MATA Framework Introduces Multi-Agent Approach for Table Question Answering

A newly developed multi-agent framework, known as MATA, aims to tackle issues in Table Question Answering (TableQA). Utilizing various reasoning methods and tools created with small language models, the system generates potential answers and subsequently refines or chooses the best one. MATA features an algorithm that reduces costly LLM agent calls, thereby boosting efficiency. It demonstrates strong performance with small, open-source models and can easily adapt to different LLM types. Comprehensive experiments were performed on two benchmarks of differing difficulty levels using ten distinct models. This strategy seeks to improve reliability, scalability, and efficiency, particularly in environments where resources are limited or privacy is a concern. Despite recent advancements in Large Language Models enhancing table understanding tasks, challenges persist.