Answering questions on real‑world documents often requires processing long inputs that mix text and tables. Optical context compression converts the context into images, promising large token savings, yet its impact on table comprehension is unclear. We investigate pixel‑level table compression for multi‑table document QA, evaluating five visual‑language models (VLMs) on two benchmarks under five visual‑token budgets. Results show that representing tables as images at native resolution matches text‑based performance and efficiency, whereas downscaling tables forces models to compensate for reduced readability with longer, less effective reasoning traces, nullifying the expected token reduction. Even heavily downscaled tables retain enough signal for the model to decide whether they are relevant to a question. Leveraging this asymmetry, we propose a training‑free two‑step approach: the model first identifies the tables needed to answer a question from a pixel‑compressed context, then reasons over those tables at native resolution. On long documents, this method saves 41% of total tokens and improves accuracy by 7 points compared to single‑step QA with native‑resolution tables, while using 15% fewer tokens than the most token‑efficient single‑step compressed setup, with no loss in accuracy.
Blogger's Review: The study presents a practical strategy that quickly filters relevant tables using low‑resolution images and then performs high‑resolution reasoning, offering an efficient pathway for large models to handle lengthy documents.