While advances in large-scale language models have improved web agents, more advanced planning and search capabilities are needed to handle complex and dynamic web environments. In this study, we enhance web agents with an explicit rollback mechanism, allowing agents to revert to previous states during their exploration trajectories. This mechanism provides the model with the flexibility to directly control the search process, resulting in an effective and efficient web exploration method. Experiments are conducted on two real-time web exploration benchmarks, using both zero-shot and fine-tuning settings, demonstrating the effectiveness of the proposed approach.
