LookZoom 196 Tokens - V-CoT-Doc: 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. Exact per-view budget 196. V-CoT-Doc. Metric: Accuracy (%).
Metric: Accuracy (%). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 21 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen3-VL-8B-Instruct (Harness: LookZoom study LookZoom) | 86.3 | |
| 2 | Qwen3-VL-4B-Instruct (Harness: LookZoom study LookZoom) | 78.3 | |
| 3 | Qwen3-VL-8B-Instruct (Harness: LookZoom study GapSight reimplementation) | 77.7 | |
| 4 | Qwen3-VL-8B-Instruct (Harness: LookZoom study LookZoom replace) | 76.3 | |
| 5 | Qwen3-VL-2B-Instruct (Harness: LookZoom study LookZoom) | 76 | |
| 6 | Qwen3-VL-8B-Instruct (Harness: LookZoom study Ground-then-crop) | 74 | |
| 7 | Qwen3-VL-4B-Instruct (Harness: LookZoom study GapSight reimplementation) | 73 | |
| 8 | Qwen3-VL-4B-Instruct (Harness: LookZoom study LookZoom replace) | 71.3 | |
| 9 | Qwen3-VL-8B-Instruct (Harness: LookZoom study ViCrop) | 69 | |
| 10 | Qwen3-VL-2B-Instruct (Harness: LookZoom study LookZoom replace) | 68.3 |
Interactive version: theaggregate.ai/benchmark?slug=lookzoom-196-tokens-v-cot-doc · How It Works · Data refreshed daily, snapshot 2026-10-27.