arXiv study shows input encodings can impose a measurable floor on interaction prediction
A new arXiv paper introduces a label-free diagnostic that computes an error floor any deterministic decoder inherits from its input encoding, demonstrated on siRNA, splicing and lipid panels.
In plain words
Researchers describe a way to calculate, before any model is trained, how much a chosen input encoding limits a predictor's ability to reproduce measured interactions. In an siRNA case, a graph neural network's feature mask collapsed many measured states into fewer classes, creating a fixed error floor. Restoring a few chemistry columns removed that limit, though refitting barely improved accuracy.
What was announced
The authors compute the attainable contrast space directly from the encoder and contrast design, without labels or predictor fitting, and project measured contrasts onto it to obtain a squared-error floor for any deterministic decoder using that encoding. On an siRNA panel, a graph neural network's training-only feature mask merged 165 measured states into 90 classes, cutting the rank of 140 interaction contrasts to 72; the resulting floor was 0.009980, about 14.63% of recorded contrast energy. Predicted interactions were nearly zero, and their mean squared error of 0.068335 slightly exceeded a zero-prediction control, so the floor added little to fitted error.
The diagnostic's scope is illustrated by two further panels: a released splicing predictor encodes injectively and keeps full design rank of 1,986, giving a zero floor, while an independent count-fingerprint encoder on a lipid panel showed rank 740 across 800 pairwise-distinct encoded contrast rows, exposing 60 restrictions that duplicate-row checks missed. The authors state the method separates representability from learned accuracy but does not by itself predict refitting gains or explain remaining error.
What it does not show
- The evidence is limited to the paper's abstract, so methods, full results and peer-review status cannot be verified from the supplied passages.
- The diagnostic does not by itself predict refitting gains or explain the remaining error, per the authors.
- The reported refitting improvement after removing the mask was only 0.000017 in interaction MSE, so the practical benefit of eliminating the floor was small in this case.
For security and IT teams
- Scope
- The findings concern encoding-imposed representational limits in specific scientific prediction panels (siRNA, splicing, lipid) and do not establish general applicability to other domains.
Sources
Source-support review completed 19 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 19 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.