Current methods create training environments for AI agents by first building fake databases, so tools behave unrealistically and tasks rarely span multiple services. ToolSpace flips that: it analyzes real traces of tool use, maps how tools depend on each other, then builds environments and tasks from those dependencies. Agents trained this way jumped from 23 to 41 average score on multiple benchmarks, matching a top model, showing that authentic tool patterns matter more than model size.
Tool-centric method synthesizes harder long-horizon agent tasks from execution traces
ToolSpace reverses the data-centric order of environment synthesis, mining tool dependencies from real traces to create cross-environment tasks that reduce frontier model scores by up to 38.2 points.
Top University
Guanting Dong · Zhichao Sun · Xinyu Yang · Rui Wang · Xin Deng · Jiajie Jin · +19 more
Renmin University of China · Fudan University · Shanghai Jiao Tong University
Research Digest··2 min read
Thread:Process Agent Benchmarks
In one line
ToolSpace synthesizes training environments for agents by mining tool dependencies from execution traces, improving agent performance.
Read it as
The authors propose ToolSpace, a framework that reconstructs a tool-composition graph from real execution traces of MCP services and open agent corpora, then synthesizes long-horizon tasks via state-aware random walks.
The authors analyze the limitations of current data-centric environment synthesis, where tools and tasks are derived from a synthesized database, leading to unrealistic tool interfaces and compositions.
How it works
before and afterThe Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.