Abstract

The learning from practice paradigm is crucial for developing capable AgenticAI systems, yet it is severely hampered by inefficient experience generation, abottleneck especially pronounced in complex benchmarks like GAIA. To addressthis, we introduce AWorld, an open-source system engineered for large-scaleagent-environment interaction. By distributing tasks across a cluster, AWorldaccelerates experience collection by 14.6x compared to standard single-node,sequential execution. This critical speedup makes extensive reinforcementlearning practical and scalable. Leveraging this capability, we trained aQwen3-32B-based agent that significantly outperforms its base model, increasingits overall GAIA accuracy from 21.59% to 32.23%. On the benchmark's mostchallenging levels, our agent achieves a score of 16.33%, surpassing theperformance of leading proprietary models. Our open-source system and resultingagent provide a practical blueprint for a complete agentic AI trainingpipeline, from efficient interaction to demonstrable model improvement.