Intel / post

Cua's RL-Trained Decision Model Hits 92.9% on GUI-360, Open-Sourced

Key takeaways · AI-distilled

Training has two stages: supervised training teaches the decision format, then agentic RL with RLOO runs the model in live cua-bench-basic environments where only a completed task earns the environment's reward.
At each step the model gets the screen state, the task goal and a fixed set of candidate actions, and returns one action; the environment then changes and the next step starts from the new state.
On held-out tasks, Cua-S1-4B-0.2 completed 17 of 18 text episodes and 13 of 18 , versus 16 and 12 for zero-shot djev, a much smaller gap than the 92.9% versus 60.1% result on the frozen 168-task GUI-360 split.
Both GUI-360 runs used the same inputs and scoring with no accessibility tree. The training code and results are merged into the Cua repo, and a Hugging Face Space lets users see how the model scores candidate actions.

Terms in this piece · Glossary

multimodal — A model that works with more than text — reading images, audio, or video, and sometimes generating them too.
benchmark — A standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.
open weights — A model whose trained parameters are published for anyone to download and run — unlike API-only models you can access but never possess.

Why it matters

A concrete, reproducible RL recipe and for training small models to make GUI decisions from raw screen state, without needing an accessibility tree.