The training methods try not to apply any particular rules to the contents of the thinking text. That's called "optimization pressure on CoT" and is thought to reduce safety by inducing the model to lie (or stop clearly printing its intentions) in the thinking text.
alt Hacker News
The training methods try not to apply any particular rules to the contents of the thinking text. That's called "optimization pressure on CoT" and is thought to reduce safety by inducing the model to lie (or stop clearly printing its intentions) in the thinking text.