WebUncertainty Framework Introduces Dual-Level Planning for Autonomous Web Agents

A novel framework for autonomous web agents, named WebUncertainty, has been introduced to overcome challenges associated with executing complex tasks. This system addresses two levels of uncertainty in planning and reasoning for agents functioning on real-world webpages. Thanks to recent progress in large language models, these agents can now interpret natural language instructions directly from the internet. Traditional agents often falter in dynamic scenarios and long-term task execution due to inflexible planning methods and reasoning that is prone to hallucinations. WebUncertainty features a Task Uncertainty-Driven Adaptive Planning Mechanism, which dynamically selects planning strategies for unfamiliar environments. Additionally, it employs an Action Uncertainty-Driven Monte Carlo tree search (MCTS) Reasoning Mechanism, utilizing the Confidence-induced Action Uncertainty (ConActU) strategy to measure both aleatoric uncertainty (AU) and epistemic uncertainty (EU), aiming to enhance the performance of autonomous web agents in complex, interactive tasks.