# AgentGen: Enhancing Planning Abilities for Large Language Model based Agent via Environment and Task Generation

**Authors:** Mengkang Hu, Pu Zhao, Can Xu, Qingfeng Sun, Jianguang Lou, Qingwei Lin, Ping Luo, Saravan Rajmohan, Dongmei Zhang  
**arXiv:** [2408.00764](https://arxiv.org/abs/2408.00764)  
**Date:** 2024-08-01  
**Summary page:** https://www.summarizepaper.com/en/paper/2408.00764v1/

## Key points

- The paper focuses on enhancing planning abilities in Large Language Model (LLM) based agents through instruction tuning, also known as agent training.
- Planning ability is crucial for LLM-based agents to interact with the environment and execute actions to achieve desired goals from an initial state.
- Existing work has limitations in generating varied and extensive trajectories due to a focus on manually designed planning tasks and environments.
- The paper introduces a framework called AgentGen, which leverages LLMs to generate diverse environments and create planning tasks based on these environments.
- To improve environmental diversity, an inspiration corpus consisting of various domain-specific text segments is used as context for synthesizing environments.
- A bidirectional evolution method called Bi-Evo is introduced to increase the difficulty diversity of generated planning tasks by evolving them from easier and harder directions.
- Evaluation results show that AgentGen significantly enhances LLMs' planning ability, with the instruction-tuned Llama-3 8B model outperforming GPT-3.5 in overall performance and even surpassing GPT-4 in certain tasks.
- This research contributes to advancing the field of large language models by automating the synthesis of diverse environments and a range of planning tasks, ultimately improving the planning abilities of LLM-based agents.

## Summary

The paper "AgentGen: Enhancing Planning Abilities for Large Language Model based Agent via Environment and Task Generation" by Mengkang Hu et al. delves into the improvement of planning abilities in Large Language Model (LLM) based agents through instruction tuning, also known as agent training. The authors emphasize the crucial role of planning ability in LLM-based agents, which involves interacting with the environment and executing actions to achieve a desired goal from an initial state. Recent studies have shown that utilizing expert-level trajectory for instruction-tuning LLMs effectively enhances their planning capabilities. However, existing work has primarily focused on synthesizing trajectories from manually designed planning tasks and environments, leading to limitations in generating varied and extensive trajectories. To address this challenge, the paper introduces a framework called , which leverages LLMs to first generate diverse environments and then create planning tasks based on these environments. To improve environmental diversity, the authors propose using an inspiration corpus consisting of various domain-specific text segments as context for synthesizing environments. Additionally, to increase the difficulty diversity of generated planning tasks, they introduce a bidirectional evolution method called . This method evolves planning tasks from easier and harder directions to synthesize a task set with a smoother difficulty curve. Evaluation results from  demonstrate that  significantly enhances LLMs' planning ability. For instance, the  instruction-tuned Llama-3 8B model outperforms GPT-3.5 in overall performance and even surpasses GPT-4 in certain tasks. This research contributes to advancing the field of large language models by automating the synthesis of diverse environments and a range of planning tasks, ultimately improving the planning abilities of LLM-based agents.