Online RL Framework Enhances Robustness of iOS UI Navigation Agents

1 reports archived1 independent sourcesupdated 11/2/2026, 08:00:00

Synthesis & Latest Updates

1 Sources Cross-Validated

To address high inference costs of frontier VLMs and brittleness of SFT-trained small models, this work proposes an online reinforcement learning framework for training UI navigation agents directly on live iOS apps. By decoupling the agent into Planning and Grounding components and utilizing LLM-as-a-Judge for dense rewards, along with a novel GiGPPO value estimation method, the approach significantly improves error recovery and robustness in dynamic interfaces.

LATEST/To address high inference costs of frontier VLMs and brittleness of SFT-trained small models, this work proposes an online reinforcement learning framework for training UI navigation agents directly on live iOS apps. By decoupling the agent into Planning and Grounding components and utilizing LLM-as-a-Judge for dense rewards, along with a novel GiGPPO value estimation method, the approach significantly improves error recovery and robustness in dynamic interfaces.

Hourly heat curve

2 fully observed hours

Fewer than 3 fully observed hours; trend not plotted yet.

Coverage timeline

Total 1 reports · Latest first

Apple Machine Learning ResearchT1Nov 2, 08:00·78 pts
Online RL Framework Enhances Robustness of iOS UI Navigation Agents
- Proposes an online RL framework to train UI navigation agents directly on live iOS apps, preventing overfitting from offline data.
- Utilizes LLM-as-a-Judge to generate dense reward signals combining step-level feedback and terminal rewards for policy optimization.
- Introduces GiGPPO, a neighbor-conditioned value estimation method that enhances critic stability by leveraging rewards from visually similar states across concurrent trajectories.
View Related Reports & Analysis Original Source