Summary
CUA-Universe is a pipeline that converts real desktop software into hybrid GUI+CLI environments (App-Forge VM adaptation across 16 applications, Task-Weave task synthesis, Path-Steer efficient hybrid-path rollout collection). A 9B model trained on the resulting data gains +39.3 CUA-Verse score with -37% steps and -60% tokens; on OSWorld it adds +16.8 SR with -57% steps, and +7.84 on OSWorld-MCP.
Why it matters
The efficiency gains are as large as the accuracy gains: hybrid GUI+CLI action spaces let computer-use agents take shell shortcuts instead of pixel-grinding through dialogs. For teams building computer-use agents, synthetic hybrid environments are now demonstrably the training-data lever.
Technical details
| Arxiv | 2609.05374 (Mon 7 Sep digest, announced 2026-09-08T00:00Z; coverage-gap recovery — same digest as ev-20260908-01..11 but not selected then, first seen 2026-09-11 run) |
|---|---|
| Pipeline | App-Forge (VM adaptation, 16 apps) + Task-Weave (task synthesis) + Path-Steer (hybrid-path rollout collection) |
| Results | 9B model: CUA-Verse +39.3 / steps -37% / tokens -60%; OSWorld SR +16.8 / steps -57%; OSWorld-MCP +7.84 |
| Code | not stated in abstract |
Tags
computer-usegui-agentsynthetic-datatraining