CoRef: A Collaborative Refinement Framework for Table Question Answering

ACL ARR 2025 February Submission2815 Authors

Abstract:

Table Question Answering (TQA) enables users to query semi-structured tables using natural language.
However, current methods struggle with two key challenges: (i) complex layouts which hinders accurate reasoning and (ii) substantial noise that disrupts table-processing code generation.
We propose CoRef, a collaborative refinement framework.
To tackle challenge (i), CoRef employs a Planner and multiple Table Curators, working alongside a Decision Trace Tree to distribute the burdens of decision-making and table curation across specialized agents, while also enabling backtracking when needed.
For challenge (ii), CoRef integrates a Code-Refining Memory module, which iteratively refines table-processing code by learning from compiler feedback.
CoRef outperforms SOTA methods in extensive experiments on three public TQA datasets (74.2\% on WikiTQ, 88.6\% on TabFact, and 74.7\% on HiTab), validating its effectiveness.

Paper Type: Long

Research Area: Question Answering

Research Area Keywords: Table QA,Agents,LLMs

Contribution Types: Publicly available software and/or pre-trained models

Languages Studied: English

Submission Number: 2815