Recently, Liu Tianyou, an undergraduate student from the Class of 2025 at the Shenzhen-Hong Kong Microelectronics Institute, Southern University of Science and Technology, has had his research, as a co-first author, accepted by EMNLP 2026—a highly esteemed international conference in the field of natural language processing. The paper, titled ".tmu: A Low-Entropy Tree-Structured Representation for LLM-Assisted Scientific Writing," delves into the representation of scientific documents. It proposes and evaluates a low-entropy tree-structured representation format, .tmu, which organizes sections, formulas, citations, and other content into tree nodes with well-defined boundaries. This approach offers a novel implementation path for large language models (LLMs) to accurately locate document content, merge manuscripts of varying styles, and rectify document errors.

The research team selected four prominent LLMs—Deepseek-v3.2, Claude-Sonnet-4.5, Gemini-3-pro, and ChatGPT-5.2—to conduct three types of experiments: document structure localization, merging documents of different styles, and error correction. The experimental results revealed that, under the evaluated conditions, .tmu achieved superior scores in both document merging and error correction tasks. Moreover, in the document structure localization task, three out of the four models utilizing .tmu obtained higher scores.

Commencing from the data representation of scientific documents, this research offers fresh perspectives for the design of LLM-assisted writing tools. It holds the potential to enhance the efficiency and reliability of content localization, manuscript integration, and error correction in scientific writing, marking a significant step forward in the field.