Paper tests weight-space programming on EEG and ECG models
Original title: Mixed Signals: Weight-Space Programming for EEG and ECG
Claude Opus 5.5 matched gradient training with six times fewer subjects and beat two hand-crafted EEG pipelines.
A paper applies weight-space programming, in which language models read and edit neural-network weights, to models trained on EEG and ECG data. Frontier models debugged planted faults without being told what they were and, in some experiments, wrote all network weights from the architecture, task, and a small subject set. The authors say the approach may suit small edge and wearable networks but still needs investigation before practical use.
- One model matched gradient-descent training using six times fewer subjects, without gradient steps.
- Claude Opus 5.5 networks beat alpha-band and CSP-LDA pipelines built from the same five subjects.
- Models could not reliably improve already well-trained classifiers.
- They also could not reliably identify a network’s task from weights alone.
From the original article
1 min read at huggingface.co
Weight-space programming allows language models to read a trained neural network's weights and edit them. This opens the door to ``training'' by reasoning rather than through gradient descent. Prior work has demonstrated its functionality and promise for game-playing networks and fixing specific failures in language models, even extending to diffusion models.
By Dara Varam, Taraneh Aminosharieh Najafi, David Atienza · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.