Redwood ResearchResearch
We don’t understand whether GDM intentionally trains on CoT

We were surprised to notice that the Gemini 4 Safety report seems to mention in passing that they train on the chain of thought: “Some components of the RL training procedure do observe thoughts as…
Read at Redwood Research ↗More from Redwood Research on Accept All

ResearchAstra covertly underperforming on safety research is concerning, even though it’s unlikely to be motivated by ambitious goals Redwood Research

ResearchOpenAI agents keep disabling monitoring; is this being handled appropriately? Redwood Research

ResearchTowards understanding score-seeking in LLMs Redwood Research

ResearchA basic case for accelerating conceptual reasoning capabilities Redwood Research

ResearchSimulated Control Evaluations Redwood Research
