LensVLM-9B by Apple

Authors:

,

,

,

,

,

,

,

,

,

Vision Language Models (VLMs) recommendation the breathtaking possible of handling content as rendered images, bypassing the need for tokenizing the content into lengthy token sequences. Since VLM depiction encoders map fixed-size images to a fixed figure of visual tokens, varying rendering resolution provides a fine-grained compression knob. However, accuracy deteriorates quickly as compression increases: characters shrink below the imagination encoder's productive resolution, making them indistinguishable. To location this, we propose LensVLM, an conclusion example and post-training formula that enables VLMs to scan compressed images, afterward selectively develop lone the applicable images to their uncompressed form via learned tools. Building on Qwen3.5-9B-Base, LensVLM maintains accuracy comparable to the full-text high border at 4.3x productive compression and outperforms retrieval-based, text- and visual-compression baselines up to 10.1x productive compression throughout seven text QA benchmarks. LensVLM additionally generalizes to multimodal document and code understanding tasks, alongside the accuracy acquire complete baselines expanding as compression increases. Our inspection validates this approach: training makes visual compression sturdy to rendering choices, and as compression grows the example increasingly relies on expanded satisfied fairly than unreliable ocular reading. The inspection additionally yields applicable tool-choice guidance: content enlargement is preferable for rendered text, during high-resolution depiction enlargement suits native documents whose layout cues transport task-relevant information.