What it is
A closed-loop suite for native visual reasoning with 300 procedural tasks, deterministic verifiable scorers, and controlled comparison across image, video, and interleaved generators. Instead of evaluating visual reasoning with ad hoc benchmarks or VLM judges, it builds a trainable testbed: 300 procedural tasks with aligned image/video/interleaved renderings and deterministic task-specific scorers usable for RL.
Key results
- 300 tasks; 3.47M images and 1.30M videos with verifiable scoring
- Often >20-point gains on 7 external visual reasoning benchmarks
- New tasks judged deeper in 113/150 blinded pairwise comparisons
Visual-state trajectories can be trained and optimized directly, and they matter more than language-only chains of thought for these reasoning tasks.