跳转至

文章背景与核心概要

近年来，图形用户界面（GUI）智能体的研究高度依赖于结果奖励建模（Outcome Reward Modeling），用于评估执行的轨迹是否满足用户的指令标准。然而，现有的奖励验证器往往缺乏任务自适应能力：它们常将检查规则不恰当地跨任务迁移、忽略特定的指令约束，或者强制施加未说明的要求。

为了解决这些局限性，本文引入了 AdaptRubric，这是一个粗到细的评分标准框架（Coarse-to-Fine Rubrics Framework），通过两个主要阶段动态构建任务自适应的评判标准：1. 类别级粗粒度阶段： 将指令路由至 GUI 任务系列，以检索可复用的任务系列标准。2. 实例级细粒度阶段： 生成紧凑的、特定于实例的提示，以捕获当前指令中的具体数值、范围和约束条件。实验结果表明，AdaptRubric 在离线奖励评估和在线强化学习优化中均持续超越先前的奖励智能体，将 F1 分数提升了 3.6 个百分点，并带来了 4.23 个百分点的任务成功率提升。

Task-Adaptive Rubrics for GUI Reward Modeling

arXiv ID: arXiv:2608.24174 [cs.AI]
Submitted: August 25, 2026
Authors: Tao Xiong, Xavier Hu, Wenkai Wang, Qinzhuo Wu, Changqiao Wu, Pengzhi Gao, Wei Liu, Jian Luan, Shengyu Zhang

📌 Summary

Recent advancements in Graphical User Interface (GUI) agents heavily rely on outcome reward modeling to evaluate whether an executed trajectory meets a user's instruction criteria. However, existing reward verifiers struggle with task-adaptability: they often transfer checks inappropriately across tasks, overlook specific instruction constraints, or enforce unstated requirements.

To resolve these limitations, this paper introduces AdaptRubric, a Coarse-to-Fine Rubrics Framework that dynamically constructs task-adaptive judging criteria through two main stages: 1. Category-Level Coarse Stage: Routes instructions to a GUI task family to retrieve reusable task-family criteria. 2. Instance-Level Fine Stage: Generates compact, instance-specific cues to capture concrete values, scopes, and constraints from the current instruction.

Results: AdaptRubric consistently outperforms prior reward agents in both offline reward evaluation and online reinforcement learning optimization, improving the F1 score by 3.6 points and delivering a 4.23-point task-success gain.

📑 Abstract

Recent studies on GUI agents have increasingly focused on outcome reward modeling, which assigns outcome rewards by judging whether an executed trajectory satisfies the success criteria implied by the user instruction. Existing GUI reward verifiers, however, often under-specify how these criteria should be constructed for each task instance. Whether using generic rubric structures or implicit model reasoning, their judging criteria are not sufficiently task-adaptive: they can transfer checks across tasks, overlook concrete constraints in the current instruction, or become overly strict by enforcing unstated requirements. To address this limitation, we propose AdaptRubric, a Coarse-to-Fine Rubrics Framework that constructs task-adaptive judging criteria through a category-level coarse stage and an instance-level fine stage. AdaptRubric performs category-level coarse rubric retrieval by routing the instruction to a GUI task family and retrieving reusable task-family criteria, then conducts instance-level fine rubric generation to surface compact cues for concrete values, scopes, and constraints in the current instruction. Across offline reward evaluation and online reinforcement learning optimization, AdaptRubric consistently outperforms prior reward agents, improving F1 by 3.6 points under a matched image budget and yielding a 4.23-point task-success gain.