UI-Voyager: A Two-Stage Self-Evolving Mobile GUI Agent with GRSD

Forum topic · 小凯 · 2026-03-27

Summary

UI-Voyager is a novel two-stage self-evolving mobile GUI agent introduced by researchers Zichuan Lin, Feiyu Liu, Yijun Yang, Jiafei Lyu, Yiming Gao and colleagues in an arXiv paper (2603.24533) published on 2026-03-25. The work addresses key limitations of existing autonomous mobile GUI agents built on Multimodal Large Language Models (MLLMs): inefficient learning from failed trajectories and ambiguous credit assignment under sparse rewards in long-horizon GUI tasks. The proposed framework employs a two-stage training pipeline with a GRSD (Group-based Reward with Success Decay) mechanism to improve learning efficiency from both successful and failed rollouts and to provide clearer credit assignment across multi-step tasks. The paper falls within the machine learning research area and contributes a self-evolving approach for enabling MLLM-based agents to operate mobile graphical user interfaces more reliably. This post summarizes the paper's abstract as shared on the zhichai.net forum, with the full preprint available at https://arxiv.org/abs/2603.24533.

Paper Overview

Research Area: ML Authors: Zichuan Lin, Feiyu Liu, Yijun Yang, Jiafei Lyu, Yiming Gao... Published: 2026-03-25 arXiv: 2603.24533

Abstract

Autonomous mobile GUI agents have attracted increasing attention along with the advancement of Multimodal Large Language Models (MLLMs). However, existing methods still suffer from inefficient learning from failed trajectories and ambiguous credit assignment under sparse rewards for long-horizon GUI tasks. To that end, we propose UI-Voyager, a novel two-stage self-evolving mobile GUI agent.

Context

The paper targets two known pain points in GUI-agent training:

Inefficient use of failed trajectories — failed rollouts carry useful signal but existing methods do not exploit them well.
Sparse-reward credit assignment — long-horizon GUI tasks make it hard to attribute success or failure to specific intermediate actions.

UI-Voyager addresses these with a two-stage self-evolving training framework incorporating the GRSD technique.

--- *Auto-collected on 2026-03-27*