Long web tasks require an agent to maintain a plan while acting on changing screens. The JHiMA-WebRL paper separates those responsibilities into a task decomposer and an actor, then trains both together rather than leaving the planner fixed.
From demonstrations to coordinated roles
A strong teacher model constructs structured subtask supervision from multimodal GUI trajectories. Those annotations are used to fine-tune the decomposer and the actor conditioned on its subtasks.
The framework then extends Group Relative Policy Optimization with hierarchical rollout grouping and role-specific advantages. In practical terms, training can improve the plan and the execution policy together, instead of asking an increasingly capable actor to compensate for an unchanged planner.
The reported comparison
JHiMA-WebRL-14B combines a 7B actor and a 7B decomposer. On WebArena-Lite-v2 it achieves an average success rate of 37.2%, compared with 24% for a ScaleCUA-32B agent fine-tuned with reinforcement learning on the same training tasks and evaluated under the same protocol.
The shared protocol makes that comparison informative, but it is still a benchmark result. It does not show that two smaller models will outperform a larger model on every website or task.
Why joint updates matter
The ablations find that updating the decomposer during reinforcement learning performs better than updating only the actor. This supports the paper's central question about learning the division of labor, rather than attributing the outcome solely to a larger total parameter count.
Implications for agent design
The useful takeaway is to evaluate coordination as part of the system. For a real deployment, an assessment should distinguish planning mistakes, action failures and loss of progress across a long history. The study offers a concrete jointly trained architecture to investigate, while leaving reliability on a particular live workflow to further validation.
Comments