Researchers at Meta have developed a versatile and powerful linguistic agent called Husky that is capable of performing complex reasoning tasks using a variety of tools. Unlike other linguistic agents that focus on specific domains or tasks, Husky is designed to handle a diverse set of tasks that require numerical, tabular, and knowledge-based reasoning.
In a recent research publication on Arxiv.org, researchers from Meta, the University of Washington, and the Allen Institute for Artificial Intelligence introduced Husky as an open source linguistic agent that can reason about a unified action space. This means that Husky is capable of breaking down tasks into a series of actions and using tools to complete each action until the task is finished or a final state is reached.
Husky’s approach involves a two-part procedure: action generation and action execution. The model predicts the action to be taken and the tool to execute it, and then performs the action while updating the state of the solution based on a predefined ontology of actions. Despite using large language models with 7 billion parameters, Husky’s performance matches or surpasses that of other advanced language models like GPT-4 in the tasks it has been tested on.
The researchers believe that Husky’s ability to generalize different types of multi-step reasoning tasks offers a promising recipe for creating open-source linguistic agents with broad capabilities. By focusing on complex reasoning tasks and using a diverse set of tools, Husky stands out as an innovative solution in the field of artificial intelligence for performing complex reasoning tasks with ease.
