Faithful Mobile GUI Agents with Guided Advantage Estimator
Haowen Hu, Pengzhou Cheng, Zheng Wu, Lingzhong Dong, Gongshen Liu, Zhuosheng Zhang · 2026
Read Faithful Mobile GUI Agents with Guided Advantage Estimator by Haowen Hu, Pengzhou Cheng, Zheng Wu, Lingzhong Dong, Gongshen Liu, Zhuosheng Zhang (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-01208.txt.
Description
Haowen Hu and five co-authors present Faithful-Agent, a framework that makes mobile GUI agents act on screen evidence rather than memorized shortcuts. Vision-language GUI agents interact strongly, yet often behave unfaithfully by ignoring displayed evidence or user instructions. Their two-stage pipeline first applies faithfulness-oriented supervised tuning to instill abstention under evidence perturbations. That teaches the agent to hold back when evidence does not support an action.
A second reinforcement finetuning stage uses a guided advantage estimator built on GRPO, with anchor-based and variance-adaptive tempering. That keeps the guided edge from collapsing under sparse rewards. A thought-action consistency reward further aligns reasoning with actions. Trap success rises dramatically over the baseline while general instruction-following performance is preserved.
Catalog record
| CREATOR | Haowen Hu, Pengzhou Cheng, Zheng Wu, Lingzhong Dong, Gongshen Liu, Zhuosheng Zhang |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 78 KB |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2605.01208