Online dense RL trains UI agents on live iOS apps
A paper by Peitong Duan, Khashayar Gatmiry, Jeffrey Nichols, Samy Bengio, and Emmanuel Abbe introduces an online dense reinforcement learning framework for training UI navigation agents directly on live iOS applications. It separates planning from UI grounding, combines step-level and terminal rewards through an LLM-as-a-Judge, and introduces GiGPPO alongside infrastructure for orchestrating hundreds of concurrent virtual iOS devices. The paper reports that its 8B agents outperformed SFT and terminal-only RL baselines, nearly achieved a sevenfold improvement in out-of-distribution Android tests, and surpassed Gemini-3.5-Flash while tying Gemini-3.1-Pro on real iOS apps.
The work presents a concrete approach to reducing the inference cost of frontier vision-language models while improving smaller agents’ robustness on interactive mobile interfaces.
Computer Science > Artificial Intelligence
arXiv:2610.39264 (cs)
[Submitted on 29 Oct 2026]
Title:Online Dense Reinforcement Learning with Planning-Grounding Agent for UI Navigation on Live IOS Apps
Authors:Peitong Duan, Khashayar Gatmiry, Jeffrey Nichols, Samy Bengio, Emmanuel Abbe
View a PDF of the paper titled Online Dense Reinforcement Learning with Planning-Grounding Agent for UI Navigation on Live IOS Apps, by Peitong Duan and 4 other authors
#GiGPPO#UINavBench#iOSWorld#AndroidWorld#Gemini-3.5-Flash#Gemini-3.1-Pro