A correct click is not enough when the goal spans many pages. The useful insight is that planning needs its own training signal: successful action sequences can be turned into executable plans, rather than assuming an instruction-following model already knows how to plan.
Key Findings
- A dedicated planner sets high-level steps while an executor maps them to concrete actions, making strategy and execution separately trainable.
- Synthetic supervision pairs successful trajectories with feasible plans, then expands those examples to improve generalization beyond a small set of demonstrations.
- The authors report 57.58% success on WebArena-Lite and 81.36% on text-only WebVoyager. These are benchmark-specific results, rather than a guarantee for arbitrary websites.
Who Should Read It
Useful for researchers designing long-horizon agents or turning execution logs into planning supervision. The main investment is collecting credible trajectories and training specialized components. Teams needing immediate browser automation should assess that data and training cost before adopting the method.