Surface / morphology
What regularities are already present in the constructed strings, roots, prefixes and tokenization?
Recovery can be real.
Composition can still fail.
The latest VARZIN experiments separate five questions that older site summaries sometimes compressed: surface clustering, frozen representation content, trained remediation, true algebraic-position recovery, and systematic composition on unseen combinations.
Current conclusion
The strongest current evidence is not “models recover LUXVAR meaning” and not “models contain no structure.” It is narrower: targeted structural information can be strongly recoverable or decodable under specified procedures, while the tested systematic-composition criteria remain unmet.
Five questions that must remain separate.
Frozen representation
Can a target variable be decoded from an unmodified model without claiming that it is already used compositionally?
Intervention
What changes after an explicit trainable projection/readout is added while the backbone remains frozen?
True group position
Does recovery track an opaque algebraic-position target beyond a correlated category label, under shuffled-label controls?
Composition
Does the recovered representation support the specified operation on combinations never seen during training?
Frozen models show strong surface capture and weak target recovery in the adversarial setup.
The dedicated Track D record reports Mistral 7B ARI −0.0337 and Llama 3 8B ARI 0.0152 against the designed orbit/axis labels under its Semantic Starvation protocol. The later Morphological Hijacking series broadens the frozen-baseline framing across GPT2-small, Pythia-410m, Mistral-7B-v0.3 and Qwen2.5-7B, with surface-prefix structure dominating the adversarial Core-30 task.
Boundary: this is a frozen/unremediated result under a deliberately adversarial construction. It is not a universal statement that language models cannot encode algebraic or structural variables.
Explicit training changes the result—and the claim.
The projection-head series trains a lightweight readout using supervised contrastive learning, hard negatives, symmetry regularization, rogue-dimension ablation and Soft-PCA initialization while keeping the backbone frozen. Under model-specific tuning, the reported trained stress-test ARI reaches approximately 0.957 for Mistral-7B and 0.880 for Qwen2.5-7B.
Supported
- Targeted intervention can substantially improve recovery of the specified structural target.
- Continuous geometry metrics move consistently with the clustering improvements in the reported experiments.
- The 300-word Appendix E variant reports pooled held-out-root ARI 0.992 ± 0.007 in its easier same-script/same-order condition.
Not implied
- Spontaneous semantic discovery by the frozen backbone.
- A single architecture-independent hyperparameter recipe.
- Natural-language semantic understanding.
- Systematic composition.
The harder Appendix F target is true group position, not just category labels.
The corrected benchmark uses 5 roots × 6 prefixes × 12 positions = 360 words with opaque suffixes and a shuffled-label control. Reported true-label ARI values are approximately 0.791 ± 0.105 for GPT2-small, 0.929 ± 0.047 for Mistral-7B and 0.996 ± 0.004 for Qwen2.5-7B; shuffled-label controls remain much lower.
Why this matters: it is stronger evidence for targeted recovery of the designed group-position variable than a correlated category label. It still does not test closure, homomorphism, or systematic composition.
The latest composition evidence is negative under the tested procedures.
Morphological Hijacking composition pilot
A capacity-controlled MLP fits seen pairs at about 0.996, then falls to TRUE 0.106 on held-out pairs, below SHUFFLED 0.281 and WRONG_OP 0.175. The diagnostic chain tests optimization, conditioning/rank, operand alignment and decoder capacity before retaining the negative generalization result.
Frozen Qwen Phase 0–2 experiment
The frozen Qwen2.5-1.5B structural variable is highly linearly decodable (surface-span mean BA 0.9666; 25/29 indices at 1.000), yet the preregistered composition phase returns 0/58 PASS, 0/58 AMBIGUOUS, 58/58 FAIL.
Current methodological takeaway: strong decodability or targeted recovery should not, by itself, be treated as evidence that the same representation is a computational substrate for systematic compositional generalization.
The public code archives preserve the diagnostic chain.
The verified v3 Zenodo ZIP contains 40 entries, including scripts 01–25 and the paper source/figures. It includes the arbitrary-scale LUXVAR generator, the 300-word experiment, the intentionally preserved rejected leaking position design, the corrected true-position benchmark with shuffled control, and the composition diagnostic sequence. VARZIN V2 separately exposes the final PDF and the large All code Phase.zip package with the frozen dataset/hashes and Phase 0–2 materials.
Older positive wording is preserved, but no longer controls the current claim.
The LUXVAR v2.2 title retains “AI-Recoverable Morpheme Families” for citation continuity. Its corrected manuscript and later audits separate inter-model morphemic/surface agreement from recovery of the intended semantic axes. The old 0.812 inter-model figure is therefore historical context, not the headline current evidence.