Mixed Signals: Weight-Space Programming for EEG and ECG
In one experiment, a language model wrote network weights that matched gradient-descent training using six times fewer training subjects.
The researchers tested whether language models could inspect and edit networks trained on EEG and ECG signals, including repairing planted faults without being told what was wrong. Networks written by Claude Opus 5.5 also outperformed two hand-crafted EEG pipelines using the same five subjects. The models could not improve already well-trained classifiers or reliably identify a network’s task from its weights alone. The authors see potential for small networks on wearables and edge devices, but say practical use needs further investigation.
HF Daily Papers' note
The researchers tested whether language models could inspect and edit networks trained on EEG and ECG signals, including repairing planted faults without being told what was wrong. Networks written by Claude Opus 5.5 also outperformed two hand-crafted EEG pipelines using the same five subjects. The models could not improve already well-trained classifiers or reliably identify a network’s task from its weights alone. The authors see potential for small networks on wearables and edge devices, but say practical use needs further investigation.
HF Daily Papers' note
score 5