Large language models can audit small neural networks by directly manipulating weights to find and fix specific bugs like hidden decision rules or threshold shifts, but this approach works better for repair than for certification.
Researchers tested whether Claude Opus can audit and repair small neural networks used in radiotherapy by reading and modifying their weights directly, without code or data access. The model successfully identified 16 of 18 intentionally flawed networks (95% accuracy) and repaired them by changing just a few parameters, outperforming fine-tuning which modified nearly all weights.