AI Models Persist on Wrong Answers, Unlike Humans Who Disengage.

New research highlights a key divergence in problem-solving strategies between humans and large reasoning models (LRMs). While both exhibit longer deliberation times for more difficult problems, their internal mechanisms differ significantly when facing incorrect solutions. Specifically, LRMs tend to expend more computational resources, measured in tokens, when they arrive at a wrong answer compared to when they solve the same problem correctly. In contrast, humans demonstrate the opposite behavior, spending less time on problems they ultimately fail to solve. This pattern suggests that LRMs' extended processing on incorrect answers is likely driven by uncertainty, as they continue to search for a solution. Humans, however, appear to employ an engagement-versus-abandonment strategy, investing more time in problems they anticipate solving and disengaging from those they expect to fail. This distinction is crucial for understanding and improving AI reasoning capabilities.