10.5281/zenodo.22287006
Nirouyar, R. (2026). Morphological Hijacking in Frozen Language Models: A Contrastive, Symmetry-Regularized Projection Head for Algebraic Structure Recovery (Version v3). Zenodo.
Third public version of the projection-head research record; the versioned DOI is retained separately for provenance.
A Zenodo deposit is a citable public record; it does not by itself imply peer review, independent validation, or broader empirical generalization.
Version v3 extends the projection-head study with a stronger emphasis on whether recovered category structure supports true algebraic composition rather than only classification or clustering. The deposited description explicitly distinguishes structure recovery from composition and adds a diagnostic pilot aimed at that harder question. This version therefore sits at the boundary between representation recovery and systematic generalization, a distinction that later VARZIN V2 work develops further.
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 v3). Zenodo.
Normalized VARZIN researcher identity: Reza Nirouyar · ORCID 0009-0000-4690-6842.
Version v3 · Preprint · open access.
Reported only when present in verified Zenodo metadata.
Zenodo metadata identifies this item as an open-access preprint published on 2026-09-03 under CC BY 4.0.
The v3 record makes composition a separate test rather than assuming it follows from category recovery. This preserves an important negative/limiting result: improved representation recovery is not, by itself, proof of compositional competence.
Version v3 is the third version-specific public deposit in the projection-head series and should be cited independently from v1 and v2.
Third public version of the projection-head research record; the versioned DOI is retained separately for provenance.
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.