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

arXiv cs.CL / 3/18/2026

💬 OpinionIdeas & Deep AnalysisModels & Research

共有:

Key Points

LLMs typically linearize 2D tables into 1D sequences, weakening row-column adjacency and layout cues.
Pure visual encoders can capture spatial cues but often struggle to preserve exact cell text.
The article shows that vision and text provide highly complementary information to LLMs, but simple fusion yields limited gains and can cause cross-modal interference.
They propose DiVA-Former, a lightweight architecture that uses visual tokens as dynamic queries to distill long textual sequences into digest vectors, effectively leveraging complementary vision–text information.
Across 13 table benchmarks, DiVA-Former improves the pure-text baseline by 23.9% and consistently outperforms baselines using visual, textual, or both inputs.

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.