What it is
A recursive training framework for computer-use agents that mines real workflows from open-world videos, turns them into controllable task primitives, and iteratively improves data and models. Instead of closed-loop self-improvement on self-generated traces, it uses a double flywheel: open-world video mining for real seed workflows, then primitive-based synthesis plus rollback-assisted rollout to build diverse, verifiable GUI/CLI training data.
Key results
- 80.2 avg on OSWorld Verified, competitive with top closed-source agents
- 43.9 avg on WeaveBench, above UI-Mate-27B's 18.3
- 25.0 full-pass on Agents' Last Exam, above Qwen-CUA's 5.3
Shows open-world seeded, recursively targeted data generation scales computer-use agents better than closed self-improvement alone.