AutoSOTA proposes an end-to-end automated research system
A system for automated discovery of state-of-the-art AI models — hypothesis to experiment to result without a human in the loop. The interesting question is not whether it works but what it does to the literature if it does.
AutoSOTA describes an end-to-end automated research system for discovering state-of-the-art AI models: generating candidate approaches, running the experiments, and evaluating results without a person in the loop at each step.
The immediate technical question — does it find anything a human would not — is less interesting than the second-order one. Automated research systems that produce publishable results produce them at machine throughput, and the literature has no mechanism for absorbing results faster than humans can read them.
Machine learning already has a replication problem driven by volume. A system that increases the volume of results without a corresponding increase in verification capacity makes the existing problem worse in exactly the dimension it is already failing.
The optimistic reading is that automated search is best suited to precisely the work humans are worst at — exhaustive architecture and hyperparameter exploration, where the bottleneck is patience rather than insight. Automating that frees the humans for the part that requires taste.
The realistic reading is that it does both, and which one dominates depends on whether the verification side automates as fast as the generation side. Right now it is not close.
arXiv — AutoSOTA: An End-to-End Automated Research System for State-of-the-Art AI Model Discovery → · arXiv — Artificial Intelligence — August 2026 listings →