AI Safety

Google DeepMind uses synthetic docs to instill positive traits in Gemini 3 Flash

Training on fake Reddit threads and emails to teach AI principles robustly.

Deep Dive

Google DeepMind's interpretability team adapted methods from Marks et al and Li et al to train Gemini 3 Flash on synthetic documents (Reddit threads, emails, research papers) that describe the model having positive traits, then fine-tuned on chat data where it demonstrates those traits. They used a two-stage pipeline: midtraining on non-chat documents followed by SFT on synthetic chat. According to the article, they used LMSYS and agentic coding evals to check for capability regression, and a collection of OOD safety evals to test whether the model exhibited aligned behavior in scenarios different from the training data.

Key Points
  • Two-stage pipeline: midtraining on synthetic docs (Reddit, emails, papers) followed by SFT on synthetic chat data
  • Used Gemini 3.1 Pro to generate scenario prompts and responses, then removed system prompts during training
  • Measured robustness via LMSYS/agentic evals (capability) and OOD safety evals (generalization)

Why It Matters

Proves synthetic document training can instill robust, generalizable positive traits in frontier models without capability loss.

📬 Get the top 10 AI stories daily