Renmin University of China, Fudan University and other institutions propose a new ToolSpace paradigm: a 744B model matches Claude Opus 5
1 day ago / Read about 0 minute
Author:小编   

Researchers from Renmin University of China, Fudan University and other universities have proposed ToolSpace, a tool-centric cross-environment agent environment synthesis framework that breaks through existing data-centric environment synthesis paradigms. The framework extracts three types of dependencies among tools—data, state, and retry—from a large number of real interaction trajectories, constructs a tool composition graph, and clusters and associates related tools as needed to generate corresponding environments and underlying data. It then generates executable cross-environment call chains through state-aware random walks and backtracking sampling, and inversely synthesizes natural language tasks with verification scripts. Experiments show that the 744B model trained on this framework achieves an average score of 56.6 on seven long-horizon tool benchmarks, matching Claude Opus 5. The team also built the more challenging AutomationBench Pro and τ³-Bench Banking Pro benchmarks based on this framework, offering a scalable new path for long-horizon agent training and evaluation.