Apple trains an 8-billion-parameter phone agent on live iOS apps

Overview

Updated 1 hour ago

Most phone-navigation agents learn by watching recorded demonstrations. Apple's researchers trained one by tapping through hundreds of virtual iPhones at once, scoring each step, and the resulting 8-billion-parameter model matched Gemini-3.1-Pro on a benchmark built from 36 real iOS apps.

On UINavBench, Apple's PPO-trained agent reached 32.4%, beating Gemini-3.5-Flash at 28.4%. On AndroidWorld, an environment it never trained on, it scored 32.8% against 4.7% for the supervised baseline. Frontier models still beat it there: Gemini-3.1-Pro scores 60.8%.

Why it matters

If live-app training works, small models can handle phone tasks locally, cutting the inference cost that keeps frontier agents out of daily use.

Questions about this story

Free account needed to ask — your question is kept and asked for you right after sign-up. Answers are public.

No questions yet — be the first to ask.

Key Indicators

32.4%

UINavBench success rate, PPO step 336

Ties Gemini-3.1-Pro and beats Gemini-3.5-Flash's 28.4% on real iOS apps.

32.8%

AndroidWorld success rate, trained on iOS only

Nearly seven times the 4.7% supervised baseline on an untrained platform.

4.7%

AndroidWorld success rate, SFT baseline

The comparison Apple's sevenfold claim rests on, and the number critics question.

Agent parameter count, in billions

Roughly a third the size of Qwen3.8-27B, which it outperforms on UINavBench.

116

UINavBench tasks

Spread across 36 real iOS apps staged with user data.

Voices

Curated perspectives — historical figures and your fellow readers.

Organizations Involved

Apple

Consumer hardware and software company

Published the framework and runs the research group behind it

Makes the iPhone these agents learn to operate, and staged the real iOS apps used in the benchmark.

Google DeepMind

AI research lab

Its AndroidWorld benchmark and Gemini models are the paper's main yardsticks

Built one of the three benchmarks Apple reports and supplies two of the model comparisons.

Timeline

2023 November 2026

3 events Latest: Today

November 2026
Apple reports online reinforcement learning results on live iOS apps
Today Research

PPO at step 336 ties Gemini-3.1-Pro at 32.4% on UINavBench. A neighbor-conditioned variant, GiGPPO, leads on the synthetic iOSWorld benchmark at 9.8%.
Today
2025
UI-TARS shows Android-specific training can pay off
Research

The open agent reached 46.6% on AndroidWorld after training specifically on Android data, while scoring 11.0% on the iOS benchmark.
2025
2023
Public Android interface data piles up, iOS data does not
Research

The paper later cites large public Android datasets as a reason frontier models score higher there. No comparable public iOS corpus exists.
2023

Scenarios

Apple ships a system-level iOS agent that acts across apps

Possible Resolves by End of 2027

Discussed by: Commentators at AI Weekly and explainx.ai both note the paper stops at benchmarks, with no product commitment announced.

The training recipe exists; the missing piece is a shipping feature. If Apple folds a planning agent into Siri or a system API, the live-app training loop becomes a product and the benchmark numbers start mattering to ordinary users. If the planners stay in the lab, the result remains a capability demonstration.

An independent lab reproduces the dense-reward gain on a public benchmark

Possible Resolves by End of 2027

Discussed by: The paper itself argues dense step-level rewards beat terminal-only rewards at equivalent training steps; no outside group has yet confirmed it on a shared benchmark.

Apple's dense-reward result rests on an LLM-as-a-Judge scoring intermediate steps. If an unaffiliated team applies the same reward structure to AndroidWorld and reports a similar edge over terminal-only training, the method becomes a general recipe rather than one lab's result.

The sevenfold transfer claim gets traced to a weak baseline

Possible Resolves by Q2 2027

Discussed by: explainx.ai flagged that the sevenfold figure may come from a low supervised baseline, and that the Gemini comparison depends on prompts and how those models were driven.

A 4.7% baseline is easy to beat. A follow-up ablation showing most of the AndroidWorld gap comes from baseline weakness, rather than online reinforcement learning, would shrink the headline without invalidating the UINavBench tie.

Historical Context

2 moments from history that rhyme with this story — and how they unfolded.

October 2017

AlphaGo Zero (2017)

DeepMind built a Go system that learned from self-play alone, with no human game records, and it beat the earlier AlphaGo that had defeated Lee Sedol 100 games to 0.

Then

Demonstrated that imitating human play was a ceiling, not a starting advantage.

Now

Self-play and reinforcement learning became the default approach for game-playing systems.

Why this matters now

Apple makes the same argument about recorded taps. Supervised training on static trajectories plateaus; step-level reward lets a model recover when the screen does something unexpected.

June 2018

OpenAI Five (2018)

OpenAI trained five neural agents to play Dota 2 from scratch, running thousands of parallel game instances and combining hand-designed sub-goals into a single reward signal.

Then

Beat experienced amateur players at The International in August 2018, then lost to professionals in 2019.

Now

Established parallel simulation plus shaped rewards as the standard infrastructure for training agents that act in interactive environments.

Why this matters now

The same two ingredients appear in Apple's setup: dense rewards from a judge, and hundreds of concurrent virtual iOS devices to generate fresh experience.