CUA-Gym: Verifiable Computer-Use Agent Tasks for Reinforcement Learning

CUA-Gym provides a collection of tasks for training computer-use agents with programmatically verifiable rewards. Each task pairs a natural-language instruction with executable setup artifacts and a Python reward function. The public preview contains a randomly sampled subset of the full task pool, authored by xlangai and last updated on 2026-05-26.

Use Cases

• Training reinforcement learning agents based on natural-language instructions
• Benchmarking agent performance based on programmatically verifiable rewards
• Developing new RLVR (Reinforcement Learning with Verifiable Rewards) algorithms based on the provided task structure
• Studying generalization in computer-use agents based on the diverse task pool

Strengths

• Tasks include programmatically verifiable Python reward functions
• Each task pairs a natural-language instruction with executable setup artifacts
• Public preview is available on Hugging Face as of 2026-05-26

Limitations

• Description metadata is limited; actual data quality requires manual inspection after download
• Row count is unknown, which may limit suitability assessment
• Column-level documentation is absent; field semantics must be inferred after download

Provenance

Source

xlangai

Collection Method

Collection of verifiable computer-use agent tasks, likely created for research purposes.

Freshness

Last updated 2026-05-26 07:49:28