10.5281/zenodo.22258644
Nirouyar, R. (2026). Morphological Hijacking in Frozen Language Models: A Contrastive, Symmetry-Regularized Projection Head for Algebraic Structure Recovery (Version v1). Zenodo.
Morphological Hijacking in Frozen Language Models: A Contrastive, Symmetry-Regularized Projection Head for Algebraic Structure Recovery
Projection-head intervention record; explicit intervention is distinct from spontaneous recovery in an unmodified frozen model.
Evidence boundary
A Zenodo deposit is a citable public record; it does not by itself imply peer review, independent validation, or broader empirical generalization.
What this record documents.
Version v1 introduces the “morphological hijacking” line of experiments on frozen autoregressive language models. The deposited description reports that surface-correlated structure can dominate clustering even when a stronger task-relevant algebraic structure is available, and documents a trainable projection-head intervention intended to recover that structure. The record also reports negative findings alongside the intervention results, making the distinction between frozen-model behavior and trained remediation central to its interpretation.
Research question
When surface morphology conflicts with a task-relevant algebraic structure, what do frozen model representations recover, and can an explicit trainable intervention change that geometry?
This question is interpreted within the documented model, dataset, implementation, and evaluation procedure; it is not generalized beyond that evidence.
Canonical record metadata.
Repository metadata, checked against Zenodo.
Nirouyar, Reza
Normalized VARZIN researcher identity: Reza Nirouyar · ORCID 0009-0000-4690-6842.
2026-09-02
Version v1 · Preprint · open access.
CC BY 4.0
Reported only when present in verified Zenodo metadata.
Zenodo metadata identifies this item as an open-access preprint published on 2026-09-02 under CC BY 4.0.
What this record adds to the research trail.
The record frames morphological hijacking as a representational failure mode and tests an explicit projection-head remediation. Any post-training recovery belongs to the intervention result; it is not evidence that the unmodified frozen model spontaneously encoded the target structure in a directly usable form.
Version v1 is preserved as the first public version-specific record in this series. Later v2 and v3 deposits are separate citable records rather than silent replacements.
Keep the record distinct from later interpretation.
Projection-head intervention record; explicit intervention is distinct from spontaneous recovery in an unmodified frozen model.
The VARZIN portal preserves the version-specific DOI, deposited creator string, publication date, and repository identity while allowing later project pages to state narrower interpretations, corrections, or negative results explicitly.