# Shortcut Learning

5 related articles

Subscribe to updates

arXiv Nov 10, 2026
A language model writes an EEG network without a single gradient step: all 2,730 weights by hand, 0.834 balanced accuracy from five subjects
Weight-space programming lets a language model read and edit a trained network's weights directly, and had been shown on game-playing networks and language models. Applied to EEG and ECG models, Claude Opus 5.5 wrote all 2,730 weights of TinyEEGNet with no gradient step and only five practice subjects, reaching 0.834 balanced accuracy on 65 unseen subjects, on par with 0.806 for the same architecture trained on 30 subjects and above 0.704 trained on five; it could not infer a network's task from its weights, nor improve an already well-trained network.
Weight-space programming lets a language model read and edit a trained network's weights directly, and had been shown on game-playing networks and language models. Applied to EEG and ECG models, Claude Opus 5.5 wrote all 2,730 weights of TinyEEGNet with no gradient step and only five practice subjects, reaching 0.834 balanced accuracy on 65 unseen subjects, on par with 0.806 for the same architecture trained on 30 subjects and above 0.704 trained on five; it could not infer a network's task from its weights, nor improve an already well-trained network.
Dara Varam（First author）
Weight-Space Programming Electroencephalography Electrocardiography Claude Opus 5.5 Shortcut Learning
researchsquare.com Oct 22, 2026
ECG models break on filter phase, not filtering: a causal all-pass shift drops conduction-block AUROC from 0.91 to 0.62
Real-time ECG monitoring must use causal filters, whose nonlinear phase misaligns the frequency components of the cardiac waveform. Across 36,583 twelve-lead records from three independent clinical databases, changing phase alone while holding spectral magnitude fixed cut ResNet-18 macro-AUROC by 0.1244 on PTB-XL, with conduction-block detection falling from above 0.91 to about 0.62. A training-time regularizer that randomly applies all-pass phase perturbations recovered 86.9% to 90.9% of that loss across cohorts without changing the architecture.
Real-time ECG monitoring must use causal filters, whose nonlinear phase misaligns the frequency components of the cardiac waveform. Across 36,583 twelve-lead records from three independent clinical databases, changing phase alone while holding spectral magnitude fixed cut ResNet-18 macro-AUROC by 0.1244 on PTB-XL, with conduction-block detection falling from above 0.91 to about 0.62. A training-time regularizer that randomly applies all-pass phase perturbations recovered 86.9% to 90.9% of that loss across cohorts without changing the architecture.
Universidad Carlos III de Madrid
Álvaro García-Casas（First author, Corresponding author）
Explainable AI Shortcut Learning Convolutional Neural Network
arXiv Oct 13, 2026
Injection probes were reading the conversation, not the injection: matched negatives cut false alarms from 80% to 17%
Linear injection probes could reach 0.91–0.98 AUROC on the standard evaluation using conversation position and text length alone; adding same-moment clean twins and harmless look-alikes to training cuts false alarms on injection-free conversations from 41–80% to 4–27% across six open models.
Georgia Institute of Technology
Kaan Kale（First author）
Matched Negatives Training Linear Probe Prompt Injection Shortcut Learning AgentDojo
arXiv Jun 2, 2026
Texture-driven visual learning suffers low-frequency shortcuts: pruning low-frequency components raises in-distribution accuracy by up to 10% and out-of-distribution accuracy by up to 40%
This work analyzes shortcut learning in texture-driven visual domains and compares it with a shape-driven standard benchmark, finding that texture-driven domains base most decisions on a few low-frequency components (LFCs) with skewed spectral behavior even though higher-frequency components (HFCs) have higher predictive power; pruning LFCs from training and test sets mitigates the shortcut and yields more balanced spectral behavior, improving in-distribution accuracy by up to 10% and out-of-distribution accuracy by up to 40% under algorithmic and real-world domain shifts, while general-purpose and domain-specific foundation models can also suffer from low-frequency shortcuts.
This work analyzes shortcut learning in texture-driven visual domains and compares it with a shape-driven standard benchmark, finding that texture-driven domains base most decisions on a few low-frequency components (LFCs) with skewed spectral behavior even though higher-frequency components (HFCs) have higher predictive power; pruning LFCs from training and test sets mitigates the shortcut and yields more balanced spectral behavior, improving in-distribution accuracy by up to 10% and out-of-distribution accuracy by up to 40% under algorithmic and real-world domain shifts, while general-purpose and domain-specific foundation models can also suffer from low-frequency shortcuts.
Harvard University
Utku Şirin（First author）
Low-Frequency Shortcuts Shortcut Learning Texture-Driven Visual Learning Foundation Models
arXiv Oct 1, 2026
Removing the overlapping-window timing shortcut lets non-invasive brain-to-text reach 36.6% word error rate with five observations per word
The work shows that most of the reported gain from jointly decoding all words in a sentence (d'Ascoli et al., 2025) is reproducible on synthetic signals containing no brain information (22.0% versus 22.3% on real MEG), because neighbouring three-second word-aligned windows overlap and implicitly leak word duration; decoding words independently instead (SimpleB2T) makes aggregation across observations and an LLM prior effective, reaching 36.6% word error rate with five observations per word on a clinically motivated benchmark.
The work shows that most of the reported gain from jointly decoding all words in a sentence (d'Ascoli et al., 2025) is reproducible on synthetic signals containing no brain information (22.0% versus 22.3% on real MEG), because neighbouring three-second word-aligned windows overlap and implicitly leak word duration; decoding words independently instead (SimpleB2T) makes aggregation across observations and an LLM prior effective, reaching 36.6% word error rate with five observations per word on a clinically motivated benchmark.
University of Oxford
Dulhan Jayalath（First author）
SimpleB2T Brain-to-Text Decoding Shortcut Learning LLM Rescoring Magnetoencephalography

You’re all caught up

A language model writes an EEG network without a single gradient step: all 2,730 weights by hand, 0.834 balanced accuracy from five subjects

ECG models break on filter phase, not filtering: a causal all-pass shift drops conduction-block AUROC from 0.91 to 0.62

Injection probes were reading the conversation, not the injection: matched negatives cut false alarms from 80% to 17%

Texture-driven visual learning suffers low-frequency shortcuts: pruning low-frequency components raises in-distribution accuracy by up to 10% and out-of-distribution accuracy by up to 40%

Removing the overlapping-window timing shortcut lets non-invasive brain-to-text reach 36.6% word error rate with five observations per word