LookZoom Reader Ablation 98 - 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 crops/decisions but base versus RL-adapted reader; not the same trained checkpoint. Exact budget 98. Five-Benchmark Average. Metric: Accuracy (%).
Metric: Accuracy (%). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | LookZoom Qwen3-VL-8B reader ablation Ours | 67.4 | |
| 2 | LookZoom Qwen3-VL-4B reader ablation Ours | 63.6 | |
| 3 | LookZoom Qwen3-VL-8B reader ablation Our reader, GapSight’s crops | 62.9 | |
| 4 | LookZoom Qwen3-VL-8B reader ablation GapSight † , base reader | 62.6 | |
| 5 | LookZoom Qwen3-VL-2B reader ablation Ours | 60.1 | |
| 6 | LookZoom Qwen3-VL-4B reader ablation Our reader, GapSight’s crops | 58.7 | |
| 7 | LookZoom Qwen3-VL-4B reader ablation GapSight † , base reader | 57.3 | |
| 8 | LookZoom Qwen3-VL-2B reader ablation Our reader, GapSight’s crops | 53.9 | |
| 9 | LookZoom Qwen3-VL-2B reader ablation GapSight † , base reader | 49.9 |
Interactive version: theaggregate.ai/benchmark?slug=lookzoom-reader-ablation-98-five-benchmark-average · How It Works · Data refreshed daily, snapshot 2026-10-27.