Visual Token Routing - Game Click: leaderboard
Computer-use and held-out game-action prediction of trained visual-token-routing systems. VTR v1 compares systems trained from Qwen3.5-2B on CUA-World training trajectories and markov-ai/gaming-500-hours, with a frozen vision encoder. Dense, compression and routing configurations remain distinct trained systems. OSWorld uses 295 offline tasks, up to 100 steps per task, averaged over three seeds. CUA-World uses 512 tasks and up to 200 steps, with checklist verification by Gemini 3 Flash. Game evaluation predicts held-out actions from 36 games: 30,000 clicks, 26,880 moves, 3,120 drags and 30,000 key presses. Costs, token counts and prefill timings are not capability scores. No row is relabeled as the unmodified Qwen checkpoint. Question access: Primary result reviewed; exact scored release availability not confirmed.
Metric: Published success/accuracy score (%). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | VTR-study Qwen3.5-2B VTR (ours) (Benchmark Sweep Run: source=VisualTokenRouting; trained-system; native November 2026 v1 configuration) | 4.9 | |
| 2 | VTR-study Qwen3.5-2B Dense, 100% area (Benchmark Sweep Run: source=VisualTokenRouting; trained-system; native November 2026 v1 configuration) | 4.74 | |
| 3 | VTR-study Qwen3.5-2B Pixel-diff (Benchmark Sweep Run: source=VisualTokenRouting; trained-system; native November 2026 v1 configuration) | 4.59 | |
| 4 | VTR-study Qwen3.5-2B VTR, 1/2 area (ours) (Benchmark Sweep Run: source=VisualTokenRouting; trained-system; native November 2026 v1 configuration) | 4.54 | |
| 5 | VTR-study Qwen3.5-2B VTR sliding (ours) (Benchmark Sweep Run: source=VisualTokenRouting; trained-system; native November 2026 v1 configuration) | 4.34 | |
| 6 | VTR-study Qwen3.5-2B VTR, 1/4 area (ours) (Benchmark Sweep Run: source=VisualTokenRouting; trained-system; native November 2026 v1 configuration) | 4.31 | |
| 7 | VTR-study Qwen3.5-2B Frozen router (Benchmark Sweep Run: source=VisualTokenRouting; trained-system; native November 2026 v1 configuration) | 4.14 | |
| 8 | VTR-study Qwen3.5-2B Dense, 50% area (Benchmark Sweep Run: source=VisualTokenRouting; trained-system; native November 2026 v1 configuration) | 3.82 | |
| 9 | VTR-study Qwen3.5-2B Dense, 25% area (Benchmark Sweep Run: source=VisualTokenRouting; trained-system; native November 2026 v1 configuration) | 3.63 | |
| 10 | VTR-study Qwen3.5-2B Dense, 12.5% area (Benchmark Sweep Run: source=VisualTokenRouting; trained-system; native November 2026 v1 configuration) | 3.44 |
Interactive version: theaggregate.ai/benchmark?slug=visual-token-routing-game-click · How It Works · Data refreshed daily, snapshot 2026-11-06.