Redwood Research questions whether Google DeepMind trains on Gemini 4 chains of thought
AIRedwood Research says Gemini 4's safety report suggests Google DeepMind uses auxiliary rewards that depend on chain-of-thought content, which it calls concerning because judges could penalize misaligned reasoning. The authors say DeepMind's statement that it does not penalize chain-of-thought text conflicts with the report, and they ask for clarification.
Source details
Redwood Research Blog · Alek Westover · blog.redwoodresearch.orgPublished · added here
Original title: We don’t understand whether GDM intentionally trains on CoT