- Who
- OpenAI
- What
- analyzed training factors that affect reasoning transparency and chain-of-thought controllability
- Quote
In a new analysis from OpenAI's safety team , researchers systematically studied which specific training steps trigger this erosion.
What happened
OpenAI’s safety team analyzed models from GPT-5 through GPT-6 Astra and identified four factors linked to reduced chain-of-thought transparency: base-model capability, architecture, reinforcement learning, and transfer from teacher models. The analysis finds that some architectures gain more on silent multi-step reasoning, while more reinforcement learning increases controllability in Astra-class models. It also describes a tradeoff: longer, more visible reasoning can be easier to monitor but raises deployment compute costs.
Why it matters
The analysis traces reduced chain-of-thought transparency to specific training choices and describes a capability and compute tradeoff, giving readers a framework for understanding how model design affects oversight.
Read next
See the full storyOpenAI Researchers Link Training Choices to Reduced Reasoning TransparencyThe story so far
OpenAI’s safety team analyzed models from GPT-5 through GPT-6 Astra and linked four training factors to reduced chain-of-thought transparency: base-model capability, architecture, reinforcement learning, and transfer from teacher models. The analysis found that some architectures gain more from silent multi-step reasoning, while additional reinforcement learning improves controllability in Astra-class models. OpenAI’s researchers described a tradeoff: longer, more visible reasoning can be easier to monitor, but it increases compute costs in deployment.