Neural-field readers stay accurate yet shift predictions when weights change equivalently
An arXiv abstract reports that visual recognition models trained on neural-field weights can be highly accurate yet sensitive to function-preserving weight transformations.
In plain words
A neural field can store the same image using different sets of internal numbers (weights). The paper's authors found that a reader trained to interpret these weights often depends on the exact numbers rather than the image content they encode. When the weights were transformed in ways that preserve the image, predictions changed even though accuracy barely moved.
What was announced
The authors measured recognition, prediction disagreement, and numerical fidelity together across shape, digit, and natural-image fields. On 10,000 official CIFAR10 test origins, one-period hidden-bias shifts changed 7.3% of a released HiddenProbe configuration's predictions while barely changing accuracy; three-period shifts changed 22.0% of predictions and reduced accuracy from 57.8% to 54.4%. Output-only ProbeGen remained unchanged. The abstract notes these effects were smaller than in a separate adapted-reader study, where matched phase and sign controls removed the corresponding sensitivities.
Controlled readout experiments suggested why invariance alone is insufficient: complete canonical hidden features reached 38.7% on MNIST, while explicitly composing those activations, outgoing weights, and output bias into field outputs reached 81.5%. The authors conclude the information is present but that the learned readout does not recover it as effectively. A separate 720-response study of language models reading numerical weights indicated that observed accuracy losses need not yield rejection of response-law equality, so stochastic audits require repeats on both inputs.
What it does not show
- The evidence is an arXiv abstract only; full methods, experimental details, and peer review status are not available in the supplied passages.
- The adapted-reader study and the 720-response language-model study are described only in summary; their designs and results cannot be independently verified from the excerpt.
- No independent confirmation of the reported figures is provided.
Sources
Source-support review completed 17 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 17 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.