CENTS: A Flexible and Cost-Effective Framework for LLM-Based Table Understanding

Large Language Models (LLMs) have recently shown impressive capabilities in a variety of applications including table understanding tasks such as column type annotation. Existing LLM-based solutions for table understanding, however, focus on developing specific framework for each individual task, or do not consider the cost-effectiveness tradeoff. In this paper, we present Cents, a unified and cost-effective framework for LLM-based solutions for table understanding tasks. Cents's key capability is an efficient and effective approach to compress the tabular LLM input in a way that reduces input token cost while improving performance compared with state-of-the-art methods. Experiment results show that Cents outperforms other LLM-based baselines on a variety of table understanding tasks at the same or lower cost.