30 words
Adversarial reference lexicon for the frozen-model morphology/prefix-hijacking experiments.
Morphological Hijacking in Frozen Language Models: A Contrastive, Symmetry-Regularized Projection Head for Algebraic Structure Recovery — v3
v3 preserves the positive targeted recovery result while adding the harder composition question. The central v3 boundary is that classification/position recovery did not become systematic composition on unseen ordered pairs under the tested pipeline.
Evidence status
Intervention study: frozen-model failure and trained remediation must remain separate claims. Publication status: Preprint; Zenodo record explicitly states not yet peer-reviewed.
Research question and current reading.
Research question
After remediation improves recovery of the target categories or geometry, does that recovered structure support the algebraic composition required for systematic generalization?
v3 preserves the positive targeted recovery result while adding the harder composition question. The central v3 boundary is that classification/position recovery did not become systematic composition on unseen ordered pairs under the tested pipeline.
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.
- The version retains the trained-remediation results from the earlier series and the explicit true-position control from v2.
- In the v3 composition pilot, a capacity-controlled MLP fits seen pairs almost perfectly (mean ≈ 0.996) but fails on 24 held-out ordered pairs.
- Held-out TRUE accuracy is 0.106, below SHUFFLED 0.281 and WRONG_OP 0.175 in the public v3 manuscript, consistent with memorization/interpolation rather than demonstrated systematic composition.
How this experiment relates to Core-30 and later scaled benchmarks.
20 roots × 15 prefixes = 300 words
Appendix E scale/generalization test. The pooled held-out-root result is ARI 0.992 ± 0.007, but under the easier same-script/same-order condition, not the hardest cross-script stress test.
5 roots × 6 prefixes × 12 positions = 360 words
Appendix F benchmark for targeted recovery of true Z12 group position with opaque suffixes and a shuffled-label control; classification success does not establish composition.
Scale boundary: Core-30 remains a canonical reference subset. Larger generated corpora and later benchmark datasets are distinct experimental artifacts and should not be silently relabeled as a single enlarged canonical lexicon.
Negative findings are part of the result.
- The composition pilot remains negative after diagnostics for optimization, numerical rank/conditioning, operand-space alignment, and decoder capacity.
- The result does not establish that composition information is universally absent; it is a scoped failure of the tested construction/readouts.
- The raw-margin layer-selection heuristic and the broad script-invariant anisotropy hypothesis are also reported as failed or narrowed hypotheses.
What this record does not establish.
- The composition conclusion is specific to the marker construction, extraction/preprocessing, tested alignment strategies, and tested decoders.
- The uploaded 21-page manuscript supplied in this chat is not byte-identical to the v3 Zenodo PDF, so this webpage is grounded in the verified public v3 Zenodo manuscript/description rather than asserting byte identity for that copy.
Boundary: a Zenodo deposit is a citable public record; it does not by itself imply peer review, independent validation, causal inference, or generalization beyond the documented experiment.
Full text, code, data, and provenance.
- Zenodo v3 exposes a version-specific research PDF and a version-specific ZIP code/paper package.
- Public package checksum: md5:847d2d8c0a7c46f277458916eefe8180.
- The page links the exact versioned DOI rather than silently treating later versions as replacements.
- Verified ZIP inventory contains 40 entries, including the paper source/figures plus scripts 01–25 for the reported experimental chain.
- Scale-specific code is present: 17_luxvar_language_generator.py builds arbitrary-scale LUXVAR variants; 18_full_experiment_expanded_luxvar.py runs the Appendix E 300-word test.
- Algebraic-control code is present: 19_true_algebraic_axis_recovery_LEAKED_rejected.py preserves the rejected leaking design, 21_FINAL_true_axis_recovery_all_models_with_control.py produces the definitive Appendix F group-position result.
- Composition diagnostic code is present: scripts 22–25 load the frozen marker artifact and culminate in 25_composition_pilot_v031_and_D4_D4L3B.py, the source for the seen-pair 0.996 / held-out 0.106 negative result.
LUXVAR_morphological_hijacking.pdf
paper · 332.49 KB · md5:0bcd538f588986a7927c61a2028173a3
luxvar_code_and_paper.zip
code-package · 444.29 KB · md5:847d2d8c0a7c46f277458916eefe8180
Zenodo REST metadata, public file inventory, sizes, and checksums re-verified 2026-09-14 for v3.
Canonical version-specific record.
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.
Verified repository identity.
Nirouyar, Reza
Normalized project identity: Reza Nirouyar · ORCID 0009-0000-4690-6842.
2026-09-03
Version v3 · 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 2026-09-03 under CC BY 4.0.
Keep this deposit distinct from later interpretation.
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.