OpenAI Scientist Says Chain-of-Thought Monitoring Remains a Core Research Goal

OpenAI Chief Scientist Jakub Pachocki said he wants to avoid a race toward models becoming harder to monitor because of misleading reporting. According to ChainCatcher, he said current frontier models, including Astra, have computational graph depth no more than twice that of GPT-4. Pachocki said OpenAI has worked since its first reasoning models to preserve and use chain-of-thought monitoring, which it believes helps observe how model alignment generalizes beyond the training distribution. He added that chain-of-thought monitoring is fragile and the trend is not encouraging, but said it can still be strengthened through research and remains a core goal of OpenAI's current research plan. The Information previously reported that Astra uses recurrent depth, allowing the same set of Transformer layers to compute repeatedly. Some reasoning can therefore occur inside the model rather than being fully written out as a chain of thought, raising concerns that models may become increasingly difficult to monitor.