Imagine an AI tasked with stacking blocks into a specific arrangement. Does it myopically focus on the very next move, or does it strategize several steps in advance, like a human playing chess? This question sits at the heart of understanding how today’s powerful Large Language Models (LLMs) plan and reason. New research from the Laboratory of Cognition and Decision Intelligence for Complex Systems sheds light on this by dissecting how LLMs tackle a classic planning puzzle called Blocksworld. This research delves into the inner workings of the LLM, examining how information flows and transforms as it plans, picking apart components like Multi-Head Self-Attention (MHSA) and Multi-Layer Perceptrons (MLPs) – the core building blocks of these models. What they discovered is intriguing: LLMs do, in fact, exhibit a form of “look-ahead” planning, encoding information about future moves in their internal representations. However, this ability isn't quite as sophisticated as human foresight. LLMs struggle to plan many steps ahead and rely heavily on recent history rather than grasping the bigger picture. This research uses clever techniques to trace how LLMs process the current state of the blocks, the desired goal state, and the sequence of moves already taken. They find that MHSA is crucial for extracting relevant information, focusing on the goal and the most recent actions. The study also probes what is stored in the LLM’s internal memory as it works through the puzzle. They discovered that the model encodes both the current arrangement of the blocks and, importantly, information about future moves. This ability to think ahead, while limited, suggests that LLMs are capable of more than just reacting to immediate stimuli. The implications of this work are significant for developing more advanced AI agents. Understanding how LLMs plan allows researchers to create more efficient, robust, and intelligent systems for complex tasks. While the research offers fascinating insights, challenges remain. Analyzing proprietary models like ChatGPT is difficult due to lack of access to their internal workings, and real-world planning tasks often lack the clear-cut right and wrong answers that simplify analysis in a controlled environment like Blocksworld. Nonetheless, this work offers a valuable peek into the mind of an AI planner, paving the way for future research that will unlock even greater planning capabilities in LLMs.