Mixed Signals: Weight-Space Programming for EEG and ECG

Mixed Signals. Weight-Space Programming for EEG and ECG. By Dara Varam and two others, from EPFL. Published on the tenth of November.

Weight-space programming means letting a language model read the weights of a trained neural network, and edit them. That opens the door to training by reasoning rather than by gradient descent. Earlier work showed it working on game-playing networks, on fixing specific failures in language models, and even on diffusion models. The claim here extends it to models trained on biosignals, specifically EEG and ECG data, and pushes it past repair. Frontier models can debug faults planted in such networks without being told what the faults are. And they can write every weight of a network from scratch, given only three things: the architecture, the task, and a minimal set of training subjects.

The strongest result comes from one set of experiments. There, a language model matched standard gradient-descent training while using six times fewer training subjects. It got there through reasoning alone, without a single gradient descent step. Put another way, the standard recipe drew on six times as many subjects and ran gradient descent to arrive at its weights. The language model was handed the architecture, the task and the smaller group of subjects, and the weights it wrote matched what that standard training achieved.

A second comparison sets model-written networks against classic hand-crafted EEG pipelines. One pipeline uses alpha band power. The other uses common spatial patterns with linear discriminant analysis. When everything was built from the same five subjects, networks written by Claude Opus 5.5 scored higher than these classic pipelines. Both sides, in other words, started from the same small pool of subjects. Alongside that sits the debugging result: specific planted faults found and fixed, with the model never told what it was looking for.

The limits are stated plainly, and they shape how far the claim reaches. Language models were not able to improve classification models that were already well trained. They also could not reliably recognize, from the weights alone, which task a network had been trained on. The abilities on show, then, are writing small networks from very little data and repairing planted faults, not polishing a strong model or reading a network's purpose out of its parameters. The authors place the usefulness with small networks, such as those deployed on edge devices and wearables. And they say directly that the concept still requires further investigation before it can be used in practice.

What shifts is the route to a working network. Instead of many subjects and many gradient steps, a language model reasons over an architecture, a task and a handful of subjects, and writes the weights out directly. For biosignal models small enough to run on a wearable, the authors find that route can be useful. Its failures on well-trained models, and on naming a network's task from its weights, mark how much of the idea is still open.