LookZoom Pointer Training - Adaptive-View Accuracy: 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. 2B pointer-training ablations; mean over 98/196/392 per-view tokens. Evidence mass is a localization diagnostic and is not substituted for answer accuracy. Adaptive-View Accuracy. 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-2B pointer <look> token ( ours ) | 70.6 | |
| 2 | LookZoom Qwen3-VL-2B pointer w/o outcome gate | 70.2 | |
| 3 | LookZoom Qwen3-VL-2B pointer Shared attention | 69.5 | |
| 4 | LookZoom Qwen3-VL-2B pointer w/o sharpening | 66.8 |
Interactive version: theaggregate.ai/benchmark?slug=lookzoom-pointer-training-adaptive-view-accuracy · How It Works · Data refreshed daily, snapshot 2026-10-27.