MobileIPL: Enhancing Mobile Agents Thinking Process via Iterative Preference Learning

TL;DR

This paper introduces MobileIPL, a novel iterative preference learning approach that enhances mobile agent reasoning by constructing CoaT-trees, leveraging GPT-4o for diverse data, and achieving state-of-the-art results on GUI benchmarks.

Contribution

The paper proposes a new iterative preference learning method with a CoaT-tree construction, rule-based scoring, and a three-stage instruction evolution to improve mobile agent reasoning and generalization.

Findings

MobileIPL outperforms strong baselines on three GUI benchmarks.

Achieves state-of-the-art performance across multiple Mobile GUI-Agents benchmarks.

Demonstrates strong generalization to out-of-domain scenarios.

Abstract

The Chain of Action-Planning Thoughts (CoaT) paradigm has been shown to improve the reasoning performance of VLM-based mobile agents in GUI tasks. However, the scarcity of diverse CoaT trajectories limits the expressiveness and generalization ability of such agents. While self-training is commonly employed to address data scarcity, existing approaches either overlook the correctness of intermediate reasoning steps or depend on expensive process-level annotations to construct process reward models (PRM). To address the above problems, we propose an Iterative Preference Learning (IPL) that constructs a CoaT-tree through interative sampling, scores leaf nodes using rule-based reward, and backpropagates feedback to derive Thinking-level Direct Preference Optimization (T-DPO) pairs. To prevent overfitting during warm-up supervised fine-tuning, we further introduce a three-stage instruction evolution, which leverages GPT-4o to generate diverse Q\&A pairs based on real mobile UI screenshots, enhancing both generality and layout understanding. Experiments on three standard Mobile GUI-agent benchmarks demonstrate that our agent MobileIPL outperforms strong baselines, including continual pretraining models such as OS-ATLAS and UI-TARS. It achieves state-of-the-art performance across three standard Mobile GUI-Agents benchmarks and shows strong generalization to out-of-domain scenarios.