Safety1d ago
Safety
SFT Drives Gemini Safety Properties
Supervised Fine-Tuning (SFT) and pretraining drive most safety properties in Gemini, rather than reinforcement learning. Google DeepMind researchers found this result counterintuitive during interpretability tests. This discovery shifts how the team approaches safety training. Practitioners should note that RL may play a smaller role in alignment than previously assumed for this model family.