From points to rooms: enhancing large language model comprehension through concrete representations

Nowadays, large language models (LLM) have made significant progress in understanding and generating natural language and have become an important focus of research in the field of artificial intelligence. The capabilities of these models have been validated in a variety of applications, including machine translation, question-answer systems, and text generation. However, there are still challenges in understanding and processing spatial data, e.g., perception of abstract graph data. This study proposes a method to convert abstract graph representations into concrete object representations to enhance the impact of LLM comprehension. In the study, correspondences between abstract graph elements and concrete objects were first identified, such as mapping nodes in the graph to concrete objects and transforming the properties of edges into distances or connectivity between these spaces. Then, the graph data are transformed into easy-to-understand textual descriptions by constructing concrete scene descriptions. The experimental results show that, compared with the direct input of abstract graph, the concrete object representation can significantly improve the understanding and information processing ability of the model. This suggests that by converting abstract data into concretized descriptions, the language comprehension capability of LLM can be utilized more effectively, thus improving its performance in processing complex data. This study not only provides new perspectives for understanding the behavior of LLM when dealing with different types of data, but also provides practical strategies for improving model performance.