8B iOS agent beats Gemini-3.5-Flash on UINavBench and matches Gemini-3.1-Pro

A team of researchers has proposed training small UI-navigation models not on recorded demonstrations but through live interaction with real iOS apps. The resulting 8B-parameter agent beat Gemini-3.5-Flash and Qwen3.8-27B on UINavBench, a benchmark of real iOS apps where a VLM judge verifies success, and tied with Gemini-3.1-Pro. On AndroidWorld, a different operating system the agent never saw during training, its success rate is almost seven times higher than that of the base model after SFT.

Why split the agent into a planner and a "grounder"

The problem the work starts from will be familiar to anyone who has tried to build UI agents. Frontier models handle screen navigation, but they can cost several dollars per task, which is too expensive for everyday use. Small models trained with SFT or offline RL memorize the trajectories in their training data. Once an app shows an unexpected screen, such an agent doesn't know how to get back on track.

The authors split the system into two parts. The planner looks at a screenshot, keeps a history of previous steps, and uses chain-of-thought reasoning to produce the current step and a long-term plan. A separate grounding model turns that step into a concrete action and tap coordinates. Grounding is a nearly deterministic task, and plain SFT handles it well. Expensive RL is therefore spent only on the planner. The planner starts from Qwen3-VL-8B-Instruct, fine-tuned on human trajectories, including the public GUIOdyssey dataset for scenarios that switch between apps. The planner also has a clarification action: if a task is underspecified, it can ask the user about their preferences.

Per-step rewards and critic stabilization

Training is built around an LLM judge that produces two signals. A dense reward scores every step: does it move closer to the goal, and is it free of hallucinations and pointless loops. A terminal reward checks whether the whole task was completed. The dense signal matters where final success is rare: the agent gets feedback even in failed episodes and learns to climb out of unexpected states.

Optimization uses PPO with generalized advantage estimation (GAE). Here the authors describe a familiar problem: when steps are rewarded, the agent starts to drag trajectories out. To counter this, they apply asymmetric discounting: local process rewards are discounted, while the task-completion reward is not.

The second novelty is GiGPPO. The method estimates a state's value using rewards from visually similar states in parallel trajectories. This stabilizes the critic. The authors stress the difference from GiGPO, which computes its baseline over exactly repeated states; in their environment, such states are almost always unique. So the signal is routed through a learned value function. According to the authors, both the dense reward and GiGPPO speed up convergence, and the result beats both SFT and RL with a terminal reward alone.

Infrastructure: 512 virtual iPhones

Online RL requires a huge number of real interactions. The authors built a pool of cloud-hosted virtual iOS devices, controlled over VNC through a set of actions such as tap, swipe, type, long press, home, and wait. Each device is configured once: apps are installed, Apple ID is signed in, and a snapshot is taken. Each episode then ends with a rollback to the snapshot rather than recreating the session, and a reset takes about 10 seconds. 512 trajectories run at once, each on its own device. The authors claim this is the first large-scale training of UI agents entirely on live iOS devices with real apps.

What to keep in mind

This is a preprint, and the available description gives no exact benchmark success percentages, so "almost seven times" refers to the low SFT baseline on AndroidWorld, not to an absolute bar. UINavBench scoring relies on a VLM judge, and training rewards come from an LLM judge, so judge errors can affect both training and measurement. The comparison with Gemini was run on one specific benchmark of real iOS apps and does not necessarily carry over to every task.

The practical takeaway for teams building mobile agents: separating planning from grounding, plus per-step rewards from a judge, can get frontier-level quality from an 8B model in a narrow domain. The price of this approach is your own farm of virtual devices, and, judging by the paper, that farm is the main condition for scaling.