Renmin University, Fudan and others propose ToolSpace: a 744B model with only supervised fine-tuning matches Claude Opus 5 on the average score across seven benchmarks
744B模型追平Claude Opus 5,人大复旦提出ToolSpace探索环境合成新范式
AI summary
Teams from Renmin University, Fudan University and Shanghai Jiao Tong University propose ToolSpace, a tool-centric environment synthesis framework; its 744B model matches Claude Opus 5 on the average score across seven long-horizon tool benchmarks. The framework extracts tool dependencies from more than 200,000 interaction trajectories and builds a composition graph, synthesizes environment data on demand, and generates cross-environment tasks that can be verified by replay; ToolSpace-744B-A40B uses only supervised fine-tuning, and its seven-benchmark average score rises from 44.6 to 56.6.
Related papers