Think AI has mastered logic? Think again. Large Language Models (LLMs) have shown impressive abilities, but how well do they *actually* reason? A new research paper, "Multi-LogiEval: Towards Evaluating Multi-Step Logical Reasoning Ability of Large Language Models," puts LLMs through a rigorous logic exam. Researchers created Multi-LogiEval, a dataset designed to test multi-step reasoning across different types of logic, including propositional logic (like "if this, then that"), first-order logic (which deals with objects and their relationships), and even non-monotonic logic, a type of reasoning that’s closer to how humans think, where conclusions can change with new information. The results? LLMs stumbled when the logic got complex. As the reasoning chains grew longer, their accuracy plummeted, especially with four or five steps involved. Interestingly, the size of the LLM didn't guarantee success. Smaller open-source models sometimes outperformed their larger counterparts, showing that bigger isn’t always better when it comes to logical thinking. Why the struggle? Analysis reveals that LLMs often misinterpret evidence or go down a rabbit hole of unnecessarily long reasoning chains, losing their way to the correct conclusion. The research also highlights the importance of the context. Longer contexts sometimes improved accuracy by giving the LLMs more information to work with, but overly long reasoning steps made errors snowball. The study uncovered how different types of logic posed different challenges, showing that LLMs struggle more with some types of reasoning than others. This research provides valuable insights into the limitations of current LLMs and paves the way for creating smarter, more logical AI in the future. It underscores that simply building bigger models won't cut it; we need to develop new techniques to truly teach AI to think logically.