EPFL: Claude writes EEG network weights by reasoning, matching gradient training with 6x fewer subjects

epfl-ch · hf · 2026-11-12

EPFL researchers extend weight-space programming to biosignal models, showing LLMs can read and write neural network weights via reasoning alone instead of gradient descent.

Key findings:

From-scratch weights: Given only the architecture, task, and a minimal set of training subjects, Claude Opus 5.5 wrote all 2,730 weights of TinyEEGNet from zero (no gradient steps), incorporating alpha 8–12Hz filtering, back-of-head channel weighting, and eye-movement detection.
Data efficiency: The LLM matched standard gradient-descent training with 6x fewer subjects. On eyes-open vs closed classification, Claude hit 0.834 balanced accuracy on 65 unseen samples from just 5 subjects, while gradient training needed 30 samples (0.806) to catch up — a +0.13 gap at 5 samples.
Beats classic pipelines: Outperformed hand-crafted alpha band power and CSP+LDA pipelines built from the same 5 subjects.
Limitations: LLMs can't improve already well-trained classifiers, and can't reliably identify a network's task from weights alone.

The authors conclude weight-space programming is promising for small edge/wearable networks but needs further study before practical use.

Original post →

EPFL: Claude writes EEG network weights by reasoning, matching gradient training with 6x fewer subjects

More from Models