Research map / Multi-agent systems and games / LLM agents

MobileIPL: Enhancing Mobile Agents Thinking Process via Iterative Preference Learning

ICLR 2026 · Multi-agent systems and games › LLM agents · approach: Self-evolving language agent learning · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes MobileIPL, a self-training mobile GUI agent that builds a CoaT thinking tree via iterative sampling, assigns rule-based rewards with backward propagation to form thinking-level DPO pairs, and uses GPT-4o-based three-level instruction evolution for warm-up SFT.

Problem

Mobile GUI agents using chain-of-action-planning thoughts (CoaT) suffer from scarce, diverse trajectories and overfitting after SFT; existing self-training either ignores intermediate reasoning correctness or requires costly process-level annotations for PRMs.

Approach

Construct a CoaT-tree by iterative multi-turn sampling, score leaf actions with a rule-based reward (type, format, click distance, input F1), propagate averaged values backward to intermediate thoughts, and derive thinking-level DPO pairs without a process reward model; add instruction evolution of GUI Q&A from screenshots to increase diversity.

Key results

• On AITZ, MobileIPL reaches 69.15% match accuracy versus 60.36% for the Qwen2-VL-7B CoaT baseline and 65.61% for UI-TARS-7B.
• On AMEX, MobileIPL achieves 74.29% overall, versus 70.33% for OS-Atlas-7B and UI-TARS-7B.
• On AndroidControl, MobileIPL reaches 72.7% Step.Acc and 70.0/72.2 on app-unseen/task-unseen OOD subsets, outperforming GRPO on all subsets.

Surprising findings

• Naive DPO on whole trajectories lowers performance relative to IPL (60.3% vs 65.4%), and DPO on unstable annotation preferences performs worse than SFT on AITZ SCROLL.
• Better SFT seed performance does not guarantee a higher IPL upper bound because sampling diversity drops as training fits the data.

Limitations

• Tree sampling and reward design depend on ground-truth golden actions, and action equivalence can mislabel valid alternatives as negatives.
• Performance on rare action types such as PRESS fluctuates strongly across IPL iterations.

Experimental setup

Datasets: AITZ AMEX AndroidControl Android in the Zoo (screenshots for instruction evolution)

Benchmarks: AITZ AMEX AndroidControl

Metrics: Step.Acc action type accuracy action match accuracy

Baselines: CogAgent AUTO-GUI SeeClick UGround OS-Atlas UI-TARS Falcon-UI GRPO SPO-Chain

Builds on: DPO ReST-MCTS ReFT CoaT Qwen2-VL

Closest papers by meaning

• Scaling, Benchmarking, and Reasoning of Vision-Language Agents for Mobile GUI Navigation (ICML 2026 · similarity 0.87)
• GUI-G1: Understanding R1-Zero-Like Training for Visual Grounding in GUI Agents (NeurIPS 2025 · similarity 0.86)
• SE-GUI: Enhancing Visual Grounding for GUI Agents via Self-Evolutionary Reinforcement Learning (NeurIPS 2025 · similarity 0.86)
• FingerTip 20K: A Benchmark for Proactive and Personalized Mobile LLM Agents (ICLR 2026 · similarity 0.86)
• UI-Genie: A Self-Improving Approach for Iteratively Boosting MLLM-based Mobile GUI Agents (NeurIPS 2025 · similarity 0.86)
• MIRAGE: Mobile Agents with Implicit Reasoning and Generative World Models (NeurIPS 2026 · similarity 0.86)