FocusVTC: Efficient and High-Performance Visual Text Compression with Adaptive Resolution

FocusVTC. Efficient and High-Performance Visual Text Compression with Adaptive Resolution. By FangZhi Zhong and six others, from the Institute of Automation, Chinese Academy of Sciences, the University of Chinese Academy of Sciences, and two other institutions. Published on the twenty-ninth of September. Its code repository is worked out: about three hundred source files, with a readme.

Long-context reasoning in large language models costs substantial computation and memory. Visual text compression shortens the input by rendering text as images. With a fixed rendering resolution, though, compression and performance pull against each other. Low DPI saves tokens at the expense of legibility. High DPI keeps the page legible, but spends tokens on content that is irrelevant. Any single setting is wrong somewhere: too blurry for the passage that matters, or too sharp for the many passages that do not.

FocusVTC breaks that trade-off by making resolution adaptive, while preserving the model's general multimodal capabilities. It reads compressed, low-DPI global views, and enhances selected regions when it needs to, integrating the enhanced views into reasoning that is already under way. It is the way a reader skims a document at arm's length and then leans in on the one paragraph that holds the answer.

Training rests on twenty-nine point four thousand high-quality chain-of-thought examples for what the authors call Reasoning-Evidence Localization. Each one links a reasoning trace to page indices and bounding boxes. Supervised fine-tuning on these examples, at multiple resolutions, teaches the model to localize the relevant regions. Group Relative Policy Optimization then learns when to enhance resolution and how to use what the enhanced view shows. No separate continual-pretraining stage is required. Fine-tuning teaches where to look. Reinforcement learning teaches whether looking closer is worth the tokens at that moment.

On RULER v1 at seventy-two DPI, FocusVTC scores eighty-seven point four at two point nine times input compression, tool observations included. Glyph, at three times compression, scores fifty-seven point five. On LongBench, FocusVTC surpasses its own text-input backbone, fifty-six point four against fifty-five point eight six. It improves the MRCR macro-average by thirteen point nine one points, and reaches a macro-average of fifty-one point one nine on VTCBench.

A latency evaluation on MRCR shows an online, end-to-end speedup of two point seven nine times over text input. General multimodal ability is preserved. MMMU rises from sixty-five point one two to sixty-six point seven three, and MME from about two thousand four hundred and twenty-four to about two thousand four hundred and fifty-seven.

A fixed resolution treats every part of a page alike. FocusVTC spends few tokens on a cheap overview and more only where the evidence sits, and the multimodal scores suggest that learning this habit did not cost the model its general abilities.