How would these penalties ensure that the CoT reflects the actual concepts the AI is reasoning about and not just any statistically likely human-sounding words?
Such as Astra thinking about a "calm visual scene" while actually doing an unrelated hidden side-task that's invisible in the CoT (see model card)?