A Table Is Worth 64 Tokens: Pixel-level Compression for Multi-Table Document Question Answering

TL;DR - This paper introduces a training-free, two-stage approach that uses heavily compressed table images for relevance filtering, then native-resolution tables for question answering. On long multi-table documents, it reduces total token use by 41% while improving accuracy by 7 points over native-resolution single-step QA.

• Native-resolution table images match text-based representations in performance and efficiency across the evaluated VLMs.
• Moderate downscaling hurts readability and triggers longer, less effective reasoning traces, offsetting expected token savings.
• Highly compressed tables retain enough information for models to identify which tables are relevant.
• The two-stage method uses 15% fewer tokens than the most efficient single-step compressed setup without sacrificing accuracy.