Neural operating systems: the research direction
A neural operating system is a research direction: learn computer behavior and interaction with generative models. It raises questions about what an interface can become when its responses are learned, how state persists during interaction, and how generated experiences might connect to useful computation. A particular renderer or recurrent architecture is one implementation choice within that direction.
From a proposal to an interactive model
In 2023, Gabriele Tolomei, Cesare Campagnano, Fabrizio Silvestri, and Giovanni Trappolini presented Prompt-to-OS. Their proposal places generative models between users and computing services, emphasizing intentions, personalized interaction, and interfaces that can be generated for the task. It is a conceptual account of a new interaction paradigm. Read Prompt-to-OS.
On July 11, 2025, Luke Rivard, Sun Sun, Hongyu Guo, Wenhu Chen, and Yuntian Deng released NeuralOS: Towards Simulating Operating Systems via Neural Generative Models. The original paper formalizes learned OS-interface simulation: screens evolve in response to user input events, with state carried through an interactive sequence. It develops an implementation, training procedure, and data-collection system for that research objective. Read the original version.
This contribution includes the modeling problem, its interaction requirements, and the direction toward adaptive generative interfaces. The implementation combines recurrent state tracking with diffusion rendering, trained on Ubuntu XFCE interaction recordings. Its low-level inputs make it possible to examine whether specific mouse and keyboard events produce the intended visual response. Original formulation and architecture.
The March 12, 2026 revision, published at ICLR 2026, also demonstrates training from synthesized data to simulate an application that was never installed in the recorded environment. Version history · Official conference record.
What part of a computer is being learned?
An interface exposes only part of a computer’s state. A hidden window may contain unsaved work. A background operation may still be running. A file can remain present after its window closes. These examples motivate three separate descriptions when reading a paper:
- Interaction: which user actions the model accepts, and when they affect its output.
- State: which information survives beyond the currently visible screen.
- Services: which operations have effects in an actual application, storage system, or external tool.
NeuralOS studies learned interface dynamics. Its limitations and future-work discussion address the gap between simulated screens and underlying system services, as well as typing and long-term memory. NeuralOS revision.
Future systems can be compared by the state they preserve, the controls they expose, the services they connect to, and the experience they generate. These questions apply to code-rendered interfaces, pixel-generating simulators, and agents operating existing software.
Read next
Follow games to computers for the connection to interactive world models, then evaluation for a framework that separates visual plausibility, correct action response, and completed tasks. The NeuralOS record provides full authors, dated source links, and citations for both the preprint and the published version.