Large Language Models (LLMs) have taken the world by storm, demonstrating impressive abilities in writing, translation, and even coding. But can they truly *reason* like humans? A fascinating new study puts LLMs to the test using "steamroller problems," logic puzzles designed to challenge deductive reasoning. Researchers explored whether LLMs could not only get the right answers but also follow the logical steps humans and automated theorem provers (ATPs) use. The results were surprising. While LLMs like GPT-3, GPT-4, and Google's Gemini could often get the correct answer, their *process* for getting there wasn't always sound. In fact, the study found a low correlation between correct reasoning and correct answers. This means that even when an LLM gets a steamroller problem right, it might not have arrived there through logical deduction. The researchers also found that LLMs are much better at "bottom-up" reasoning, starting with basic facts and building toward a conclusion, rather than "top-down" reasoning, which starts with a goal and works backward. This aligns with how LLMs are typically trained. These findings raise important questions about the reliability and explainability of LLM reasoning. Even state-of-the-art models can't always be trusted to think logically, even when they seem to give the correct answer. The research suggests that future development might explore neuro-symbolic methods, which combine LLMs with more traditional, rule-based systems to achieve more robust, trustworthy reasoning capabilities. This is particularly important in areas like law, healthcare, and finance where explanations and justifications are critical.