The research introduces EnvCraft, a system for generating synthetic environments for Agentic Reinforcement Learning. The framework utilizes an environment synthesis engine to create isolated sandbox workspaces and a data generation engine to produce task trajectories. A total of 139 interactive environments were synthesized, containing approximately 20,000 complex tasks. Qwen3/3.5 models (8B-32B) were evaluated, demonstrating improvements of up to +11.9% on claw-style benchmarks and +8.0% on general tool-use benchmarks. The research indicates that synthesized environments provide effective learning signals for training agent systems.
Source: https://arxiv.org/abs/2609.05576