What Accuracy Hides: Measuring Adapter Drift in CUR-Factored Parameter-Efficient Fine-Tuning

Retention under parameter-efficient fine-tuning is almost always judged by one number: accuracy on a held-out benchmark after the model has been adapted to a new task. This paper is a measurement study of what that number hides. Using CUR-factored adapters as our case study (the update is written $\Delta W = CUR$, with $C$ and $R$ frozen columns and rows of the pretrained weight and only the $r \times r$ core $U$ trained), we find that benchmark accuracy can conceal the behavioural drift it is meant to measure, and that a baseline not matched to the numerical treatment of the models it is compared against can manufacture most of an apparent retention gap. We study CUR adapters with INT8 compression of the frozen factors (31,744 trainable parameters, $137\times$ fewer than LoRA) on Qwen2.5-1.5B-Instruct. First, after fine-tuning both a LoRA and a CUR adapter to \emph{identical} MMLU accuracy, they disagree with the base model on $13.9\%$ and $7.9\%$ of questions respectively: equal accuracy, materially different behaviour, a factor-of-two difference the accuracy column reports as zero. Second, we document a measurement artifact that manufactured $80\%$ of an apparent retention gap in our own earlier run, when the baseline did not match the numerical treatment of the models compared against it. Third, on a two-task retention probe (MRPC $\rightarrow$ SST-2 $\rightarrow$ MRPC, three seeds) with task-specific heads isolating adapter drift, the CUR adapter shows no measurable forgetting on any seed ($-0.003 \pm 0.002$) while LoRA forgets more on all three ,but by a seed-dependent amount ranging from 3 F1 points to a full collapse to the majority-class floor, so the dramatic collapse is itself a single-seed event. The CUR adapter also fits the interfering task 4.4 points less well, an alternative explanation we can characterize but not eliminate. We report where the approach does not win: an MRPC F1 gap to LoRA of about 11.5 points (parameterization, not compression: at a matched learning rate per-row INT8 and FP16 score identically), a peak-training-memory regression, and no saving from 4-bit storage. The MMLU study is single-seed; the retention probe and the RTE and MRPC quantization ablations use three seeds (two for the MRPC INT4 row). We scope our claims accordingly.