MobileGym wants to be the training ground phone agents have been missing
If you've tried to train an agent to use a phone, you've hit the same wall everyone hits. Real devices are slow, flaky, and impossible to run a thousand of in parallel. App state drifts, a login expires, a popup appears, and your reward signal turns to noise. You can't do reinforcement learning on a substrate you can't reset and can't trust. MobileGym, a new paper with eleven authors, is built to fix exactly that.
It's a simulation platform for mobile GUI agents with two properties that matter more than they sound. Verifiable, meaning a task either provably succeeded or it didn't, so the reward isn't a guess. And highly parallel, meaning you can spin up huge numbers of environments at once instead of babysitting physical handsets. Together those are the two things that make serious RL on phone agents actually possible rather than a research demo on five devices.
This is the unglamorous infrastructure layer, and it's exactly the layer that decides who wins. The agents that learned to use computers and browsers got good once people built gyms to train them in, repeatable, resettable, measurable worlds. Phones have lagged precisely because that environment was missing. Real hardware doesn't scale and screenshots-and-prayer doesn't verify.
The phone is the most personal computer most people own, and an agent that can actually operate one, book the appointment, file the expense, navigate the broken checkout flow, is a huge chunk of the 100x promise. None of it trains without a gym. MobileGym is a bet on building the gym before the headline agent.
Paper: arxiv.org/abs/2605.26114
← Back to all articles
It's a simulation platform for mobile GUI agents with two properties that matter more than they sound. Verifiable, meaning a task either provably succeeded or it didn't, so the reward isn't a guess. And highly parallel, meaning you can spin up huge numbers of environments at once instead of babysitting physical handsets. Together those are the two things that make serious RL on phone agents actually possible rather than a research demo on five devices.
This is the unglamorous infrastructure layer, and it's exactly the layer that decides who wins. The agents that learned to use computers and browsers got good once people built gyms to train them in, repeatable, resettable, measurable worlds. Phones have lagged precisely because that environment was missing. Real hardware doesn't scale and screenshots-and-prayer doesn't verify.
The phone is the most personal computer most people own, and an agent that can actually operate one, book the appointment, file the expense, navigate the broken checkout flow, is a huge chunk of the 100x promise. None of it trains without a gym. MobileGym is a bet on building the gym before the headline agent.
Paper: arxiv.org/abs/2605.26114
Comments