Refining Guideline Knowledge for Agent Planning Using Textgrad

Runnan Fang, Shuofei Qiao, Zekun Xi

Abstract: Guideline Knowledge is helpful for LLM (Large Language Model) based Agent in embody task planning. In this work, we introduce Textgrad to optimize the Guideline Knowledge for the agent’s embodied tasks. This allows the model to automatically optimize the Guideline Knowledge by calculating the text gradients in the Guideline Knowledge and analyzing the loss in failed trajectories.