Adaptive Planning on the Web: Using LLMs and Affordances for Web Agents

We investigate the adaption of agents using plans on the Web despite its large and dynamic nature, as well as agents’ constrained perception. Based on Semantic Web technologies and affordances, we compare how agents choose appropriate actions to adapt to their environment by condition-action rules or suggested actions of large language models. We conduct experiments on execution cost and plan stability distance to see whether agents choose appropriate actions to adapt their plans. We find that cost and stability of rule-based and LLMs for adaptation with affordances are close together, while performance differs greatly.