable method for aligning model behavior without relying solely on human-curated datasets.