AdaptRubric Enhances GUI Agents with Task-Adaptive Rewards

Recent advancements in GUI agents often rely on outcome reward modeling, where success is judged by how well a trajectory meets user instructions. However, existing GUI reward verifiers frequently lack task-adaptiveness, either using generic rubrics or implicitly reasoning, which can lead to overlooking specific constraints or enforcing unstated requirements. The AdaptRubric framework addresses this by introducing a Coarse-to-Fine Rubrics approach for constructing highly task-adaptive judging criteria. It begins with a category-level coarse stage, routing the instruction to a GUI task family and retrieving relevant reusable criteria. This is followed by an instance-level fine stage, which generates compact cues for concrete values, scopes, and constraints specific to the current instruction. This method ensures that the judging criteria are precisely tailored to each task instance. AdaptRubric consistently outperforms previous reward agents in both offline reward evaluation and online reinforcement learning optimization, demonstrating significant improvements in F1 scores and overall task success rates.