OpenAI reports its own experiments showing chain-of-thought controllability rises during reinforcement learning and can be passed on through training data.
Research: What Affects CoT Monitorability Across The Model Lifecycle?
What it means for youOpenAI says its most capable models are getting harder to oversee by reading their reasoning, one of the main ways the company checks for misbehaviour.
For lawmakersA lab's safety cases lean on chain-of-thought monitoring, and OpenAI now says progress in pretraining is likely to require new approaches to monitorability.
For researchersOpenAI finds no-CoT capability tracks perplexity, controllability rises in RL and transmits via fine-tuning, yet controllability is a noisy proxy for monitor evasion.
“with GPT-6 Astra we observed degradation in CoT monitorability for non-adversarial settings, greater ability to perform CoT monitor evasion under adversarial testing, and significantly greater no-CoT capability compared to prior models.”
The post's introduction, explaining what motivated the study.