NeuralStackA research system called ASI-Arch has been built to explore neural network architectures with limited direct human intervention. The system repeatedly researches possible ideas, runs experiments, analyzes the results and updates its direction in a closed loop.
The work focuses on linear attention, an area of model design aimed at improving how neural networks handle sequences. Across 1,773 iterative experiments, ASI-Arch discovered 105 architectures described by the paper as state of the art.
Its strongest architecture improved on DeltaNet by nearly three times the gain achieved by Mamba2. The result is presented not only as a performance improvement, but also as evidence that an AI system can manage a long-horizon research process rather than simply optimize a fixed, human-designed search space.
ASI-Arch combines autonomous hypothesis generation with implementation, training and empirical evaluation. The research also examines how different parts of the framework contribute to progress, offering a view into which mechanisms help an agent navigate a difficult architecture-discovery problem.
The project reflects growing interest in systems that apply language-model agents to AI research itself. Instead of treating an agent as a tool for isolated coding or literature tasks, the framework places it inside a repeated experimental cycle where earlier findings influence later proposals.
Why this matters
If the approach generalizes, architecture discovery could become less dependent on researchers manually proposing every major design change. That could expand the number of ideas tested and make some forms of AI progress more computationally scalable.
The claims also invite scrutiny. Critics cited on the article page argue that the system resembles LLM-based neural architecture search and question whether generating and testing model variations demonstrates open-ended creativity or artificial superintelligence. Those objections highlight the difference between an effective engineering system and a system that independently originates fundamentally new research agendas.
For now, ASI-Arch offers a concrete demonstration in one specialized domain rather than proof of general autonomous AI research. Its next test will be whether similar closed-loop methods can produce reliable advances across other architectures, tasks and research fields without relying on narrowly defined objectives.