Researchers have developed a method to preserve scientific sentence meaning through hierarchical JSON representations generated by lightweight LLMs. The team fine-tuned a language model using a novel structural loss function to transform complex scientific sentences into structured JSON formats that maintain their original meaning across diverse scientific domains.

Scientists collected sentences from academic papers and converted them into hierarchical JSON structures. These structured representations then served as input for a generative model tasked with reconstructing the original text, enabling the researchers to test whether JSON formats could effectively capture the nuanced information present in scientific language.

Comparing original and reconstructed sentences using semantic and lexical similarity metrics, the team demonstrated that hierarchical JSON formats successfully retain scientific text information with remarkable accuracy. This approach offers a new way to structure and process complex scientific language, potentially improving how researchers manage and analyze academic literature.