p2-TQA: A Process-based Preference Learning Framework for Self-Improving Table Question Answering Models

The introduction of p2-TQA marks a significant advancement in table question answering (TQA) systems, which are designed to answer questions based on tabular data. Traditional methods often under-utilize available data and overlook post-training enhancements. p2-TQA addresses these issues by automatically generating process-based preference data through a table-specific pipeline, thus eliminating the need for costly manual data collection. The framework employs contrastive learning to optimize models, resulting in notable performance gains—up to 5% on in-domain datasets and 2.4% on out-of-domain datasets—with just 8,000 training instances. This efficiency is further highlighted by the fact that models enhanced with p2-TQA maintain up to five times higher efficiency compared to their larger counterparts. The competitive results achieved by p2-TQA against state-of-the-art TQA systems underscore its potential to revolutionize the field, making advanced TQA capabilities more accessible and…