ToolSpace: Tool-Centric Environment Synthesis Scaling for Long-Horizon Agents
ToolSpace: Scaling Tool-Centric Environment Synthesis for Long-Horizon Agents
The paper proposes ToolSpace, a tool-first framework that synthesizes multi-environment tool-use benchmarks and training data by mining execution traces from real MCP services and open agent corpora. In contrast to data-centric approaches that first synthesize databases and then derive tools, it mines data, state, and retry dependencies from adjacent calls using deterministic rules, clusters a weighted tool-combination graph into environments, and generates tasks via state-aware random walks with rollback, decompiling them into natural-language tasks with assertion checks and retaining only samples that can be successfully replayed from the initial state. On AutomationBench and tau3-Bench Banking, this pipeline generates harder Pro versions, with 7 frontier models dropping by up to 38.2 points. On seven long-horizon tool-use benchmarks, ToolSpace raises the average score of Qwen3.6-35B-A3B from 23.0 to 41.0 (43.2 after RL) and GLM-5.2 from 44.6 to 56.6, comparable to Claude Opus 5.
Related events and news
1 events · 1 news mentions
News that mentions this paper
人大复旦等提出ToolSpace智能体环境合成框架
- 公众号 机器之心cites the paper's id or link人大复旦等提出ToolSpace:744B模型仅经监督微调,七项基准平均分追平Claude Opus 5
论文标题:ToolSpace: Scaling Tool-Centric Environment Synthesis for Long-Horizon Agents