EPFL's Claude Opus 5.5 Writes a Brain Signal Classifier With Zero Training
An EPFL team gave frontier LLMs trained EEG and ECG network weights as text and asked them to debug, repair, and even write networks from scratch, without any gradient descent.
Takeaways
- EPFL used frontier LLMs to write and edit biosignal network weights directly, no gradient descent involved. Paper
- Claude Opus 5.5 wrote a TinyEEGNet from 5 subjects that matched gradient descent trained on 30 subjects.
- LLM-written networks beat classic EEG pipelines (alpha power, CSP+LDA) on the same 5 practice subjects.
- Opus, GLM, and DeepSeek diagnosed hidden ECG recording artifacts without being told the cause.
- No model could improve a healthy network or reliably identify a network's task from weights alone.
- Full experiments, prompts, and reasoning traces available on the project page.
Researchers at EPFL’s Embedded Systems Laboratory report that Claude Opus 5.5 wrote all 2,730 weights of a neural network using recordings from five practice subjects and zero gradient updates. On an eyes-open versus eyes-closed task, the network scored 0.834 on 65 unseen samples, compared with 0.806 for a conventional model trained on 30 subjects. That comparison underpins the sixfold data-efficiency claim in the team’s Mixed Signals paper.
Gradient descent normally learns a network’s weights by calculating how each value contributed to an error, then adjusting those values to reduce it. The EPFL team tested weight-space programming, or WSP, in which a language model inspects weight tensors through a Python sandbox and writes code to modify them. Earlier studies applied WSP to toy game agents and targeted patches for language and diffusion models. This study asks whether domain knowledge about physiology and recording hardware can guide the process.
Why biosignals fit the method
Electroencephalography, or EEG, measures electrical activity from the brain, while electrocardiography, or ECG, records the heart’s electrical signals. Their common patterns and failure modes are extensively documented: electrode placement changes signal shape, power lines introduce predictable interference, muscle activity creates artifacts, and posterior alpha rhythms strengthen when the eyes close. A frontier language model can draw on that technical literature when choosing filters and connections.
Model size makes the experiment tractable because TinyEEGNet contains only 2,730 weights, allowing the complete network to fit within the workflow. The study’s ECGResNet contains about 941,000 parameters, so the language models made localized repairs to that network. Architecture details, prompts, and reasoning traces appear on the project page.
Five tests, zero gradient steps
The authors tested Claude Opus 5.5, GLM-5.3, and DeepSeek-V4.1-Flash on five tasks of increasing scope. A Python sandbox held the tensors and exposed an evaluate