Can a language model write a brain-signal network?

The question behind the weights

A neural network’s behavior depends on its weights: numbers that shape how signals pass through its layers. Usually, an optimizer adjusts these numbers through gradient-descent steps, using labeled examples. In weight-space programming, a language model receives information about a network and proposes weight changes—or writes weights from scratch—without taking gradient steps itself.

Dara Varam, Taraneh Aminosharieh Najafi and David Atienza tested this approach on electroencephalogram (EEG) recordings of brain activity and electrocardiogram (ECG) recordings of heart activity. Their experiments asked whether language models could identify a network’s task, improve a functioning model, repair planted faults, diagnose a hidden failure, or create a network from scratch.

How the tests worked

The researchers compared Claude Opus 5.5, GLM-5.3 and DeepSeek-V4.1-Flash. They used a compact EEG classifier with 2,730 weights and an ECG classifier with 941,189 trainable weights. Depending on the experiment, models received network details, practice recordings and labels, and opportunities to run the network and receive feedback. They were instructed not to use gradient descent or iterative fitting. Evaluation used data excluded from the model’s practice data, though the precise split differed across experiments.

The tests included EEG tasks such as distinguishing eyes-open from eyes-closed recordings. Some networks had deliberate faults, such as altered weights or exchanged electrodes. In another test, an ECG network learned a misleading shortcut: an artificial recording artifact was correlated with heart-attack labels during training but not in the held-out data. For the from-scratch challenge, a model received an all-zero EEG network, its architecture and data from five practice subjects, then had to supply every weight.

A promising result, within one setup

For eyes-open versus eyes-closed classification, Claude Opus wrote the EEG network’s weights using five practice subjects. It achieved 0.834 balanced accuracy on 65 unseen subjects. The same architecture trained with gradient descent reached 0.704 using five subjects and 0.806 using 30. Claude’s result therefore exceeded that particular 30-subject baseline. Its network also scored above two hand-crafted EEG methods fitted on the same five subjects. This is evidence about one task and setup, not evidence that language models generally outperform conventional training.

The models also showed some troubleshooting ability, with important differences between tests. On planted EEG faults, success varied by fault and model. For faults that were invisible in the weights, feedback enabled Claude Opus and DeepSeek to recover performance, but they generally did not identify the fault’s mechanism; their edits were guided by performance feedback. The ECG electrode-swap repair was harder, and only Opus substantially changed the network’s behavior.

For the hidden ECG shortcuts, Opus and GLM identified all three artifact types, while DeepSeek identified two. The edits reduced how often the artifacts flipped heart-attack predictions. Yet no model’s edit beat the best standard, gradient-based fix. DeepSeek’s repair for the muscle-filter case was a notable weaker outcome: its reported MI-AUC was 0.79, compared with 0.83–0.86 for the other model-and-artifact results described by the paper.

What the models could not do

In a specific task-identification experiment, models saw rounded EEG-network weights and chose among candidate tasks. They did not reliably identify the network’s task; their answers were similar to guesses on control networks. None improved the already well-trained EEG classifiers in the study. These results do not show that language models cannot improve any trained network, but they do mark limits in these experiments.

The authors suggest that successful cases may draw on recognizable network structure and prior knowledge of the task, rather than a general ability to infer what arbitrary weights mean or optimize any model. The experiments do not establish which explanation accounts for each result.

Why more evidence is needed

Several experiments used just one run per model and setting, leaving run-to-run consistency unknown. The datasets came from PhysioNet and are public; the authors note that they may have appeared in the language models’ training data. That could contribute to apparent familiarity, but the study does not establish that it caused the results. A single language model graded textual diagnoses, and the hand-crafted EEG comparisons were added after the language-model runs rather than specified in advance alongside the main tests.

There is also a practical trade-off: writing the small EEG network took tens of minutes, while conventional training took seconds in the reported setup. Success on these held-out data does not establish medical reliability or performance at other hospitals and populations.

This is an arXiv preprint describing early research, not clinical validation. Its results make direct weight editing an interesting topic for further study, particularly for small networks, but do not support using it as a clinical tool.