AI reasoning models were supposed to be the industry’s next leap, but now a string of research is calling that into question.