The paper
Most UI agent scaling is currently throttled by the cost of human time. OS-Genesis took a much more scalable path by using Reverse Task Synthesis. Instead of recording a user completing a task, they started from a terminal state and worked backwards to hypothesize the intent: an agent explores the app with no task in mind, and the task description is written afterwards from what it did OS-Genesis.
Discussion
The intuition here is that it’s easier to verify a goal if you already have the state. By starting with the app state and generating the task description, every synthesized task is one the environment is known to reach, and a reward model grades the resulting trajectories rather than throwing incomplete ones away OS-Genesis.
The jump on AndroidWorld is the part that stood out to us the most. Going from 9.82% to 17.41% just by training on this synthetic data instead of task-driven synthetic data OS-Genesis suggests we’re nowhere near the ceiling for what small models can do if the data diversity is high enough.
For our work on CUAs, the reverse synthesis logic is a potential fix for the ground truth problem. If you can automate the trajectory, you can train on environments that humans haven’t manually mapped out yet.
What we took from it
- Reverse synthesis gives tasks with a known-reachable goal state, which makes them easier to verify (Environment Generation).
- Data diversity, not just volume, is what moved small GUI agents here; worth testing on agent-built environments (WorldSmith).
- We’re very interested in how synthetic data trajectories will define the next generation of CUAs. If you can automate the training data, you can automate the agent (Web Agents).