Rendering Tables as Images Only Pays Off If You Downscale for Selection and Then Re-Read at Native Resolution

arXiv 2608.26949·medium signalXBlueskyLinkedInCopy link

Testing five VLMs across two benchmarks and five visual-token budgets, representing tables as images at native resolution matches text on both accuracy and efficiency, but downscaling backfires because models compensate for lost readability with longer, weaker reasoning traces that cancel the token savings. The asymmetry the authors exploit is that heavily downscaled tables still carry enough signal to decide whether a table is relevant to the question. Their training-free two-step method (identify relevant tables from the compressed context, then reason over those at native resolution) saves 41% of total tokens and gains 7 accuracy points over single-step native-resolution QA on long documents.