Reproduce “ToolSpace: Scaling Tool-Centric Environment Synthesis for Long-Horizon 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
Guanting Dong, Zhichao Sun, Xinyu Yang, Rui Wang, Xin Deng, Jiajie Jin, Zhixiong Yang, Junhao Huang and 17 more
ToolSpace builds executable cross-service task paths from observed tool-call traces, linking adjacent calls through data, state, and retry dependencies. It samples paths whose inputs are available, replays them from initial states, and retains those passing checks. ToolSpace-35B-A3B averages 41.0 across seven benchmarks, versus 23.0 for its backbone and 43.2 after on-policy RL; replay validates chains, not agent success.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.