Speculative Rollback Correction for Quality-Diverse Web Agent Imitation

The article introduces Speculative Rollback Correction (SRC), a novel imitation learning framework designed for training interactive web agents in resettable environments. SRC employs a fixed-horizon branch review strategy that allows agents to execute speculative actions before expert intervention, enabling the identification and correction of harmful deviations while preserving successful trajectory prefixes. The approach was evaluated on the WebArena-Infinity benchmark, resulting in the collection of 977 verifier-passing trajectories and 9,183 next-action examples, demonstrating improved recovery efficiency compared to traditional step-level reviews, which is crucial for practitioners aiming to enhance agent performance and robustness in dynamic environments.