Large Language Models (LLMs) have revolutionized natural language processing tasks, and their potential in physical world planning tasks is beginning to be leveraged. However, these models often encounter problems in understanding the actual world, resulting in hallucinatory actions and a reliance on trial-and-error behavior. Researchers have noted that humans perform tasks efficiently by leveraging global task knowledge and local state knowledge, thus avoiding blind trial-and-error actions.
Many approaches within LLM-based agent systems such as agent planning, external tool utilization, and code generation, often result in trial-and-error actions. Moreover, present approaches like knowledge-augmented agent planning also run into issues when trying to transfer tasks.
In an effort to optimize agent planning, researchers from the Zhejiang University – Ant Group Joint Laboratory of Knowledge Graph, the National University of Singapore, and Alibaba Group have constructed a parametric World Knowledge Model (WKM). The WKM incorporates expert and explored trajectories of tasks and integrates this knowledge into expert trajectories to aid in the efficient execution of tasks.
The agent model compares expert and sampled trajectories to self-synthesize task knowledge. By using state knowledge accumulated from expert trajectories at each planning step, the model can form a state knowledge base that ensures effective and accurate agent planning.
The researchers used unseen tasks on the ALFWorld, WebShop, and ScienceWorld datasets to evaluate the method. The performance surpasses GPT-4 on ALFWorld and WebShop, showing that the integration of world knowledge outperforms simple fine-tuning. The experiments proved the WKM’s ability to reduce trial-and-error behavior, improve unseen task generalization, and extend unified world knowledge training.
In conclusion, the parametric WKM approach has demonstrated superior performance in planning language agent models. By providing global planning task knowledge and local state planning knowledge, the model reduces trial-and-error actions considerably. It also exhibits improved performance on state-of-the-art models, indicating the model’s high efficiency and accuracy. As a result, the development of the WKM serves as a milestone in refining agent models for planning tasks.
