The researchers of Meta have developed Husky, an open source linguistic agent that, instead of focusing on a single task, has been designed to perform complex reasoning tasks with different tools.
Linguistic agents are characterized, in general, by the use of proprietary large language models (LLM) or because they are aimed at performing tasks within a specific domain. In this scenario, researchers from Meta, along with others from the University of Washington and the Allen Institute for Artificial Intelligence, have presented a different approach.
Their proposal is called Husky and is an open source linguistic agent capable of “reasoning about a unified action space to address a diverse set of complex tasks that involve numerical, tabular and knowledge-based reasoning,” as indicated in the research text published on Arxiv.org.
To do this, it uses a two-part procedure: the generation of the action, which predicts the action, that is, “the highest level step to be taken and the tool that executes it”; and action execution, whereby the model and tool perform the action and “update the state of the solution” with a “predefined ontology of actions” until it adopts the final state.
In a clearer way, Husky breaks down each task into a series of actions, and in each of them it uses a tool to carry it out until it completes the task, or reaches the final state.
The researchers have highlighted that it uses large language models (LLM) with 7 billion parameters, but that its performance “equals or exceeds” that of other frontier language models such as GPT-4 in the tasks that are performed. have analyzed.
“Our work presents a solid recipe for creating open-source linguistic agents that generalize different types of multi-step reasoning tasks,” the researchers conclude.