OpenAI studies training effects on GPT-6 Astra CoT monitorability

Researchers at OpenAI analyzed how training choices across pretraining, architecture, reinforcement learning, and fine-tuning affect chain-of-thought monitorability in AI models. They found that decreasing base model perplexity improves no-CoT capability, while reinforcement learning and training on model-generated data increase CoT controllability. These findings suggest that as frontier models scale, monitoring them through their chain-of-thought becomes increasingly challenging. Open article →