Redwood Research

Redwood reports measurements suggesting GPT-6 Astra can perform more substantial cognition without verbalizing it in chain-of-thought when prompts include filler tokens, raising monitoring concerns.

Redwood argues that chain-of-thought controllability evaluations may be under-elicited and that prompt optimizations can materially change measured performance.

Redwood describes an operationalization (a proxy metric) for “opaque serial depth” between text bottlenecks as a way to reason about latent-reasoning-like architectures and monitorability risks.

A shorter Redwood blog post pointing readers to the full report and summarizing key takeaways.

Redwood publishes its independent investigation report with a detailed timeline and scope, including datasets reviewed and methods for analyzing agent collaboration and behavior during the stated in-scope window.

Redwood announces ControlConf 2026 and lists proposed themes for talks and discussion around monitoring, auditing, and AI control research challenges.

Redwood describes experiments where short, strong advice can improve weak LLM performance in a control context, including evaluations run on SWE-bench Verified and BashArena.