I asked ChatGPT to restate this in more laymen's terms (posted below) and I am not to surprised at the answer.<p>"Lately, some AI models have shown impressive abilities to solve complex problems, and many people credit this to a method called Chain of Thought (CoT), where the model is trained to think through steps like a human might. In this paper, we take a closer look at that idea to see if it's really what's driving better performance.<p>We focus on the model’s step-by-step thinking (the words it generates along the way) — often treated like human "thoughts" — and examine whether these actually help the model solve problems more accurately. To test this, we train AI models using clean, correct step-by-step reasoning paths and final answers, all based on a known solving method (A* search). This lets us check both the final answers and the reasoning steps to see how they relate.<p>Interestingly, we find that even when a model gives the right answer, its reasoning steps can still be wrong or messy. To go further, we even train models using completely random and incorrect reasoning steps — and surprisingly, they still perform about the same, and sometimes even better, than those trained on correct steps.<p>This suggests that the step-by-step "thoughts" the model shows aren’t as meaningful or reliable as many assume. In short, just because a model looks like it’s reasoning through a problem doesn’t mean it actually is — and we should be careful not to treat its outputs as if it thinks like a human or follows strict logic."