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. In this work, we present its applicability to models trained on biosignal data, specifically EEG and ECG data. We show that frontier models are not only able to debug specific faults planted in biosignal-based models without being told what these faults are, but are even capable of writing all the weights of a network from scratch, given only the architecture, the task, and a minimal set of training subjects. In particular, in one set of experiments, a language model was able to match standard gradient-descent training using $6\times$ fewer training subjects, through reasoning alone and without any gradient descent step. Networks written by Claude Opus 5.5 also scored higher than classic hand-crafted EEG pipelines (alpha band power, and common spatial patterns with linear discriminant analysis) when built from the same 5 subjects. We also highlight some of the limitations, which include 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. Overall, our results indicate that 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.
Alex: A language model wrote a working brain-signal classifier without taking a single gradient step. But the same study found it could neither improve well-trained classifiers nor reliably tell what their weights were for.
Sam: That separates building something useful from understanding something already built. Whose study is this, and how broadly should we take that result?
Alex: It’s an experimental preprint by Dara Varam and colleagues at EPFL, called “Mixed Signals: Weight-Space Programming for EEG and ECG.” They tested language models on networks processing electrical recordings from the brain and heart.
Sam: For researchers short of labeled recordings, the attraction would be needing less data. Is that the practical promise here?
Alex: Yes, especially for compact classifiers that could run inside wearables. But this is an early demonstration, not a deployment recipe. We’ll unpack what the models actually did, the strongest comparison, and why apparent physiological knowledge needs a careful qualification.
Sam: Start with the alternative to ordinary training. What does weight-space programming actually ask a language model to do?
Alex: Weights are the numerical settings that determine a neural network’s computations. Usually, gradient descent adjusts them incrementally using labeled examples. Weight-space programming instead asks a language model to inspect those settings and write new ones, much as a programmer edits code.
Sam: The authors didn’t invent that idea here. What changes when they bring it to brain and heart recordings?
Alex: Earlier work had demonstrated game-playing networks, repairs to particular model failures, and networks written for algorithmic tasks. Here, the proposed advantage is physiological knowledge. A language model may already know what meaningful electrical patterns should look like, while an optimizer has to extract useful structure from the available recordings.
Sam: And these recordings vary across people. A classifier could learn something about the recording setup rather than the physiology.
Alex: That’s central to the study. The authors test everything from identifying an existing network’s task to writing its weights from zero. Between those extremes, they ask whether models can improve healthy classifiers, repair planted faults, and diagnose failures on unseen recordings.
Sam: “By reasoning” could hide quite a lot of computation. Could the language model run training software behind the scenes?
Alex: The rules prohibited gradient descent, optimizers, and iterative fitting. Most experiments allowed a restricted Python session, with limits on network evaluations. Depending on the experiment, feedback included accuracy, individual predictions, or internal activity; some heart-signal experiments also allowed one-shot linear algebra.
Sam: So this wasn’t necessarily staring at numbers and producing an answer immediately. Some tasks allowed a cycle of proposed edits and feedback.
Alex: Yes. And in the network-writing experiment, the model could analyze practice recordings and design filters, but not fit or search network components against recordings or labels. The strongest result came from that constrained, tool-assisted process—not from reading weights alone.
Sam: Let’s make that result concrete. What was the model trying to recognize?
Alex: Whether someone’s eyes were open or closed, using electroencephalography, or EEG: electrical activity recorded through scalp electrodes. Claude Opus wrote every weight of a small classifier starting from zeros. It had recordings from five practice subjects.
Sam: What did it build into those weights, rather than learn through conventional training?
Alex: Its network looked for alpha rhythm, a pattern of brain activity, over the back of the head. It separated frequency bands, compared neighboring electrode rows, and turned those comparisons into amplitude detectors. Strong alpha activity relative to neighboring bands indicated closed eyes.
Sam: That sounds like assembling a physiological signal-processing pipeline inside a neural network. Did it work beyond the people whose recordings it could inspect?
Alex: On sixty-five unseen test subjects, its balanced accuracy was about eighty-three percent. That measure averages accuracy equally across the classes. The same architecture trained conventionally on thirty subjects reached about eighty-one percent; the uncertainty around the difference included no advantage.
Sam: So the defensible headline is comparable performance with fewer subjects, not established superiority over the larger training set.
Alex: That’s the distinction. Opus also scored above the conventional network trained on the same small number of subjects. And it outscored both hand-crafted comparisons fitted on those practice subjects: an alpha-power threshold and a spatial-filtering pipeline followed by a linear classifier.
Sam: Eye closure has a recognizable physiological signature. Did the approach work on anything less direct?
Alex: They also tested left-versus-right fist movement. Opus wrote a mirror-symmetric network, so exchanging left and right electrodes also reversed its decision. It worked on unseen subjects, but fell below the network trained on the full training group. The open-weight language models wrote working networks too, generally with weaker performance.
Sam: For someone building a small device, was this a faster route to the classifier?
Alex: Not in these experiments. Writing the network took roughly thirty-eight minutes, while conventional training with the small subject set took only seconds on the authors’ GPU. The written and conventionally trained networks had the same architecture, so their running cost after deployment was identical.
Sam: That covers construction. What happened when the models had to understand an existing network?
Alex: They failed to identify its task reliably from the weights. Giving them electrode names sometimes helped, but performance on untrained controls suggested guessing. Opus often recognized the dataset’s layout and sampling rate, then chose a familiar task regardless of what the weights contained.
Sam: That makes the repair results particularly useful. Could they fix a network without understanding its original purpose?
Alex: Concrete faults were easier. Opus and DeepSeek could spot and undo an abnormally enlarged filter or shifted output bias, even without feedback. But faults that left ordinary-looking weights—such as exchanging class outputs—were nearly undetectable without running the network.
Sam: Did feedback reveal those hidden faults, or just give the models a way around them?
Alex: Mostly a way around them. With feedback, Opus and DeepSeek recovered nearly all the lost performance across the planted EEG faults. Yet they rarely named the hidden faults correctly. Successful repair did not mean a correct explanation.
Sam: The heart-recording tests sound closer to the shortcut problem we started with. What did the authors plant there?
Alex: They trained networks to associate heart-attack labels with recording artifacts, including electrical hum and a calibration pulse. When that association disappeared in unseen recordings, performance suffered. Opus and another model identified every artifact type, then edited the networks to largely stop relying on them.
Sam: Diagnosing a shortcut is useful, but was the resulting repair better than ordinary retraining?
Alex: None beat the best training-based fix using the same labeled practice data. Some of those fixes required knowing the cause, which the language models discovered independently. And no language model improved the already healthy brain-signal classifiers.
Sam: Before treating this as evidence of physiological reasoning, what uncertainty matters most?
Alex: The datasets are public and likely appeared in the language models’ training material. The authors acknowledge that apparent physiological knowledge could partly be knowledge of these particular datasets. Also, network writing and the larger debugging tests used a single run per model and setting, so repeatability remains unmeasured.
Sam: Given those limits, who should read the full document, and where should they begin?
Alex: Researchers developing biosignal classifiers or compact wearable models should start with “A network written from zero.” Then read the experimental rules and limitations, especially the held-out-subject comparisons and possible dataset familiarity. Those details distinguish a promising construction method from a general replacement for training.
Sam: And what should everyone else remember?
Alex: Writing a useful classifier is not the same as understanding a trained one. This study supports the former in a narrow setting, while exposing clear limits on the latter.
Sam: Keep that distinction in mind when the next weight-editing claim arrives.