[Faint]

Remember that these models generate one token at a time. They do not "think ahead" much more than maybe a few tokens in beam search. So if the problem requires search - actual comparison of approaches, and going back-and-forth between draft and thinking through the implications - the model can't do it (except in a limited sense, if you prompt it to give it's "train of thought"). So it's comparable of you being in front of whiteboard, hit with a question, and you would have to start answering immediately without thinking more than you can while talking through your answer at the same time. Doable if you know the material well. If it's a new problem, that approach is doomed. Given that, I think the language models do remarkably well. A little bit of search, and maybe trying to generate the answer in different order (like short draft -> more detailed draft -> more detailed draft... etc.) will improve things a lot.