ViPanelTR: A Multi-Agent Framework for Vietnamese Table Question Answering

Abstract

Table question answering over real-world data remains challenging in zero-shot settings, where models must reason over irregular table structures and handle unanswerable queries. This is particularly evident in Vietnamese benchmarks such as Open-ViTabQA, which contain merged cells, reasoning-intensive questions, and explicit Null cases. We present ViPanelTR, a new multi-agent inference framework for zero-shot Vietnamese Table QA. ViPanelTR uses five role-conditioned agents for structural, logical, numerical, evidential, and synthesis-oriented reasoning, applies an early answerability gate, and performs bounded self-review, peer deliberation, and answer normalization. Experiments on Open-ViTabQA show consistent improvements over matched single-agent baselines across three backbones, reaching up to 80.66% in terms of F1-score. Additional comparisons with CoAgt and CoQ show that the gains are not only due to using multiple agents, but also to the proposed role design, answerability handling, and normalization strategy.