Large language models (LLMs) have shown impressive abilities, but their reasoning skills often fall short, especially when complex, multi-step thinking is required. They tend to rely on intuition, predicting the next word rather than truly reasoning. Think about solving a tough math problem – humans plan, calculate, and verify, while LLMs often rush to a response. Researchers have now developed a new framework called SWAP (Structure-aware Planning with Accurate World Model) to address this limitation and make LLMs reason more deliberately. SWAP guides the LLM's reasoning using structural information, like a roadmap to the solution, and uses a ‘world model’ to simulate the consequences of different actions. Imagine the LLM building an internal map of the problem, testing different routes before deciding on the best path. A key innovation in SWAP is its Generator-Discriminator architecture. The generator predicts the next step, while the discriminator acts like a quality control expert, ensuring the steps are logically sound. This process helps to refine the LLM's world model, leading to more accurate predictions and better reasoning. The developers found that SWAP significantly boosted LLMs' performance on various tasks, including math, logic, and coding challenges. This suggests that teaching LLMs to think step-by-step, like humans, could be the key to unlocking their full reasoning potential. While SWAP marks a significant leap, the quest for human-level reasoning in AI is far from over. Future research could explore more dynamic interactions with the world model and even teach LLMs to identify and correct their own errors, paving the way for more robust and reliable AI.