Former DeepSeeker and collaborators release new method for training reliable AI agents: RAGEN
1 min read
Summary
Advancements in large language models (LLMs) mean the development of AI agents that can complete complex, multi-task assignments is closer than ever.
However, for these agents to be effective in real-world, enterprise environments, they require more sophisticated training and a move away from simply learning by rote.
To this end, a group of researchers from organisations including Northwestern University, Microsoft and the University of Washington have developed RAGEN, a system for evaluating and training AI agents that allows them to learn from experience, vital for adapting to uncertainty and ongoing change.
The system has been tested across three tasks: Bandit, Sokoban and Frozen Lake, and the researchers found that the model works “across all 3 tasks, relieves collapse, and attains better reward”.
As well as being made open-source, the researchers have also created an interactive demonstration to illustrate the thought processes behind the agent’s decision-making.