StanislavKrym comments on The ‘People Pleaser’ Problem in LLMs

[StanislavKrym]

I think that it was actually already done with reasoning models. They type some tokens in the CoT, then proceed to write the actual answer which they show the users.