How to Utilize Complementary Vision-Text Information for 2D Structure Understanding

TL;DR

This paper introduces DiVA-Former, a novel architecture that effectively combines visual and textual information to improve 2D table understanding, outperforming existing methods across multiple benchmarks.

Contribution

The paper proposes DiVA-Former, a lightweight model that uses visual tokens as dynamic queries to better integrate vision and text for 2D structure understanding.

Findings

DiVA-Former improves table understanding accuracy by 23.9% over text-only baselines.

It achieves consistent gains over existing multimodal approaches.

The model effectively exploits complementary vision-text information.

Abstract

LLMs typically linearize 2D tables into 1D sequences to fit their autoregressive architecture, which weakens row-column adjacency and other layout cues. In contrast, purely visual encoders can capture spatial cues, yet often struggle to preserve exact cell text. Our analysis reveals that these two modalities provide highly distinct information to LLMs and exhibit strong complementarity. However, direct concatenation and other fusion methods yield limited gains and frequently introduce cross-modal interference. To address this issue, we propose DiVA-Former, a lightweight architecture designed to effectively integrate vision and text information. DiVA-Former leverages visual tokens as dynamic queries to distill long textual sequences into digest vectors, thereby effectively exploiting complementary vision--text information. Evaluated across 13 table benchmarks, DiVA-Former improves upon the pure-text baseline by 23.9\% and achieves consistent gains over existing baselines using visual inputs, textual inputs, or a combination of both.