Large Language Models (LLMs) excel at many tasks, from writing poems to summarizing complex topics. But can they plan? Can they figure out the sequence of actions needed to achieve a specific goal, like a robot navigating a room or a character in a game solving a puzzle? New research dives into this question, exploring the limits of LLMs as planning formalizers. The surprising finding? While LLMs struggle to generate plans directly, they show promise in a different role: translating human-like descriptions into formal logic that computers *can* use for planning. This approach, called 'LLM-as-formalizer,' involves converting natural language into a structured language like PDDL (Planning Domain Definition Language). Think of it as bridging the gap between how humans describe problems and how computers solve them. The study tested several LLMs, including open-source and larger models like GPT-4, on the classic Blocks World puzzle. They were given descriptions of block arrangements and asked to generate the corresponding PDDL, which was then fed into a planning solver. The results revealed a clear trend: Larger models were significantly better at generating usable PDDL, outperforming smaller, open-source models. More interestingly, when an LLM *could* generate valid PDDL, this method significantly outperformed asking the LLM to devise the plan directly. However, the research also highlighted challenges. The more natural and human-like the problem descriptions became, the harder it was for the LLMs to translate them into formal logic. This suggests that implicit information, easily understood by humans, still trips up AI. The study also explored whether LLMs were simply memorizing patterns from training data by testing them on a modified Blocks World with nonsensical names. LLMs acting as formalizers proved robust to this change, suggesting they genuinely understand the underlying logic, not just the specific vocabulary. This research sheds light on a crucial aspect of AI development: bridging the gap between human language and computer logic. While LLMs still have much to learn about planning in complex real-world scenarios, this approach offers a promising path forward for building more capable and human-understandable AI systems. Future research aims to improve open-source models and tackle more intricate domains with partially observed environments that require the AI to interact and gather information, bringing us closer to truly intelligent agents.