Reproduce “Scaling Verifiable Environments for Long-horizon Work 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
Jiazheng Zhang, Long Ma, Yunxian Yang, Zhiheng Xi, Zhikai Lei, Yajie Yang, Chenyang Liao, Enyu Zhou and 18 more
WorkForge turns expert workflows and accessible real resources into verifiable workspace-based training tasks. It retrieves files, profiles them, and uses checkable facts to shape tasks and verification. Real-resource workspaces average 42.3 files versus 7.9, while fine-tuning on WorkForge trajectories raises Qwen3.5-35B-A3B-Base GDPval rubric score from 45.5 to 73.6 and APEX score from 5.0 to 21.3. These gains do not isolate causes.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.