Claude Opus 5.5 writes EEG classifier weights from scratch

Claude Opus 5.5 wrote every one of 2,730 weights in a small EEG classifier by reasoning, with no gradient-descent step at any point, and hit 0.834 balanced accuracy on an eyes-open/closed task from only 5 practice subjects.

That is the central result of Mixed Signals: Weight-Space Programming for EEG and ECG, a preprint posted to arXiv on November 10 by Dara Varam, Taraneh Aminosharieh Najafi, and David Atienza. The reference network trained the conventional way with gradient descent on 30 subjects reached 0.806. Classical hand-crafted EEG pipelines, alpha band power and common spatial patterns with linear discriminant analysis, came in at 0.746 and 0.779 on the same task.

"In one set of experiments, a language model was able to match standard gradient-descent training using 6× fewer training subjects, through reasoning alone and without any gradient descent step," the authors write.

The EEG architecture, TinyEEGNet, has 2,730 weights. The ECG network, ECGResNet, has 941,189. The authors also ran a debugging experiment on the ECG side: they planted hidden recording artifacts inside the classifier, and Opus and GLM-5.3 each identified three of three without being told what to look for. After edits, myocardial-infarction AUC moved from 0.67–0.69 to 0.83–0.86.

Three language models were tested: Claude Opus 5.5, GLM-5.3, and DeepSeek-V4.1-Flash. The EEG data came from PhysioNet (64-electrode scalp recordings at 160 Hz from 109 subjects); the ECG data from PTB-XL (21,388 clinical recordings at 100 Hz).

The paper is equally direct about where the trick breaks. The authors list two headline limitations: "the inability of language models to improve already well-trained classification models, and their incapacity to reliably recognize, from the weights alone, the task a network was trained on." Their closing is bounded: weight-space programming "can be useful for small networks, such as those deployed on edge devices and wearables, but the concept still requires further investigation before it can be used in practice." Code is on GitHub.