核心概念
ConAgents framework enhances tool learning by enabling cooperative and interactive agents to adaptively calibrate themselves, improving task performance.
要約
ConAgents introduces a Cooperative and interactive Agents framework for tool learning tasks. It modularizes the workflow into Grounding, Execution, and Observing agents. The Iterative Calibration (IterCali) method enables adaptive calibration based on feedback from the tool environment. Experimental results show superior performance over baselines in Success Rate and Correct Path Rate metrics.
統計
Experiments demonstrate 6 point improvement over SOTA baseline.
Human evaluation shows superiority in Executability and Correct Rate of Parsing.
ConAgents achieves 13.2% relative improvement in Success Rate.
引用
"We propose the ConAgents, a Cooperative and interactive Agents framework."
"Experiments conducted on three datasets demonstrate the superiority of our ConAgents."
"Our proposed IterCali method can enable the agents to adapt their actions to complete tasks."