AgentGen: Automating Environment and Task Generation to Enhance Planning Abilities in LLM-Based Agents with 592 Environments and 7,246 Trajectories

AGENTGEN is a novel framework proposed by researchers from the University of Hong Kong and Microsoft Corporation to automate the generation of environments and planning tasks for LLM-based agents. This framework uses a two-stage process involving environment generation from a diverse inspiration corpus and task generation through a bidirectional evolution method called BI-EVOL. The system successfully created 592 unique environments and 7,246 high-quality trajectories, significantly improving the planning capabilities of LLMs like Llama-3 8B, surpassing models like GPT-3.5 and GPT-4 in certain tasks. AGENTGEN offers a scalable and efficient alternative to manual design, enhancing the overall training and performance of intelligent systems.