The researchers at Meta have developed an open source linguistic agent called Husky that is designed to perform complex reasoning tasks using a variety of tools, rather than focusing on a single task. This sets Husky apart from traditional linguistic agents, which typically rely on large language models or are limited to specific domains. In collaboration with researchers from the University of Washington and the Allen Institute for Artificial Intelligence, Husky has been developed to address a diverse set of tasks that involve numerical, tabular, and knowledge-based reasoning.
Husky’s unique approach involves breaking down each task into a series of actions, using tools to carry out each action until the task is completed or reaches its final state. The two-part procedure involves generating actions and executing them, updating the state of the solution with a predefined ontology of actions. Husky utilizes large language models with 7 billion parameters, and its performance matches or exceeds that of other language models like GPT-4 in the tasks that have been evaluated.
The researchers believe that Husky provides a solid foundation for creating open-source linguistic agents that can generalize across different types of multi-step reasoning tasks. By combining a diverse set of tools and reasoning capabilities, Husky offers a new approach to linguistic agent development that has the potential to advance the field of artificial intelligence.