- Build verifiable data synthesis pipelines using rendering engines, symbolic systems, and code sandboxes to generate challenging visual reasoning data and multi-turn agent trajectories. Establish a continuously improving data flywheel through real-world data mining, rejection sampling, difficulty stratification, deduplication, contamination prevention, and data mixture optimization.
3 days
