Vision-Language Models can process text as rendered images, but accuracy degrades with compression; LensVLM addresses this by scanning compressed images and selectively expanding relevant parts through learned tools, maintaining high accuracy even at high compression ratios.

Vision Language Models[-1–> (VLMs) offer the exciting possibility of processing text as rendered images, bypassing the need for tokenizing the text into long token sequences. Since VLM image encoders map fixed-size images to a fixed number of [0–>visual tokens[-1–>, varying [0–>rendering resolution[-1–> provides a fine-grained compression knob. However, accuracy deteriorates quickly as compression increases: characters shrink below the vision encoder’s effective resolution, making them indistinguishable. To address this, we propose LensVLM, an inference framework and post-training recipe that enables VLMs to scan compressed images, then selectively expand only the relevant images to their uncompressed form via learned tools. Building on Qwen3.5-9B-Base, LensVLM maintains accuracy comparable to the full-text upper bound at 4.3x effective compression and outperforms retrieval-based, text- and visual-compression baselines up to 10.1x effective compression across seven [0–>text QA benchmarks[-1–>. LensVLM also generalizes to multimodal document and [0–>code understanding[-1–> tasks, with the accuracy gain over baselines growing as compression increases. Our analysis validates this approach: training makes [0–>visual compression[-1–> robust to rendering choices, and as compression grows the model increasingly relies on expanded content rather than unreliable visual reading. The analysis also yields practical tool-choice guidance: text expansion is preferable for rendered text, while high-resolution image expansion suits native documents whose layout cues carry task-relevant information.