Researchers built a three-agent AI system that automatically tunes the internal components of specialized neural networks, no human tuning required.
The system assigns three roles to large language models: an analyst that decides what to try, a coder that writes the component, and a recorder that logs every experiment in a sandboxed environment. Crucially, the agents only rewrite small, specified pieces of a model - never the overall architecture - so the same framework works across very different problems. The team tested it on five tasks: classifying handwritten digits with spiking neural networks on MNIST and Digits, and solving two stubborn physics problems, the generalized nonlinear Schrodinger equation and the Kuramoto-Sivashinsky equation, with physics-informed neural networks. Tuned by the agents, classification accuracy rose 13 percentage points on MNIST and 6 points on Digits, while error reductions across the physics tasks ranged from 14-fold (mass drift in the Kuramoto-Sivashinsky run) to roughly 2,491-fold (matching the analytical soliton solution), with the invariant-preservation error falling more than 100-fold on average.
That spread matters more than any single headline number. A 14-fold gain on one physics metric and a 2,491-fold gain on another suggest the system is finding real, if wildly inconsistent, improvements across genuinely different problem types rather than overfitting to one lucky case. Unlike full neural architecture search, this approach leaves the big structural decisions to humans and automates only the fiddly component-level choices experts usually hand-tune themselves.
That is a more modest claim than most AI-redesigns-AI pitches make, and five tasks is a thin evidence base - but it is also a far more believable one.