# OpenAGI: When LLM Meets Domain Experts

**Authors:** Yingqiang Ge, Wenyue Hua, Jianchao Ji, Juntao Tan, Shuyuan Xu, Yongfeng Zhang  
**arXiv:** [2304.04370](https://arxiv.org/abs/2304.04370)  
**Date:** 2023-04-10  
**Summary page:** https://www.summarizepaper.com/en/paper/2304.04370v1/

## Key points

- Pursuit of Artificial General Intelligence (AGI) has led to development of large, comprehensive intelligent models
- Equipping these models with ability to harness domain-specific expert models for complex task-solving is crucial
- Recent developments in Large Language Models (LLMs) make them promising as controllers to select, synthesize, and execute external models to solve complex tasks
- OpenAGI is an open-source AGI research platform designed to offer complex multi-step tasks accompanied by task-specific datasets, evaluation metrics and a diverse range of extensible models
- OpenAGI formulates complex tasks as natural language queries that serve as input to the LLM which selects, synthesizes and executes models provided by OpenAGI to address the task
- Reinforcement Learning from Task Feedback (RLTF) mechanism uses the task-solving result as feedback to improve the LLM's task-solving ability creating a feedback loop for self-improving AI
- Open sourcing the code benchmark and evaluation methods of OpenAGI on GitHub at https://github.com/agiresearch/OpenAGI facilitates long-term improvement and evaluation of AGI's ability within the community
- This project offers significant potential in advancing AGI research by providing an open source platform that enables researchers worldwide access to tools necessary for developing more advanced AI systems capable of solving increasingly complex problems.

## Summary

The pursuit of Artificial General Intelligence (AGI) has led to the development of large, comprehensive intelligent models. However, it is equally crucial to equip these models with the ability to harness various domain-specific expert models for complex task-solving. Recent developments in Large Language Models (LLMs) have demonstrated remarkable learning and reasoning abilities, making them promising as a controller to select, synthesize, and execute external models to solve complex tasks. In this project, a team of researchers developed OpenAGI - an open-source AGI research platform specifically designed to offer complex multi-step tasks accompanied by task-specific datasets, evaluation metrics and a diverse range of extensible models. OpenAGI formulates complex tasks as natural language queries that serve as input to the LLM. The LLM subsequently selects, synthesizes and executes models provided by OpenAGI to address the task. This approach allows LLMs to operate various expert models for complex task-solving which is a promising approach towards AGI. Furthermore, the researchers proposed a Reinforcement Learning from Task Feedback (RLTF) mechanism that uses the task-solving result as feedback to improve the LLM's task-solving ability. Thus creating a feedback loop for self-improving AI where the LLM synthesizes various external models for solving complex tasks while RLTF provides feedback to improve its task-solving ability. To facilitate long-term improvement and evaluation of AGI's ability within the community, they open sourced the code benchmark and evaluation methods of OpenAGI on GitHub at https://github.com/agiresearch/OpenAGI . Overall this project offers significant potential in advancing AGI research by providing an open source platform that enables researchers worldwide access to tools necessary for developing more advanced AI systems capable of solving increasingly complex problems.