Preprint reports gains from AI agents tuning specialized neural networks

Researchers report in a preprint that a system of AI agents improved specialized neural networks across five classification and differential-equation tasks. Classification accuracy exceeded baselines by 13 percentage points on MNIST and 6 percentage points on Digits; error against an analytical single-soliton solution fell approximately 2491-fold. The system assigns analyst, coder and recorder roles to large language models, using an experiment log and a sandboxed execution environment. Agents generate code for specified model components while leaving the overall network structure unchanged. The same framework tuned spiking neural networks and networks trained to satisfy physical equations, but the reported gains concern these selected tasks rather than unrestricted model design.