dailyai.report

Google DeepMind trained Gemini 1.5 Flash to adopt specific values by midtraining it on synthetic documents before applying chat finetuning. This two-step process ensures traits remain robust even in out-of-distribution scenarios. The research provides a blueprint for steering model behavior without relying solely on human-curated datasets. Practitioners can now improve SFT effectiveness via targeted synthetic pre-conditioning.