LookZoom Reproduction - TextVQA Binarized VQA: leaderboard
LookZoom v1: Qwen3-VL-Instruct at 2B/4B/8B. Fixed 500-question subsets (seed 0) of DocVQA val (ANLS),InfographicVQA val (ANLS),ChartQA test (relaxed accuracy),TextVQA val (VQA accuracy), plus 300 held-out V-CoT-Doc (strict match). Per-view budget B means B * 32^2 pixels preserving aspect ratio, not total tokens. Single view/AttWarp/replace answer from 1 view; ViCrop/ground-then-crop/GapSight/LookZoom use 2. LookZoom reader uses GRPO LoRA; pointer trained from own correct answers; GapSight is authors reimplementation. Distinct trained systems remain separate. Appendix S2: separate 1000-question subsets with paper prompt, Qwen2.5-VL-7B at 512x512 pixels, official baseline implementations. Only authors re-evaluated rows admitted; reported-source rows use different support. Metric: VQA accuracy (%).
Metric: VQA accuracy (%). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 3 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | LookZoom reproduction Qwen2.5-VL-7B ViCrop | 77.2 | |
| 2 | LookZoom reproduction Qwen2.5-VL-7B AttWarp | 76.7 | |
| 3 | LookZoom reproduction Qwen2.5-VL-7B Base | 75.4 |
Interactive version: theaggregate.ai/benchmark?slug=lookzoom-reproduction-textvqa-binarized-vqa · How It Works · Data refreshed daily, snapshot 2026-10-27.