Most RL research for computer-use or GUI agents struggles with reliably measuring success at scale because rewards are hard to define and verify. CUA-Gym addresses that gap by packaging each task with both the executable setup artifacts and a programmatic Python reward function that deterministically checks completion, letting researchers scale verifiable training and evaluation for desktop and web agent behaviors.
What Sets It Apart
- Verifiable, executable tasks: every task pairs an instruction with setup files (scripts, documents, mocks) plus a reward.py that computes a programmatic score — so evaluations measure final state correctness, not heuristic proxies. This reduces ambiguous human labeling and enables automated large-scale runs.
- Viewer-friendly metadata layout: the release exposes a single Parquet tasks table for fast browsing, filtering, and sampling (7,897 tasks in the current subset), while storing original task bundles inside a compressed artifacts archive to avoid tens of thousands of tiny files in the hub repo.
- Designed for computer-use agents and RLVR: targets desktop and mock-web apps (e.g., libreoffice, vscode, instagram_mock) and supports research on verifiable rewards, synthetic task generation, and post-training filtering.
Who It's For & Trade-offs
Great fit if you build or evaluate GUI/web agents and need reproducible, programmatic rewards for large-scale training runs. It suits researchers who can run tasks in isolated VMs/containers and who want a metadata-first workflow for sampling and analysis. Look elsewhere if you need an out-of-the-box hosted evaluation service — several web task endpoints are stored as placeholders and require you to deploy the CUA-Gym-Hub mock apps or use the provided hosted defaults only for small smoke tests. Also note rewards check final state (not full episode traces), so CUA-Gym is optimized for end-state verification rather than detailed behavioral attribution.
Where It Fits
Think of CUA-Gym as a middle layer between synthetic task generators and full production evaluation harnesses: it gives reproducible, programmatic scoring and compact metadata for experiment orchestration, but expects users to provide sandboxed execution environments and to materialize web endpoints for large-scale runs.
