Large Language Models (LLMs) have impressed us with their abilities in various domains, but planning remains a significant challenge. While they can generate impressive results on familiar tasks, LLMs often falter when faced with novel or complex scenarios that require genuine reasoning about actions and consequences. This post delves into recent research exploring why LLMs struggle with planning and reveals some promising strategies for improvement. Researchers investigated several techniques to enhance LLM planning abilities using an extended version of the PlanBench dataset, a benchmark for evaluating planning in LLMs. They found that simply fine-tuning LLMs on problem-plan pairs leads to brittle performance on out-of-distribution tests, especially those involving longer plans. This suggests that LLMs may be relying on memorization and pattern matching rather than true understanding of the planning domain. However, the research also uncovered some promising directions. While techniques like Chain-of-Thought (CoT) prompting and self-correction didn’t significantly improve overall plan validity, they *did* boost plan executability. This means the plans generated were more likely to be logically coherent and follow the rules of the domain, even if they didn't always achieve the desired goal. This suggests that LLMs are learning some fundamental reasoning skills, even if they struggle to put them all together for a successful plan. Interestingly, the study found that a variant of CoT, “State CoT,” helped the LLM understand state transitions better, improving performance in short problems. This shows that supplying more context and guiding the LLM’s reasoning process can be beneficial. The most significant improvements came from reinforcement learning (RL) using a novel reward function called Longest Contiguous Common Subsequence (LCCS). This reward encouraged the LLM to generate plans that were closer to valid reference plans, even when they weren't perfect. RL improved both the validity and executability of plans, particularly on longer, more challenging problems. The research suggests that LLMs *are* capable of improving their planning abilities, but they need better guidance and more sophisticated training methods. Simply throwing data at them isn't enough. Techniques like CoT and RL, combined with more informative rewards and better evaluations beyond simple validity checks, offer promising avenues for developing LLMs that can reason effectively and generate plans that work in the real world. The journey towards robust AI planning is still ongoing, but this research illuminates the path toward more strategic and effective approaches.