LookZoom Window Placement 4B 196 - Five-Benchmark Average: 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. Same 0.5 min(W,H) square window. Reader is the RL-adapted reader; compare single/random/pre-answer/look placement. Ground-truth evidence-box oracle is excluded. Metric: Accuracy (%).
Metric: Accuracy (%). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | LookZoom Qwen3-VL-4B square-window pre-answer pointer | 75.4 | |
| 2 | LookZoom Qwen3-VL-4B square-window look pointer | 75.1 | |
| 3 | LookZoom Qwen3-VL-4B square-window single view | 66 | |
| 4 | LookZoom Qwen3-VL-4B square-window random placement | 65.8 |
Interactive version: theaggregate.ai/benchmark?slug=lookzoom-window-placement-4b-196-five-benchmark-average · How It Works · Data refreshed daily, snapshot 2026-10-27.