AgentGen - A framework developed by Microsoft in collaboration with the University of Hong Kong to enhance the planning capabilities of large AI models.

What is AgentGen?

AgentGen, an AI project framework jointly launched by the University of Hong Kong and Microsoft, significantly enhances the planning capabilities of Large Language Models (LLMs) by automatically generating diverse environments and tasks. Utilizing the BI-EVOL method, AgentGen creates task sets with increasing difficulty, enabling 8-parameter Llama-3 models to approach GPT-4 performance on certain tasks. AgentGen performs exceptionally well in both in-domain and out-of-domain task evaluations and holds promise for future applications in complex planning tasks such as robot control and smart homes.

AgentGen's main functions

• Adaptive environment buildingThe framework can autonomously conceive and create diverse virtual environments, providing rich interactive scenarios for intelligent agents.
• Intelligent task customizationBased on advanced language models, AgentGen can intelligently customize tasks and generate adaptive environments.
• Dynamic difficulty adjustmentThrough the innovative BI-EVOL strategy, the framework can dynamically adjust the task difficulty, promoting the agent's learning and growth in tasks of varying complexity.
• No need for a large amount of training dataAgentGen's zero-shot generation capability reduces the need for large-scale training datasets and speeds up the model training process.
• Precision Skill EnhancementBy fine-tuning instructions, the framework can precisely enhance the agent's skills in specific tasks and improve its problem-solving abilities.
• Comprehensive performance monitoringAgentGen achieves optimal performance across various tasks through a meticulous performance evaluation system.

AgentGen's technical principles

• Environment generationAgentGen uses Large Language Models (LLMs) to generate environment specifications, including the definitions of state space, action space, and transition functions. Then, it implements the concrete code for these environments using code generation techniques.
• Heuristic rules and corpusIn the environment generation process, AgentGen uses heuristic rules and diverse corpora to guide LLM in generating diverse environments.
• Task generationBased on the generated environment, AgentGen further uses LLM to generate corresponding planning tasks, ensuring that the tasks match the environment.
• BI-EVOL methodAgentGen employs a bidirectional evolutionary (BI-EVOL) approach to adjust task difficulty, including easy-evol (simplifying tasks) and hard-evol (complexing tasks), forming a set of tasks with increasing difficulty.
• Zero-shot learningIn the initial stage of task generation, AgentGen prompts the LLM to generate a set of initial planning tasks through zero-shot learning.
• Command fine-tuningAgentGen uses instruction fine-tuning technology to fine-tune LLM using synthetic trajectory data (action-observation pair sequences), thereby improving its planning capabilities.
• Performance evaluationAgentGen evaluates the model's performance on planning tasks using metrics such as success rate and progress rate, ensuring the effectiveness of training.

AgentGen's project address

Application scenarios of AgentGen

• Robot controlIn the fields of automation and intelligent manufacturing, AgentGen can be used to improve the autonomous planning and decision-making capabilities of robots, enabling them to perform complex tasks more effectively.
• Smart home systemAgentGen can be integrated into smart home systems, helping the system automatically adjust device settings based on user behavior and preferences, improving living comfort and energy efficiency.
• Personal AssistantAs a personal assistant, AgentGen can help users schedule appointments, plan tasks, and set reminders, improving their life and work efficiency.
• Transportation planningAgentGen can be used to optimize transportation systems, helping to plan more rational driving routes, reduce congestion, and improve travel efficiency.
• Game AIIn video game development, AgentGen can be used to generate more intelligent non-player characters (NPCs), providing a richer gaming experience.