8B-Parameter Agent Operates iOS Apps at Gemini 3.1 Pro Level

A team of researchers has shown that a small 8B-parameter model can be taught to use real iOS apps with confidence, and to do it more cheaply than flagship neural networks. The agent was not trained on recorded examples. It learned by trial and error directly on live virtual iPhones. On UINavBench, a test with real iOS apps, it matched Gemini 3.1 Pro and beat Gemini 3.5 Flash and Qwen3.8-27B. The tasks are along the lines of "find a pair of sneakers and add them to the cart." The agent has to tap buttons, scroll through screens, and not get lost on its own.

Why small interface agents usually break

Agents that look at a screenshot and tap the screen themselves are attracting growing attention. Flagship models handle this, but by the authors' estimate, a single task often costs several dollars. That is too expensive for everyday use.

Hence the demand for small models. But the typical ways of training them have weak points:

• Fine-tuning on demonstrations (SFT). The model copies recorded human routes. Those recordings contain almost no mistakes, so the model does not know what a failure looks like or what to do after one. When it strays from the familiar path, the agent gets lost.
• Reinforcement learning with a reward only at the end. The agent gets a signal only when the task is finished. If successes are rare, there is almost no signal. Steps that genuinely moved the agent closer to the goal earn nothing if the task fails anyway.
• Single-step reinforcement learning. The model is rewarded for resembling a reference chain of actions, and the problem is the same as with SFT.

Planner, executor, and judge

The authors split the agent into two parts. The planner receives the instruction, a screenshot, the history of previous steps, and the remainder of the plan. It reasons through a chain of thought and outputs the current step and a plan for what comes next. Its available actions include asking the user a question when a task is vaguely worded, for example, which sneakers to look for.

The grounding model takes the text of a step and turns it into a concrete action with touch coordinates. This is a nearly mechanical mapping from intent to a point on the screen, and standard fine-tuning handles it well. The expensive reinforcement learning is reserved for the planner, which needs reasoning and the ability to improvise its way out of trouble.

Rewards are handled by another language model acting as judge. It produces two signals:

• a dense reward for each step: did it move the agent closer to the goal, and was it free of fabrications and useless loops;
• a final reward for completing the whole task.

Training uses the PPO algorithm with step-level advantage estimation (GAE), which computes how much better or worse a given step turned out than expected. To keep the agent from dragging out a session to collect rewards for intermediate steps, the authors introduced asymmetric discounting. Step rewards lose value over time, while the final reward for success does not.

A separate contribution is called GiGPPO. PPO includes a critic, a model that estimates how good the current state is. The critic is unstable, so it is fed signals from visually similar screens across parallel runs. The related GiGPO method averages results only over exactly matching states. In live apps such matches almost never occur, and the groups collapse into singletons.

Hundreds of virtual iPhones and the results

Training on live devices requires infrastructure. The authors use cloud-hosted virtual iOS devices controlled over VNC: tap, swipe, text input, long press. Each device is set up once: apps are installed, Apple ID is signed in, and a snapshot of the virtual machine is saved. After an episode, the system is rolled back to the snapshot rather than restarted, which takes about 10 seconds. During training, 512 trajectories run at the same time, each on its own device. According to the authors, this is the first training of interface agents at such a scale entirely on live iOS devices with real apps.

The starting point is Qwen3-VL-8B-Instruct, fine-tuned on human recordings. The authors used two of their own datasets (Apple apps and third-party apps) and the public GUIOdyssey. It contains Android apps, but it includes tasks that switch between several apps, which the iOS data lacks.

What the measurements on UINavBench, iOSWorld, and AndroidWorld showed:

• the agent clearly beat both baselines: fine-tuning on demonstrations and reinforcement learning with a reward only at the end;
• dense rewards and GiGPPO speed up training convergence;
• on AndroidWorld, where the agent was not trained on iOS, the success rate is nearly seven times higher than after standard fine-tuning;
• on UINavBench, the agent outperformed Qwen3.8-27B and Gemini 3.5 Flash and matched Gemini 3.1 Pro.

What this means in practice

The sevenfold gain on Android suggests that the agent learned general navigation skills rather than the routes of specific iOS apps. For companies building mobile assistants, this is a signal: a compact model, potentially suitable for on-device use, can approach a flagship if it is trained in live environments with step-by-step feedback.

There are caveats. The results come from benchmarks, and the judge used in training is also a language model, so the quality of the rewards depends on it. A sizable fleet of virtual devices is also required. But the recipe is clear: split planning from grounding, reward progress rather than only the final outcome, and train on real apps.