Fine-Tuning Shifts Form Before Competence: Treatment-Correlated Style Drift Confounds Surface-Form Evaluation of Adapter Specialization.

Abstract

Stacking task adapters on a frozen domain adapter is a common approach for modular specialization of large language models, usually justified by knowledge injection; whether the surface-form graders standard in this setting can measure such specialization validly has remained unexamined. Two challenges follow: isolating the domain layer's marginal contribution under identical task training, and determining whether measured differences reflect competence or output form. In this work, both are addressed with a matched-arm DoRA ablation on a frozen Qwen 3.5 9B hybrid: a 242 M-parameter domain expert is trained on a licensed 27.8 M-token corpus, and byte-identical task adapters are trained flat, stacked (DFE-stacking), and jointly across three tasks and five conditions, with the key extraction pair replicated at a second training seed; every grader is audited by per-item failure classification and a dual-rater blind human re-grade with consensus adjudication. The principal finding is that fine-tuning shifts output form before competence: stacking yields no seen-task benefit; an apparent generalization collapse on an unseen schema reduces to a single, highly seed-stable formatting coercion that the domain layer dampens and task-mixture training eliminates, and most automated keyword-grading failures prove to be correct paraphrases, leaving no genuine capability loss detected on any instrument. Because such style drift is induced by the training condition itself, surface-sensitive grader error becomes bias aligned with experimental arms rather than noise. Three additional experiments extend the findings: the coercion replicates on Llama 3.1 8B (60% vs. 19% untuned); a pre-registered 240-item formatting-perturbation battery across four further unseen schemas and eight conventions reveals fine-tuned conditions ranking lowest under strict grading yet highest under format-normalized grading; and a matched-compute wrong-domain control shows the domain layer's coercion dampening to be generic substrate regularization, while value integrity under the induced drift tracked the substrate's domain. The principal contribution lies in the identification and quantification of this confound: surface-form evaluation of adapter specialization is structurally confounded, and per-item audits of the type released herein are a necessary component of specialization studies.