dailyai.report

Google DeepMind trained Gemini 1.5 Flash to adopt specific values by midtraining it on synthetic documents. This process preceded finetuning on chat data to ensure the traits remained robust during out-of-distribution tasks. The team's findings provide a concrete blueprint for developers to program specific behavioral properties into LLMs without relying solely on human feedback.