Mixed Signals: Weight-Space Programming for EEG and ECG

A paper on weight-space programming for EEG and ECG signals examines whether language models can read and edit the weights of trained neural networks to "train" them by reasoning instead of gradient descent, extending an approach previously demonstrated on game-playing networks and on fixing specific failures. The work focuses on applying weight-space editing to biomedical signal models, where the prior demonstrations did not cover the domain.

Weight-space programming allows language models to read a trained neural network's weights and edit them. This opens the door to training'' by reasoning rather than through gradient descent. Prior work has demonstrated its functionality and promise for game-playing networks and fixing specific fai