FocusVTC: Efficient and High-Performance Visual Text Compression with Adaptive Resolution
FocusVTC adaptively compresses text as images, beating Glyph on RULER at similar compression.
FocusVTC compresses long text by rendering it as images and raising resolution only on relevant regions, avoiding a fixed-DPI trade-off. Trained on 29.4K Reasoning-Evidence Localization examples and optimized with Group Relative Policy Optimization, it scores 87.4 on RULER v1 at 2.9x compression versus 57.5 for Glyph at 3.0x. It slightly beats its text backbone on LongBench (56.40 vs 55.86), lifts MRCR by 13.91 points, scores 51.19 on VTCBench, and shows a 2.79x online speedup while MMMU and MME stay at or above the backbone.
- Scores 87.4 on RULER v1 at 2.9x compression versus Glyph’s 57.5
- Beats its text backbone on LongBench, 56.40 versus 55.86
- Raises MRCR macro-average by 13.91 points and reaches 51.19 on VTCBench
- Reports a 2.79x end-to-end latency speedup over plain text
- Trains on 29.4K REL-CoT examples with REL-SFT and GRPO
From the original article
1 min read at huggingface.co
Long-context reasoning in large language models incurs substantial computation and memory costs. Visual text compression (VTC) reduces input length by rendering text as images, but fixed-resolution rendering creates a compression-performance trade-off: low DPI saves tokens at the expense of legibility, whereas high DPI spends tokens on irrelevant content.
By FangZhi Zhong, Xuerui Qiu, Yuqi Pan · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.