When Diverse Training Hides Benchmark Leaks

Two findings about large language models pull in opposite directions. Work on dataset diversity says that training on varied data, rather than simply more data, pushes models toward general solutions and away from brittle shortcuts. Work on benchmark contamination says that test questions routinely leak into training data, which inflates scores. It also says the best current leak detector works only because training on a question makes the model's answers to it rigid and repetitive. If diverse training keeps a model's answers varied, then the habit we want for generalization may erase the signal auditors rely on. We do not know whether diverse training makes leaks less harmful, harder to see, or both, and in what proportion.

This proposal measures that trade-off directly. We train small reasoning models on datasets of fixed size but controlled diversity. We plant known benchmark questions into the training data at controlled doses. Each planted question is paired with a never-seen "twin" of equal difficulty. For every diversity level we then measure two quantities: how much the leak inflates accuracy, and whether current detectors can still find it. The result is a set of dose-response curves. They tell practitioners when a high score from a diversely trained model can be trusted. They also yield a detector variant, calibrated against clean twins, built for models whose answers are diverse by design.

Core Claim

We claim that training diversity changes the cost of contamination on two separate axes. The first is the memorization premium, defined as

$$\Delta = \mathrm{Acc}(\text{leaked items}) - \mathrm{Acc}(\text{clean twins}),$$

where both accuracies are measured on the same model, and the twins are items of matched difficulty that never appeared in training. The second is detectability: how well an auditor can tell leaked items from clean ones. We hypothesize the following. Structural diversity, meaning varied problem shapes and solution strategies rather than varied wording, shrinks $\Delta$ for paraphrased leaks. It also lowers the detection power of generation-diversity detectors, and it lowers that power faster than it shrinks $\Delta$. The consequence is a regime of "hidden inflation": a premium that is still several points large while every detector sits near chance.

If the hypothesis is wrong in the reassuring direction, the field learns something just as useful. Detection power and $\Delta$ would then fall together, so that whatever leaks are undetectable are also nearly harmless.

Grounding in Prior Work

Several papers motivate the gap. Detecting RLVR Training Data via Structural Convergence of Reasoning introduces Min-kNN Distance. The method samples 32 answers per prompt and flags a prompt as seen in training when those answers are unusually similar to one another. It reports an average AUC of 0.70, and robustness to paraphrase (0.72 to 0.71). That paper treats low answer diversity as a fixed fingerprint of training. It never asks what happens when training is designed to keep answers diverse.

Diversity-Incentivized Exploration for Versatile Reasoning does exactly that. Its method, DIVER, adds sequence-level diversity rewards to reinforcement learning on math. Here reinforcement learning means learning from verifiable rewards, or RLVR: updating the model with a reward for checkably correct answers. DIVER reports gains on MATH-500, AIME and similar sets. The contamination literature considers those sets leaked, so DIVER's own numbers cannot separate reasoning from varied recall.

On The Fragility of Benchmark Contamination Detection in Reasoning Models constrains our design. It shows that one epoch of clean reinforcement learning after contaminated supervised fine-tuning pushes likelihood detectors such as LiRA toward chance. It also shows that contamination with chain-of-thought data leaves all ten detectors it tests between 50% and 58% AUROC. We therefore cannot treat likelihood detectors as the reference standard, and our pipeline must include a reinforcement learning stage.

When Is Benchmark Contamination Detectable? Information Limits and Power-Calibrated Audits supplies the analytic lens. It shows that audit power scales with $\alpha \rho \sqrt{m}$. Here $\alpha$ is the fraction of the benchmark that leaked, $\rho$ is how separable the model's behavior is on seen versus unseen items, and $m$ is the number of audited items. Our hypothesis amounts to the statement that training diversity lowers $\rho$. We will estimate $\rho$ directly in each condition.

Three papers supply methods we reuse. Measuring Diversity in Synthetic Datasets contributes DCScore, a diversity measure computed by asking how well each sample can be classified apart from the others in embedding space. We use it to select training subsets. DyVal and BeyondBench generate reasoning problems from graphs and solver-verified templates. That lets us vary problem structure deliberately and draw unlimited clean twins.

Rethinking Benchmark and Contamination for Language Models with Rephrased Samples supplies two things. The first is the paraphrase-injection protocol: training on rephrased MMLU test questions lifts Llama-2-13B from about 55 to about 86. The second is the strongest filtering baseline we compare against.

Research Program

We work with Qwen2.5 at 1.5B and 7B parameters, the family used in both the fragility study and Spurious Rewards Paradox. The pipeline has two stages. A supervised fine-tuning stage teaches step-by-step math solutions. A GRPO reinforcement learning stage on verifiable math rewards follows it. This mirrors the multi-stage setting in which the fragility paper found detectors collapse.

The training pool is drawn from OpenR1-Math-220k together with DyVal- and BeyondBench-generated problems. From this pool we build training sets of identical size and compute budget at three or four diversity levels, defined along two separate dimensions. Lexical diversity is varied by paraphrase-augmenting a fixed set of problem structures. Structural diversity is varied by sampling different graph shapes, task families and solution strategies, and we check each level with DCScore over solution embeddings. Separating these dimensions is the core ablation. The diversity literature, and the contrast between DyVal's training gains and BeyondBench's finding that math-specialized tuning hurts algorithmic reasoning, both suggest that structure rather than wording is what matters. That claim has not been tested under contamination. During reinforcement learning we also cross each data condition with three objectives: standard GRPO, token-level entropy regularization, and DIVER.

Into each condition we inject benchmark items at contamination fractions of 0, 1%, 5% and 20% of the target benchmark. Each injection takes one of three forms: verbatim, paraphrased, or answer-only. These are the categories that Shihab et al. found to be ordered in detectability.

We use two kinds of target benchmark, each with exact clean twins. The first is natural. MATH-500 and GSM8K are each randomly split in half, and only one half is ever injected. Because the split is random, the halves match in difficulty in expectation. Since these benchmarks may already be in pre-training data, we take the premium as the difference of each model's leaked-minus-clean gap from the same gap in the zero-injection model. The second is synthetic. We fix a sample of DyVal and BeyondBench instances and inject it, then measure on fresh draws from the same generator settings. This gives a perfect counterfactual that pre-training cannot have touched.

We also build one new tool. Min-kNN uses an absolute similarity threshold, and diversity training raises answer diversity on every item. A model trained for diversity may therefore look "unseen" everywhere. Our twin-calibrated variant instead compares each suspect item's answer similarity with the distribution over difficulty-matched clean twins scored by the same model. It reports a paired statistic rather than a global threshold. This is a small change to an existing detector. Its value is that it is the natural fix if the main hypothesis holds.

Baselines

The naive reference point is the zero-injection model in each diversity condition. It fixes the clean accuracy level and the detector false-positive rate. Every premium and AUC is reported relative to it, and it also measures how much pre-existing contamination of MATH-500 and GSM8K the base model brings in.

The quantity baseline is a homogeneous training set, low in DCScore, scaled to the same compute as the diverse sets by adding more near-duplicate problems. It tests whether any reduction in $\Delta$ comes from diversity rather than from leaked items being diluted by sheer volume. It is run with the same optimizer, number of steps and hyperparameter search budget as the diverse arms. Diversity beats it if the premium falls further at matched compute.

The strongest detection baseline is Min-kNN Distance from Detecting RLVR Training Data via Structural Convergence of Reasoning, with its reported average AUC of 0.70. It is the only published detector built for models trained with reinforcement learning, and the fragility paper rules out the likelihood family. We run it unchanged, with 32 samples per prompt and the same sampling temperature as our twin-calibrated variant. Alongside it we run LiRA and plain loss as the likelihood family. For LiRA, the fragility paper reports 89.1% AUROC after contaminated fine-tuning and a drop of roughly 14% after one reinforcement learning epoch. Our variant beats Min-kNN if its AUC is higher in the diversity-trained conditions, and no worse in the low-diversity conditions where Min-kNN is expected to work.

The strongest diversity-training baseline is DIVER Diversity-Incentivized Exploration for Versatile Reasoning, which reports a six-benchmark math average of 43.1 for Qwen2.5-Math-7B, 1.2 points above entropy-regularized reinforcement learning. We re-run it inside our pipeline under the same rollout budget. The aim is not to beat it on accuracy. It is the condition most likely to produce hidden inflation, because it rewards answer diversity directly.

The strongest prevention baseline is the LLM decontaminator from Rethinking Benchmark and Contamination for Language Models with Rephrased Samples. It retrieves training items by embedding similarity and then has a judge model decide whether each is a rephrased test item, with reported F1 between 0.94 and 1.0. We apply it to every contaminated training set and record how much of $\Delta$ remains after filtering.

The key ablation compares lexical against structural diversity at identical DCScore where possible. It isolates which kind of diversity drives any change in $\Delta$ and in detectability.

Why not just…?

Why not just decontaminate the training data with an LLM filter? For data you control, you should, and we report how much of the premium the filter removes. The filter answers a different question from ours. It needs the test set in hand, and it cannot help someone auditing a released model whose training data, including pre-training and distillation sources, they cannot see. That is the situation in which detectors matter. Our measurement of residual premium after filtering will also show whether paraphrases the filter misses behave differently under diverse training.

Why not just evaluate on dynamic benchmarks and ignore detection? This is a real threat to the detection half of the proposal. Dynamic suites such as LiveBench and BeyondBench sidestep contamination by construction. BeyondBench reports an open-source ceiling of 30–35% on its Hard suite, which shows how much static scores overstate ability. These suites only cover verifiable domains, though, and they do not say how much of a given static score was inflated. We include them as generalization endpoints. If diversity training lifts BeyondBench and DyVal scores while $\Delta$ stays negligible, we will conclude that the honest recommendation is "trust dynamic benchmarks, stop auditing," and say so.

Why not just add more data instead of diverse data? The quantity arm settles this directly at matched compute. Evidence from Beyond Quantity: Trajectory Diversity Scaling for Code Agents is relevant: 500 diverse samples reached 36.66% on BFCL Multi-turn against 35.91% for Qwen3-Coder-480B-A35B-Instruct. That evidence concerns generalization, not leaks, which is why we test it here.

Why not just raise sampling temperature or add token-level entropy? DIVER's authors report that token-level entropy adds surface noise without producing different strategies. We still include the entropy-regularized arm. If entropy alone lowers Min-kNN's AUC as much as structural diversity does, then the detector is fooled by surface noise. That would make the twin-calibrated fix more important, not less.

Proposed Experiments and Decisive Checks

The primary result is a grid. For each diversity level and reinforcement learning objective, we plot the premium $\Delta$ against injection dose and form, for MATH-500 halves, GSM8K halves, and synthetic DyVal/BeyondBench twins. Next to each premium we report AUC for loss, LiRA, Min-kNN and the twin-calibrated variant. From the AUCs we estimate $\rho$ and check it against the $\alpha \rho \sqrt{m}$ power prediction. For generalization, every condition is scored on LiveBench math and reasoning categories, the BeyondBench Easy and Hard suites, and fresh DyVal problems at complexity levels D1–D4.

The decisive check for the main claim is whether any condition shows a premium of at least three accuracy points, on paraphrased leaks at 5% dose, while Min-kNN's AUC is below 0.6. That combination is hidden inflation. If it occurs only under structural diversity and not lexical diversity, the core ablation explains the cause.

A second check tests the reassuring alternative. Across all conditions, we fit the relation between $\Delta$ and detector AUC. A tight monotone fit through the origin would mean undetectable leaks are harmless.

A third check tests the twin-calibrated detector. It must recover AUC in the high-diversity arms without losing ground in the homogeneous arms. If it cannot, we conclude that generation-diversity auditing has a limit set by training diversity.

As a sanity anchor, the verbatim, homogeneous, supervised-fine-tuning-only condition should reproduce the fragility paper's pattern. That means a sizable premium (they report 8.8% average pass@1 inflation) and high LiRA AUROC that falls after reinforcement learning. If we cannot reproduce that, the rest of the grid is not interpretable.