Morphological Hijacking in Frozen Language Models: A Contrastive, Symmetry-Regularized Projection Head for Algebraic Structure Recovery — v2

Research question and current reading.

v2 strengthens the intervention result and adds a direct group-position target with a shuffled-label control. The safe claim is targeted group-position classification recovery under this specific synthetic construction—not proof of a homomorphic group representation or systematic composition.

What was actually tested.

• Adversarial LUXVAR Core-30 with algebraic target labels placed in conflict with salient surface prefixes.
• Frozen hidden-state extraction and k-means baseline scored by orbit ARI and Prefix Hijacking Ratio.
• Frozen-backbone attention/projection head trained with supervised contrastive loss, hard negatives, symmetry regularization, rogue-dimension ablation, and Soft-PCA initialization.
• Held-out, cross-script, word-order, and related stress tests; later versions add explicit group-position and composition controls.

Strongest supported findings.

• Frozen baselines recover the adversarial prefix at ceiling while target orbit recovery is near failure in the reported Core-30 setup.
• After model-specific retuning, Mistral-7B and Qwen2.5-7B report trained ARI ≈ 0.957 and 0.880 respectively in the stated stress-test configuration.
• The dedicated opaque-position experiment reports true-label ARI roughly 0.79–1.00 versus shuffled-control ARI roughly 0.02–0.30 across GPT2-small, Mistral-7B, and Qwen2.5-7B.

Negative findings are part of the result.

• The group-position success does not test group composition, closure, or homomorphism.
• The same fixed hyperparameters do not transfer across architectures; key settings were retuned per model.

What this record does not establish.

• The direct position test uses one synthetic construction and a limited set of architectures.
• Shuffled controls are non-zero on the 7B models, reinforcing the need for matched-control interpretation rather than absolute ARI alone.

Full text, code, data, and provenance.

• Zenodo v2 exposes a version-specific research PDF and a version-specific ZIP code/paper package.
• Public package checksum: md5:d5b73639e3ca492ee6ddb60f717f5899.
• The page links the exact versioned DOI rather than silently treating later versions as replacements.

Zenodo REST metadata, public file inventory, sizes, and checksums re-verified 2026-09-14 for v2.

Canonical version-specific record.

Verified repository identity.

Zenodo metadata identifies this item as an open-access preprint under CC BY 4.0.

Keep this deposit distinct from later interpretation.

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.