Core Concepts
Large language models can effectively learn to use external tools through cooperation and interaction with the ConAgents framework.
Abstract
Tool learning empowers large language models (LLMs) to extend their capabilities using external tools.
Existing methods face challenges in addressing complex tasks due to limitations of single LLM agents.
ConAgents framework modularizes tool learning into Grounding, Execution, and Observing agents for superior performance.
Iterative calibration method (IterCali) enables agents to adapt based on feedback, showcasing a 6% improvement over baselines.
Experiments on three datasets demonstrate the efficiency and consistency of the ConAgents framework.
Stats
ConAgents demonstrates a 6 point improvement over the SOTA baseline.
Quotes
"Tool learning empowers large language models as agents to use external tools to extend their capability."
"ConAgents framework modularizes the workflow of tool learning into Grounding, Execution, and Observing agents."