as far as I understand CoT - thinking tokens don't truly mean what model is thinking about, that's why interpretability is such a big deal still
and on top of that frontier providers don't let you see real CoT only showing little summaries
add a skeleton here at some point
about 19 hours ago