Table-Critic: A Multi-Agent Framework for Collaborative Criticism and Refinement in Table Reasoning

TL;DR

This work tackles error propagation in table reasoning by introducing Table-Critic, a four-agent framework (Judge, Critic, Refiner, Curator) that collaboratively identifies, critiques, and refines intermediate reasoning steps. It leverages a self-evolving template tree to accumulate critique patterns and guide future reflections, enabling iterative improvement until correctness. Empirical results on WikiTableQuestions and TabFact show notable accuracy gains over decomposition-based and critic-based baselines across multiple LLMs, with robust performance and manageable cost increases. The approach demonstrates effective error localization, correction, and continual learning for structured data reasoning, offering a scalable path toward more reliable table reasoning in real-world applications.

Abstract

Despite the remarkable capabilities of large language models (LLMs) in various reasoning tasks, they still struggle with table reasoning tasks, particularly in maintaining consistency throughout multi-step reasoning processes. While existing approaches have explored various decomposition strategies, they often lack effective mechanisms to identify and correct errors in intermediate reasoning steps, leading to cascading error propagation. To address these issues, we propose Table-Critic, a novel multi-agent framework that facilitates collaborative criticism and iterative refinement of the reasoning process until convergence to correct solutions. Our framework consists of four specialized agents: a Judge for error identification, a Critic for comprehensive critiques, a Refiner for process improvement, and a Curator for pattern distillation. To effectively deal with diverse and unpredictable error types, we introduce a self-evolving template tree that systematically accumulates critique knowledge through experience-driven learning and guides future reflections. Extensive experiments have demonstrated that Table-Critic achieves substantial improvements over existing methods, achieving superior accuracy and error correction rates while maintaining computational efficiency and lower solution degradation rate.

Figures (13)

• Figure 1: An illustration of Table-Critic, a multi-agent framework for table reasoning tasks, where the Judge identifies errors, the Critic provides detailed critique, the Refiner corrects the reasoning process, and the Curator updates a self-evolving template tree to accumulate critique knowledge and improve future performance.
• Figure 2: Analysis of Model Convergence and Iteration Requirements on WikiTQ and TabFact Datasets.
• Figure 3: Computational cost and Effectiveness Comparison between SC (Self-Consistency Based on Chain-of-Table) and our Table-Critic.
• Figure 4: An example of self-evolving mechanism in our Template Tree.
• Figure 5: Instructions for the Judge Agent. These instructions outline the procedure for verifying the correctness of a predicted answer and identifying errors within the reasoning process.