Large language models (LLMs) have shown impressive abilities in solving complex reasoning tasks, especially when using chain-of-thought (CoT) prompting, where they explicitly lay out their reasoning steps. However, CoT is computationally expensive. Researchers are exploring "implicit CoT," where LLMs arrive at answers without showing their work. But do these models *actually* reason through the problem internally? A new study challenges this assumption. Researchers investigated how LLMs handle multi-step arithmetic problems using implicit CoT, analyzing the models' internal states to see if they genuinely perform step-by-step calculations. Surprisingly, when simply *prompted* to use implicit CoT, LLMs often guessed the correct answer without actually calculating the intermediate steps, especially in problems with multiple steps. However, when specifically *trained* to use implicit CoT, the models did demonstrate internal step-by-step calculations. This suggests that simply asking an LLM to give a direct answer doesn't guarantee genuine reasoning. There's a critical difference between learned implicit reasoning and prompted implicit reasoning. Furthermore, the study reveals a significant weakness of implicit CoT: both prompted and trained models struggled when the problem format was slightly altered (e.g., reversing the order of equations), even though the underlying difficulty remained the same for a human. This fragility highlights the limitations of current implicit CoT methods and suggests that explicit CoT remains essential for reliable and robust reasoning in LLMs. While implicit CoT holds the promise of faster computation, it’s not yet a reliable replacement for showing your work. For now, explicit CoT seems to be the most effective method for complex tasks, ensuring that LLMs aren't just taking educated guesses but genuinely understanding the problem.