Large volumes of data are stored in various types of tables, for example spreadsheets, web tables, or databases. Both industry and academia aim to leverage Large Language Models (LLMs) or Visual Language Models (VLMs) to automate table processing tasks. A research team lead by Professor Jing Zhang and Professor Yueguo Chen provide a comprehensive review of recent academic research and published their review article on 15 Feb 2025 in Frontiers of Computer Science co-published by Higher Education Press and Springer Nature.

They classify four prevalent table types: spreadsheets, web tables, databases, and documents.

They analyze which stages in the data lifecycle can benefit from a specific engineering task powered by AI and LLM.

For training methods, they summarize techniques for LLMs and VLMs training tailored for table processing. The key problem is curating table datasets and pre-train or fine-tune a table model. For instance, transform existing online datasets into formats suitable for instruction-tuning, or guide more powerful LLMs to synthesize task-specific datasets.

For prompting techniques, the key problem is building an LLM-powered agent, utilizing the robust reasoning capabilities of large language models.

In addition to these AI methods and approaches, they summarize the current datasets and benchmarks available.

Although many AI methods have been applied in various applications like Microsoft Excel or Google Sheet, and have achieved high rankings on certain benchmarks, they are still unable to address complex table processing tasks or ambiguous user queries. Moreover, LLMs should give some slow and deep thinking as table processing tasks need complicated reasoning.

DOI: 10.1007/s11704-024-40763-6