10.5281/zenodo.22262388
Nirouyar, R. (2026). Morphological Hijacking in Frozen Language Models: A Contrastive, Symmetry-Regularized Projection Head for Algebraic Structure Recovery (Version v2). Zenodo.
Second public version of the projection-head research record; preserve version-specific DOI citation.
A Zenodo deposit is a citable public record; it does not by itself imply peer review, independent validation, or broader empirical generalization.
Version v2 expands the morphological-hijacking study around an adversarial LUXVAR Core-30 lexicon in which salient surface form is deliberately placed in conflict with algebraic orbit structure. The deposited description reports near-total prefix recovery under unsupervised clustering while the target orbit structure is not recovered in the frozen baselines, then investigates anisotropy, token-length effects, and trained remediation. The record therefore combines a negative frozen-model result with a separate intervention result.
This question is interpreted within the documented model, dataset, implementation, and evaluation procedure; it is not generalized beyond that evidence.
Nirouyar, R. (2026). Morphological Hijacking in Frozen Language Models: A Contrastive, Symmetry-Regularized Projection Head for Algebraic Structure Recovery (Version v2). Zenodo.
Normalized VARZIN researcher identity: Reza Nirouyar · ORCID 0009-0000-4690-6842.
Version v2 · Preprint · open access.
Reported only when present in verified Zenodo metadata.
Zenodo metadata identifies this item as an open-access preprint under CC BY 4.0.
The v2 record strengthens the separation between surface-form clustering and algebraic target recovery, while adding diagnostics intended to explain the failure mode. The project does not treat trained recovery as equivalent to spontaneous structure recovery.
Version v2 is a distinct public deposit dated 2026-09-02. It should be cited separately from v1 and v3 whenever version-specific methods or results are discussed.
Second public version of the projection-head research record; preserve version-specific DOI citation.
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.