Can large language models (LLMs) truly understand the nuances of meaning in sentences? A fascinating new study delves into this question by exploring how well LLMs grasp "thematic fit." Thematic fit describes how well a word or phrase fits into a specific role within a sentence's action. For example, in the sentence "The chef cooked the meal," "chef" fits perfectly as the agent doing the cooking, while "meal" fits naturally as the thing being cooked. This research probes whether LLMs can recognize these subtle relationships between words, actions, and roles. Researchers experimented with various prompting methods, including simple direct questions and more complex, step-by-step reasoning prompts. They also tested how giving LLMs more context, like complete sentences instead of just word pairs, influenced their understanding. Interestingly, the study found that while LLMs like GPT-4 show a remarkable ability to grasp thematic fit, exceeding previous models, their performance isn't always consistent. Step-by-step reasoning helped the models on some tasks but hindered them on others. Similarly, providing full sentences sometimes improved accuracy but in other cases, led to more confusion, potentially due to the models' difficulty in filtering out nonsensical sentences. The research reveals that while LLMs have come a long way in understanding meaning, challenges remain, particularly in deciphering complex relationships between words and actions. This work opens exciting avenues for future research in enhancing LLMs' ability to reason and truly grasp the subtleties of human language. The findings offer valuable insights for developing more nuanced and reliable AI systems capable of truly understanding the richness of human expression.