ProCUA-SFT Technical Report

A new dataset called ProCUA-SFT, containing 3.1 million step-level supervised fine-tuning samples, has been released to train computer-use agents that interact with graphical desktops. The dataset enables a 7-billion-parameter model to achieve a 45.0% success rate on the OSWorld benchmark, an 18.7 percentage-point gain over its base performance [1][2]. The dataset was produced by a fully automated pipeline that synthesizes grounded tasks on live desktops seeded with real-world content, including 912 spreadsheets from SpreadsheetBench and approximately 10,000 permissively-licensed presentations from Zenodo [1][2]. A single vision-language model, Kimi-K2.5, served as goal generator, precondition judge, and trajectory executor, which the authors state eliminates planner-actor capability gaps [1][2]. Each of the 93,000 synthetic trajectories was expanded into step-prefix samples that reproduce the context layout seen at inference time [1][2]. The work addresses a documented shortcoming of the largest prior public resource, AgentNet, which contains 22,500 human trajectories. When used for supervised fine-tuning, continuing training on AgentNet caused the OSWorld success rate of UI-TARS 7B to fall from 26.3% to between 8% and 10%, a phenomenon the researchers describe as negative transfer [1][2]. Fine-tuning the same base model on ProCUA-SFT for a single epoch reversed that decline, producing the 45.0% score and surpassing AgentNet-trained counterparts by more than 35 percentage points [1][2]. A subset of ProCUA was incorporated into the training data for the Nemotron 3 Nano Omni model, contributing to its computer-use capabilities [1][2]. The paper appears on arXiv, which since 2022 has integrated with Hugging Face Spaces to allow researchers to link interactive demos directly from abstract pages, making it possible for readers to test models without writing code [3][4][5]. The computer-use agent domain sits within the broader field of large language models, which are trained with self-supervised learning on vast amounts of text to perform natural language processing tasks [7].