Fine-tuned language models often score well on a benchmark because paraphrases of its test questions leaked into their training data. Two findings appear to offer cheap protection. Low-rank adaptation (LoRA) fine-tunes a model by training only two small matrices whose product is added to each frozen weight, and it is reported to barely memorize its training data. Large training batches are reported to "dilute" leaked examples and reduce how much the model profits from them. We think both protections may come from one simpler cause: how far a single leaked example moves the model each time it is seen. In plain SGD, one occurrence of an example in a batch of size $B$ moves the weights by $\eta/B$ times its gradient, where $\eta$ is the learning rate. Standard LoRA starts with one adapter matrix at zero and uses one shared learning rate for both matrices, which makes its early updates unusually small. If per-example step size is what matters, then LoRA's leak resistance and the batch-size protection are not structural safeguards. They would disappear as soon as practitioners follow the linear scaling rule (multiply the learning rate by $k$ when the batch grows by $k$) or adopt better-tuned LoRA variants.
The stakes are practical. Many leaderboard models are LoRA fine-tunes; Platypus trained its 13B and 70B models with LoRA and relied on heuristic decontamination filters. Current "LoRA is leak-resistant" evidence was gathered at fixed hyperparameters. If that resistance is really just a small effective step, then the community's move toward better-scaled adapters will quietly increase how much benchmark scores are inflated by contamination. A result showing the opposite would also matter: if the rank bottleneck genuinely blocks contamination even at matched step size, it could be recommended as a deliberate defense.
Core Claim
In LoRA fine-tuning, how much a model profits from contamination depends on the measured per-exposure influence of a leaked example, not on batch size or rank as such. Here "profits" is exploitation, defined following Magar and Schwartz as accuracy on leaked test items minus accuracy on matched unleaked test items. We define per-exposure influence as the average drop in the model's loss on a leaked item caused by one gradient step that contains it. Under plain SGD, this influence is governed by
$$\Delta\theta_x = -\frac{\eta}{B}\,\nabla_\theta \ell(x;\theta),$$
where $\theta$ are the trainable parameters and $\ell$ is the loss on example $x$. Two concrete predictions follow. First, when $\eta$ is scaled linearly with $B$, the protection from large batches reported by Magar and Schwartz should vanish. They changed $B$ at fixed $\eta$, which confounds dilution with lowering $\eta/B$. Second, when we compare standard LoRA, LoRA+ and the rank-invariant μA parameterizations, their exploitation curves should collapse onto one curve when plotted against measured per-exposure influence. The comparison must be made at matched out-of-distribution accuracy. If the curves collapse, LoRA's reported leak resistance is a step-size artifact. If LoRA stays below full fine-tuning at matched influence and matched clean utility, the rank bottleneck is doing real work.
Research Program
The work rests on a controlled contamination setup for autoregressive models, extending the design of Magar and Schwartz from BERT pre-training to instruction fine-tuning. We split a benchmark's test set randomly into a "leaked" half and a "clean" half. Paraphrased versions of leaked-half items are then injected into the fine-tuning mixture at controlled rates and duplication counts. The paraphrases are in-distribution contamination in the sense of DICE, meaning rewordings with altered numbers. Exploitation is the accuracy gap between the two halves. Because the split is random, this gap cannot come from question difficulty.
Three factors are crossed. The first is batch size, under three learning-rate policies: fixed $\eta$ (the Magar and Schwartz regime), $\eta \propto B$ (linear scaling), and $\eta \propto B^{p}$ with $p$ swept. The temperature theory of Jastrzebski et al. and the batch-invariance results of Nacson et al. were derived for SGD. The research file flags that AdamW's per-coordinate normalization changes how gradient noise and batch size interact. We therefore run SGD and AdamW arms, and for AdamW we measure which exponent $p$ makes exploitation batch-invariant rather than assuming $p=1$.
The second factor is adapter parameterization. We compare standard LoRA with a shared learning rate, LoRA+ with its recommended ratio of 16 between the two matrices' learning rates, and μA with the second matrix initialized randomly and a scaling factor of 1. We also include full fine-tuning. Chen et al. report that the μA setting makes the optimal learning rate rank-invariant and transferable to full fine-tuning. The third factor is rank, from 4 to 256.
Throughout, we log per-exposure influence directly as the loss drop on leaked items across steps containing them. This measured quantity, not a theoretical formula, is the x-axis on which collapse is tested. It works the same way for Adam and for any parameterization.
The analysis compares methods at matched utility, because a method that simply learns less will also exploit less. For every run we record clean-half accuracy and out-of-distribution accuracy alongside exploitation. We then compare methods on the frontier of exploitation versus out-of-distribution accuracy, never at a single hyperparameter setting.
A secondary question comes from CANARY. CANARY reports that rank-4 LoRA partially evades hidden-state contamination auditing, with AUROC around 0.83. We will test whether that evasion also tracks per-exposure influence, using a DICE-style late-layer probe. If it does, the evasion comes from undertrained adapters rather than from the low-rank constraint.
Baselines
The naive reference point is standard LoRA with the learning rate held fixed as batch size varies. This reproduces in the LoRA setting the regime in which Magar and Schwartz observed batch-size protection. We will also include a contamination-free run at each configuration, which gives the zero point for exploitation and the clean utility ceiling. Full fine-tuning at a tuned learning rate is the reference for "maximal leakage." Wang and Li and the federated memorization study report that LoRA memorizes far less than full fine-tuning; the latter cites up to a 10x reduction in verbatim canary reproduction. Neither measured benchmark exploitation, and we found no reported LoRA exploitation result on GSM8K or MMLU on Lacuna.
The strongest existing mitigation for this question is low-rank LoRA with dropout. Wang and Li report that this combination limits extreme semantic similarity to training data without hurting ROUGE-L. We also include Flat-LoRA, which injects random weight noise to find minima that are flat in the full weight space. It is the strongest published method aimed at robustness under corrupted training data, and it reports a 3.18% GSM8K accuracy gain over standard LoRA on Llama 2-7B. On the utility side, LoRA+ reports a 1.3% MMLU gain over best-tuned LoRA on Llama-7B. For contamination strength, DICE reports that 2% in-distribution contamination yields roughly a 10% absolute jump on GSM8K. Our naive arm should reproduce an effect of that size before any other comparison is trusted.
For fairness, every method sees identical data mixtures, contamination seeds and token budgets. Every method also gets the same number of learning-rate trials, chosen by clean-half validation accuracy, never by exploitation.
A mitigation "beats" another only if it gives lower exploitation at out-of-distribution accuracy within one standard error across seeds. For our own claim, the decisive result is different. The claim is confirmed if the parameterization and batch-size arms collapse onto one exploitation-versus-influence curve, with spread across arms no larger than seed variance. It is refuted if standard LoRA sits clearly below that curve.
The ablation that isolates the key ingredient keeps $\eta/B$ fixed and changes only how duplicated leaked copies are placed. In one arm all copies go in the same batch; in the other they are spread across batches. Batch dilution predicts little difference, because the copies are averaged away either way. The per-exposure account predicts that clustering matters differently from spreading. This is because Adam's normalization, and the size of the parameter change each time a copy is seen, depend on placement.
Why not just…?
Why not just decontaminate the fine-tuning data? Platypus did exactly this with keyword and embedding-similarity filters. However, DICE reports that n-gram, perplexity and Min-k style detectors perform near chance on paraphrased, number-altered contamination, the same kind that produced its 10% inflation. MMLU-CF found about a 14% average accuracy drop once questions were rephrased and closed, despite prior decontamination claims. Filtering is complementary to this work, not a substitute. We will still report how much of our injected contamination an embedding filter catches, so readers can see what gets through.
Why not just trust that LoRA doesn't memorize? This is the claim under test, and it is a real possibility that it holds. The federated study already shows it weakening at rank 1024. The existing evidence also comes from verbatim reproduction of private text at fixed hyperparameters, not from benchmark exploitation, and it predates better-scaled adapters. If standard LoRA sits below the collapse curve at matched utility in our experiments, we will report that the bottleneck protects, and the proposal's central hypothesis fails.
Why not just say that any method that learns less leaks less? This is the most serious objection. It could make the whole finding trivial. That is why every comparison is made on the frontier of exploitation versus out-of-distribution accuracy, never at single settings. The non-trivial question is whether, at equal useful learning, the step-size knobs and the rank knob trade off utility against leakage differently.
Why not just use Flat-LoRA or sharpness-aware training? Flat-LoRA targets brittleness from sharp minima, not the profit a model gets from specific leaked items. There is no evidence in the file that flatness reduces exploitation. We run it on the same frontier. If it dominates, flatness rather than step size is the lever, and we will say so.
Proposed Experiments
Models are Llama-3.2-1B and Qwen2.5-3B, which Chen et al. used to validate μA. The primary domain is GSM8K. The fine-tuning mix is its training set plus paraphrased leaked-half test items at rates of 0, 1, 2 and 5%, with 1, 5 and 20 duplicates. We measure exploitation on the GSM8K halves and on GSM-hard, and out-of-distribution accuracy on MATH, SVAMP and ASDiv.
The second domain is MMLU, using the same split-half design. Clean utility is additionally checked on the MMLU-CF validation split. That split does not share questions with MMLU, so it shows whether the model gained real knowledge rather than benefiting from the leak.
Batch sizes span 8 to 512 under the three learning-rate policies. Every configuration is run with three seeds.
Three results decide the question. The first is whether exploitation stays flat across batch size under linear (SGD) or best-exponent (AdamW) scaling. If it does, the dilution explanation of Magar and Schwartz is refuted in this setting. The second is whether exploitation versus measured per-exposure influence collapses across standard LoRA, LoRA+, μA, the ranks and full fine-tuning at matched out-of-distribution accuracy. The third is whether the drop in probe AUROC for low-rank adapters, the evasion effect CANARY reported, disappears once influence is matched.
References
- Platypus: Quick, Cheap, and Powerful Refinement of LLMs
- Data Contamination: From Memorization to Exploitation
- LoRA+: Efficient Low Rank Adaptation of Large Models
- Learning Rate Scaling across LoRA Ranks and Transfer to Full Finetuning
- DICE: Detecting In-distribution Contamination in LLM's Fine-tuning Phase for Math Reasoning
- Finding Flatter Minima with SGD
- Stochastic Gradient Descent on Separable Data: Exact Convergence with a Fixed Learning Rate
- CANARY: Zero-Label Detection of Fine-Tuning Contamination in Language Models
- Leaner Training, Lower Leakage: Revisiting Memorization in LLM Fine-Tuning with LoRA
- Mitigating Unintended Memorization with LoRA in Federated Learning for LLMs
- Flat-LoRA: Low-Rank Adaptation over a Flat Loss Landscape
- MMLU-CF: A Contamination-free Multi-task Language Understanding Benchmark