LLM reads and edits EEG and ECG model weights

The paper's authors apply weight-space programming to models trained on EEG and ECG biosignal data, showing that LLMs can read and edit the weights of trained networks, completing debugging and training through inference alone. The authors state that frontier models can debug and fix specific faults implanted in biosignal models without being told what the fault is, and can also write out all the weights of a network from scratch using only the architecture, the task, and a very small number of training subjects. In one set of experiments, a language model, without using any gradient descent steps, matched standard gradient descent training with 6 times fewer training subjects; a network written by Claude Opus 5.5 scored higher with the same 5 subjects than classic handcrafted EEG pipelines (alpha-band power, common spatial patterns combined with linear discriminant analysis). The authors also note limitations: LLMs struggle to improve already well-trained classification models, and cannot reliably identify from weights alone what task a network was trained on; the authors believe the method has practical potential for small networks on edge devices and wearables, but further research is needed before practical use.

Generated by AI from reports · Updated 7 hr ago