CompoWorld: Compositional Environment Scaling for General Agents
CompoWorld trains agents on verified multi-service workflows built from reusable tool environments.
The paper describes 448 services exposing 10,130 tools, with tasks generated by connecting services through dependency graphs. Its training set includes 3K supervised trajectories and 1K reinforcement-learning tasks for Qwen3.6-35B-A3B. The authors report a 9.17-point average gain over the backbone across eight benchmarks, and say the model tops compared 35B-A3B agent-specialized models on AutomationBench. Source: HF Daily Papers' note.
The paper describes 448 services exposing 10,130 tools, with tasks generated by connecting services through dependency graphs. Its training set includes 3K supervised trajectories and 1K reinforcement-learning tasks for Qwen3.6-35B-A3B. The authors report a 9.17-point average gain over the backbone across eight benchmarks, and say the model tops compared 35B-A3B agent-specialized models on AutomationBench. Source: HF Daily Papers' note.
score 5