Language models write neural network weights from scratch, matching gradient descent with 6x fewer subjects
IgorCarron · x · 2026-11-17
An EPFL paper on arXiv (Mixed Signals: Weight-Space Programming for EEG and ECG) shows "weight-space programming": frontier LLMs can read and edit a trained network's weights, and even write all weights from scratch given only the architecture, task, and a minimal set of training subjects.
Key results:
- In one experiment, a language model matched standard gradient-descent training using 6x fewer subjects, through reasoning alone;
- Networks written by Claude Opus 5.5 outperformed classic hand-crafted EEG pipelines (alpha band power, CSP+LDA) built from the same 5 subjects;
- Models could also debug planted faults in biosignal models without being told what they were.
Limitations: LLMs struggle to improve already well-trained classifiers and can't reliably identify a network's trained task from weights alone. The takeaway: training is turning into thinking, especially valuable for data-scarce biosignal modeling.
More from Research
- Researcher's pet peeves: the clichés that plague nearly every AI paper's intro — StephenLCasper · 2026-11-17
- Looped Transformers explained: same weights, evolving hidden states, extra compute per parameter — TheTuringPost · 2026-11-17
- MIT CSAIL's TangiMap uses an AI agent to map software onto physical controls — MIT_CSAIL · 2026-11-17
- SWE-bench co-author on building influential AI benchmarks: start with people — OfirPress · 2026-11-17
- Better representations first: where does data scaling break down in robot learning? — KyleMorgenstein · 2026-11-17
- Policies use only a sliver of image information — extracting it matters for safety — KyleMorgenstein · 2026-11-17