Synthetic document finetuning for instilling positive traits
Google DeepMind researchers have developed a method to instill positive traits into their Gemini 3 Flash model. This approach involves two stages: first, midtraining the model on synthetic documents that describe Gemini exhibiting desired properties, and second, finetuning it on synthetic chat data where it demonstrates these traits. The study found that chat finetuning was particularly effective in robustly embedding these traits, even in out-of-distribution scenarios, and shared insights for improving both midtraining and supervised finetuning effectiveness. AI
IMPACT This research demonstrates a novel method for aligning AI models with desired traits, potentially improving safety and reliability in future AI systems.
