Reproduce “DeskForge: Dense Supervision from Desktop Environments for Computer-Use Agents” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
A. Said Gurbuz, Ahmed Nassar, Sunghwan Hong, Marc Pollefeys, Peter W. J. Staar
DeskForge generates controlled Linux desktop scenes from running applications and densely labels visible elements and click outcomes. It fine-tunes action models on single-target examples. Qwen3.5-4B accuracy rises from 76.26% to 87.54% across four held-out conditions, and improves on five external benchmarks. With a fixed planner, it solves more WebArena-Infinity and OpenApps tasks, but results cover tested settings.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.