Muscle-signal interfaces decode what a person intends to do from the electrical activity of their forearm muscles, recorded by electrodes on the skin. This recording is called surface electromyography (sEMG). These decoders fail on people they were not trained on. Fat thickness, muscle geometry and electrode placement differ enough between people that each user's signals look different, so today every new user, and often every new day, needs a recalibration session. The large-language-model literature has recently shown that one model's behavior can be moved into another model without any gradient training. The method records both models' internal activations on the same roughly 1,000 prompts and solves a single least-squares problem for a linear map between them. We propose to use that same closed-form step, with no backpropagation and no optimizer, to map a new user's internal activations inside a pretrained sEMG Transformer onto the activations the model produces for its training population. The idea in two sentences is this. Instead of fine-tuning the decoder for each new person, we treat the person as a "different model" of the same underlying motor behavior and fit a linear converter between their hidden states and the population's hidden states, in a few seconds on a CPU. Whether this works depends on a scientific question that no paper in our reading answers: at what depth of a trained sEMG Transformer, and at what model scale, does the difference between people become a linear, low-rank transformation?
The timing is good because Scaling and Distilling Transformer Models for sEMG showed that causal Transformers trained on raw sEMG improve log-linearly up to 109M parameters and generalize better to unseen users as they grow. That paper reports a character error rate (CER) of 30.5% for the 109M teacher and 31.9% for a distilled 2.2M student on emg2qwerty. Neither model is adapted to the individual. If larger models do generalize by making people "look alike up to a linear map," closed-form alignment should get better with scale. That is a falsifiable link between the scaling result and the cost of personalization.
Core Claim
We claim that per-user sEMG personalization of a pretrained Transformer can be done by solving one regularized least-squares problem at an intermediate layer. The data are pairs of activations, matched by what the user was doing, not by what their signal looked like. This step recovers most of the accuracy that gradient-based adaptation provides, at a small fraction of its compute and with no learning-rate or step-count tuning. We further claim that the fraction of between-user variation that a linear map can remove rises with model size and peaks at a middle depth. The prediction is that early layers are still tied to electrode-level idiosyncrasies and late layers are already task-specific.
What is new relative to Command-V: Training-Free Representation Finetuning Transfer is the pairing problem. Command-V aligns two text models by feeding both of them identical prompts. Two humans never produce identical sEMG, so there is no shared input. We instead pair by shared content. In emg2qwerty, a calibration window of the new user typing a given character, aligned to the keystroke, is paired with the population's mean activation for that character at the same relative time. In gesture datasets, pairing is by gesture class and phase. Let $H_u \in \mathbb{R}^{n \times d}$ be the new user's hidden states at layer $\ell$ for $n$ calibration windows, and $H_{\text{ref}}$ the matched population targets. We fit
$$ W_u = \arg\min_W \; \lVert H_u W - H_{\text{ref}} \rVert_F^2 + \lambda \lVert W - I \rVert_F^2 = (H_u^\top H_u + \lambda I)^{-1}(H_u^\top H_{\text{ref}} + \lambda I), $$
where $d$ is the hidden width, $I$ is the identity and $\lambda$ controls how far the map may move from "do nothing." We then insert $W_u$ (plus a bias term) at layer $\ell$ during inference. A rank-limited variant, $W = I + BA$ with low rank $r$ obtained by truncated SVD of the ridge solution, keeps the stored per-user state small. This makes it a closed-form analogue of the low-rank adapters (LoRA) used for language models.
Relation to Prior Work
Three sEMG papers motivate the gap. The scaling paper above builds large decoders but personalizes nothing; it compresses statically by distillation. TEMGNet reaches 82.93% on NinaPro DB2 with about 65k parameters, but it trains a separate model per subject, which is exactly the recalibration burden we want to remove. CPEP aligns sEMG embeddings with hand-pose embeddings for zero-shot gesture transfer, and it evaluates only with linear probes on frozen features. We align inside the backbone rather than at its output, and we align people to people rather than sEMG to pose.
One sEMG paper is the closest competitor on adaptation without backpropagation. SpGesture adapts to arm-posture shifts without source data by pseudo-labeling in the membrane-potential space of a spiking network, and gains 4.1–4.5 points. Its adaptation is unsupervised and tied to spiking architectures. Ours needs a short labeled calibration session but works on any standard Transformer.
Command-V supplies the method: closed-form pseudoinverse converters between activation spaces. It also constrains us, because it documents oversteering and a cross-family failure in which a transferred adapter raised jailbreak success above 60%. We therefore regularize toward identity and report degradation cases explicitly.
The efficiency literature defines the comparison class. KerZOO cuts zeroth-order (forward-pass-only) fine-tuning cost substantially relative to MeZO. MaCP halves activation memory relative to LoRA by tuning cosine-transform coefficients. Sparsity-Accelerated Training skips inactive neurons during training. All three still run an iterative optimizer. Our proposal removes iteration entirely, at the price of assuming the user shift is linear at some layer. The experiments are designed to test that assumption, not to take it for granted.
Research Program
The first phase is diagnostic. We train, or reproduce, the emg2qwerty Transformer family across the size range studied in the scaling paper, including the distilled 2.2M student. For each held-out user and each layer, we measure "linear alignability." This is the fraction of the gap between that user's content-matched activations and the population targets that the ridge map removes on held-out calibration windows. We plot it against depth and model size. The scaling hypothesis predicts that the peak alignability rises with parameters. We also estimate the effective rank of $W_u - I$ to see whether user identity lives in a few directions.
The second phase turns the diagnostic into a personalization procedure. We pick the layer using a small validation set of users. We then sweep calibration length from roughly 30 seconds to 10 minutes of typing, and record CER, CPU wall-clock time for the solve, and per-user stored parameters. We test three placements: single-layer alignment, alignment at the best two layers solved one after the other, and the rank-limited variant. We also test whether a map fit on the 109M teacher can be transferred to the 2.2M student through a second Command-V-style converter between the two models. The student's own map is the comparison.
The third phase tests where linearity breaks. We transfer the procedure to NinaPro DB2 with leave-one-subject-out splits. We move it to the SpGesture posture dataset, where the shift is arm position within a person rather than identity across people. Where recordings allow, we examine within-session fatigue, which the sEMG literature describes as a nonlinear warping of the amplitude–force relationship. That is the case in which we expect a single linear map to fail.
Baselines
The naive reference point is the unadapted generic model. On emg2qwerty the file reports 30.5% CER for the 109M teacher, 31.9% for the distilled 2.2M student and 35.9% for the 2.2M model trained from scratch, all from Scaling and Distilling Transformer Models for sEMG. Any adaptation must first beat these on the same held-out users.
The strongest existing competitor is gradient-based LoRA personalization of the same backbone on the same calibration data. We found no reported result on Lacuna for personalized CER on emg2qwerty with LoRA or full fine-tuning, so we will establish it ourselves. For fairness it gets the same calibration windows. It also gets a per-user tuning budget of a learning-rate and step-count grid chosen on validation users, not test users. We report both its best-case accuracy and its compute (optimizer steps, peak memory, CPU and GPU seconds). Full fine-tuning is run as an upper bound. A zeroth-order variant in the style of KerZOO is run as the backprop-free iterative competitor, under matched forward-pass budgets. Beating LoRA outright is not our success criterion. Success means recovering at least half of the CER gap between the generic model and LoRA personalization at under 5% of LoRA's adaptation compute. The alternative success criterion is matching LoRA in the shortest-calibration regime, where gradient methods overfit.
On gesture data, the reference points are TEMGNet's 82.93% within-subject accuracy on NinaPro DB2 from TEMGNet, and SpGesture's 89.26% overall accuracy and 4.1–4.5 point posture-adaptation gain from SpGesture. No reported cross-subject NinaPro DB2 result with a pretrained Transformer was found on Lacuna.
Two ablations isolate the key ingredients. The first replaces content-paired regression with unpaired moment matching, meaning we match the new user's activation mean and covariance to the population's without labels. This tests whether pairing by behavior is what makes the method work. The second moves the same closed-form map from the hidden layer to the raw input channels, fitting a 16-by-16 channel-mixing matrix per wrist. This tests whether depth matters or whether a classical input-space correction already suffices.
Why not just…?
Why not just retrain the output head on frozen features? This is the standard move in sEMG and costs even less than our method. It is our lower gradient-based rung, and it is also a closed-form special case at the last layer. CPEP found that linear probing of supervised baselines generalized poorly to unseen gestures, which suggests the features themselves shift. For people rather than gestures, our layer-wise alignability curve answers the question directly. If the curve peaks at the final layer, head retraining is the right answer and we will say so.
Why not just run LoRA with backpropagation? The 2.2M student is tiny. This is the real threat. The memory argument for avoiding backprop largely disappears at 2.2M parameters, since backprop there fits on a phone. We therefore do not rest the case on memory. The case is threefold. First, the method has no step count, learning rate or early-stopping choice, which matters when a clinician calibrates a prosthesis without supervision. Second, least squares with an identity prior may overfit less than gradient descent on one minute of data. Third, the map can be solved once on the teacher and reused. The calibration-length sweep settles the question. If LoRA dominates at every calibration length at comparable wall-clock time, our contribution shrinks to the diagnostic finding about where user identity lives, and we will report that.
Why not just correct the signal at the electrodes with a channel-mixing matrix or per-channel renormalization? Electrode rotation and displacement are roughly linear mixtures of channels, so an input-space fix may fully handle within-person session shifts. We expect it to fall short across people, whose anatomy changes the signal nonlinearly before it reaches the electrodes. The input-space ablation above measures exactly this. If it matches hidden-layer alignment across subjects on emg2qwerty, we will drop the depth claim and retarget the work to the regimes where it does not.
Proposed Experiments and Decisive Checks
The primary benchmark is emg2qwerty, using the scaling paper's held-out users and CER as the metric. Secondary benchmarks are NinaPro DB2 (accuracy, leave-one-subject-out) and the SpGesture posture dataset (accuracy across postures, compared with SSFDA). For every method we report accuracy, adaptation wall-clock time on a laptop CPU, peak memory, per-user stored parameters and minutes of calibration required.
Three outcomes would be decisive. First, if the peak linear alignability does not rise with model size across the scaling family, the proposed link between scale and cheap personalization is false. Second, if closed-form alignment closes less than a third of the gap to LoRA at every calibration length, the method is not competitive and the paper becomes a negative result about linear transfer outside language. Third, if content-paired alignment beats unpaired moment matching by a clear margin, and hidden-layer alignment beats input-layer alignment across subjects, then the two ingredients that distinguish this proposal from Command-V and from classical channel correction are each shown to matter. We will also report every user for whom alignment raised the error rate. This is the sEMG counterpart of Command-V's oversteering failure, and a deployable method has to show those cases.
References
- Scaling and Distilling Transformer Models for sEMG
- Command-V: Training-Free Representation Finetuning Transfer
- TEMGNet: Deep Transformer-based Decoding of Upperlimb sEMG for Hand Gestures Recognition
- CPEP: Contrastive Pose-EMG Pre-training Enhances Gesture Generalization on EMG signals
- SpGesture: Source-Free Domain-adaptive sEMG-based Gesture Recognition with Jaccard Attentive Spiking Neural Network
- KerZOO: Kernel Function Informed Zeroth-Order Optimization for Accurate and Accelerated LLM Fine-Tuning
- MaCP: Minimal yet Mighty Adaptation via Hierarchical Cosine Projection
- Sparsity-Accelerated Training for Large Language Models