Multimodal Tabular Reasoning with Privileged Structured Information

Han-Jia Ye (Nanjing University) · Qingguo Chen (Alibaba Group) · Weihua Luo (Alibaba Group) · Kaifu Zhang (Alibaba Group) · Jun-Peng Jiang (NanJing University) · De-Chuan Zhan (Nanjing University) · Yu Xia (Alibaba Group) · Hai-Long Sun (Nanjing University) · Shiyin Lu (Nanjing University)
logical inferencelogical stepsmodality-bridged informationmulti-step information extractionmultimodal large language modelsprivileged structured informationreasoning skillsreasoning tracesstate-of-the-art performancestructure-aware reasoning trace generatorstructured text tablestable imagestabular reasoningvisual inferencevisual representations

Tabular reasoning requires complex, multi-step information extraction and logical inference, such as aggregation, comparison, or calculation over tabular data. While recent advances have leveraged large language models (LLMs) for reasoning over structured text tables, such high-quality textual representations are often unavailable in real-world settings, where tables typically appear as images. In this paper, we tackle the task of tabular reasoning directly from table images. Our core strategy is to leverage privileged structured information