Contamination Hides at the Point, Not in the Neighborhood: Auditing Reasoning Models with Weight-Noise Stability

Benchmark contamination means test questions leaked into a model's training data, so its score reflects recall rather than skill. The usual way to detect it is to ask whether the model finds the leaked items unusually likely. That signal can be removed. Reinforcement learning (RL) fine-tuning of reasoning models, using methods such as GRPO that clip how far each update can move the model's output probabilities, closes the likelihood gap between seen and unseen items while the inflated score remains. After this kind of training, ten standard detectors fall to near-chance accuracy. This proposal tests a different question. Instead of asking how confident the trained model is on an item, ask how quickly that confidence collapses when the model's weights are jittered with small random noise. The hypothesis is that memorized items sit on narrow ridges in weight space, while items the model genuinely solves sit on broad plateaus. RL that only reshapes the outputs at the trained weights may leave those ridges intact.

Learning theory already has a name for the quantity being measured. PAC-Bayesian bounds are generalization guarantees that hold for a randomized model, meaning a distribution of weight vectors drawn at prediction time. Turning them into guarantees for the single deterministic model people actually deploy costs something, called the stochastic-to-deterministic conversion gap. Theorists treat this gap as slack to be minimized. Here it becomes the measurement: on each benchmark item, how much better is the deterministic model than its noisy neighbors? The proposal is to compute this per-item gap on a reasoning model, test whether it survives RL cloaking, and check whether it gives benchmark-level audits enough statistical power on small frontier benchmarks.

Core Claim

For a trained model with weights $\theta$, a benchmark item $x$, and the model's own deterministic chain-of-thought plus answer $y$, define the per-item conversion gap at noise scale $\sigma$ as

$$ g_x(\sigma) = \log p_\theta(y\mid x) - \mathbb{E}_{\epsilon \sim \mathcal{N}(0,\ \sigma^2 D_\theta)}\big[\log p_{\theta+\epsilon}(y\mid x)\big], $$

where $D_\theta$ is a diagonal matrix that scales noise relative to each neuron's incoming-weight norm. This makes the measurement invariant to rescaling weights.

The contamination score $s(x)$ is $g_x(\sigma)$ averaged over a small grid of $\sigma$ values. It is then residualized against $\log p_\theta(y\mid x)$ and against accuracy on clean items of matched difficulty. Without that residualization, the score would be ordinary likelihood under another name.

The claim is that after supervised contamination followed by clipped RL, $s(x)$ separates contaminated from clean items substantially better than likelihood-based detectors do. A theoretical argument makes this plausible. The margin-based derandomization results in On Margins and Derandomisation in PAC-Bayes bound how often a noisy predictor disagrees with its mean by the ratio of noise scale to output margin, divided by the model's local sensitivity to its weights. Clipped RL acts directly on the output margin at $\theta$. Nothing in the clipping objective constrains the curvature that determines how fast that margin erodes in a neighborhood of $\theta$.

How the Prior Work Shapes This

Two papers motivate the gap. On The Fragility of Benchmark Contamination Detection in Reasoning Models shows that one GRPO epoch cuts LiRA detection AUROC by about 14% and loss-based detection by about 16%. It also shows that chain-of-thought contamination drives ten detectors to 50–58% AUROC, while inflating DeepSeek-R1-Distill-Llama-8B's pass@1 by 11.7%. It studies only output-probability detectors and never examines weight-space geometry, which is the gap this proposal fills. Data Contamination: From Memorization to Exploitation shows that memorization and score inflation come apart (about 15% memorization with under 1% exploitation, rising to 40% exploitation at 200 duplicates). That is a warning: a sharpness signal may track memorization that does not affect scores. The evaluation therefore measures detection of score-inflating items, not merely seen items.

Three papers supply the method. Normalized Flat Minima derives per-neuron-normalized PAC-Bayes sharpness. That sharpness rises monotonically with random-label memorization in small vision networks and correlates at 0.92–0.98 with the generalization gap on CIFAR-10. The noise scaling $D_\theta$ is borrowed from it, but applied per item on an autoregressive model rather than as one global number. Its Hessian machinery is replaced with Monte Carlo forward passes, because second-order approximations are unreliable and expensive at 8B parameters. A general framework for the practical disintegration of PAC-Bayesian bounds and Non-Vacuous Generalisation Bounds for Shallow Neural Networks show how to relate a single deterministic network to its randomized neighbors. They are used as interpretive lenses, not as sources of bounds; no certificate is claimed for an 8B model. A PAC-Bayesian Link Between Generalisation and Flat Minima controls complexity with expected gradient norms under noise. This motivates a cheap variant that replaces the expectation in $g_x$ with a single-draw gradient-norm estimate.

The closest existing method is EPGS from From Flat Facts to Sharp Hallucinations. EPGS adds Gaussian noise to entity-token embeddings and measures how much the gradient changes, arguing through an input–parameter equivalence that this mimics weight noise. It reaches 0.7373 AUROC against 0.58 for semantic entropy on a "stubborn error" subset of SQuAD with Llama-2-7B. The proposal differs from EPGS in three ways. It targets contamination rather than hallucination. It perturbs weights directly and uses only forward passes, with no named-entity extraction, so it applies to long math and science reasoning traces that have few named entities. And it tests survival under RL cloaking, which EPGS never examines. EPGS also notes a complication that this proposal faces too: rare true facts the model has memorized look sharp. The design handles this by evaluating against contaminated-versus-clean labels with difficulty matched, rather than assuming sharpness implies leakage.

Research Program

The ground truth comes from re-creating the contamination pipeline of the Fragility paper. Starting from DeepSeek-R1-Distill-Llama-8B, a random half of the items from GPQA Diamond, OlympiadBench's text-only problems, and AIME 2024 is placed into a supervised fine-tuning set. Two forms are used: question-plus-answer only, and question-plus-full-chain-of-thought. Then one to three epochs of GRPO run on clean math data. Snapshots are taken after supervised contamination, after each RL epoch, and for a never-contaminated control trained on the same schedule. This yields item-level labels, a cloaking dose (RL epochs), and a contamination type. Because the paraphrase result in When Is Benchmark Contamination Detectable? predicts weaker signals for weaker exposure, a third arm injects paraphrased questions.

For each item, the method draws $K$ noise vectors at each of four or five scales. The scales are chosen so that clean-item accuracy degrades from about 100% to about 50% of its unperturbed value. For each draw, one teacher-forced forward pass scores the fixed deterministic trace $y$. This keeps the cost at roughly $K$ forward passes per item rather than $K$ regenerations of a long reasoning trace. On a subset, the method also regenerates answers under noise. This checks that the teacher-forced gap predicts the accuracy gap, meaning the deterministic answer is correct while the noisy models' answers are not, which is what matters for score validity.

The analysis has three parts. First, item-level detection AUROC for $s(x)$ as a function of RL dose. Second, a decomposition of where along the trace the gap concentrates (question restatement, intermediate steps, final answer). This tests the Fragility paper's account that chain-of-thought contamination spreads the memorization signal across long traces. Third, a benchmark-level audit. Using the sparse-mixture model of Shihab et al., in which detectability scales with $\alpha \rho \sqrt{m}$ (contamination fraction times behavioral separability times audit size), the method estimates the separability $\rho$ for each detector. From that it computes the smallest contamination fraction detectable at 80% power on 198-item GPQA Diamond, and whether the AIME 2024 versus AIME 2025 contrast from MathArena is visible in public models without injection.

Baselines

The naive reference point is raw sequence loss at the unperturbed weights. This is exactly $s(x)$ with $\sigma = 0$, so any improvement over it is attributable to the neighborhood measurement. The Fragility paper reports that loss-based AUROC drops by about 16% after one GRPO epoch.

The strongest existing reference-free detector is Min-K% Prob from Detecting Pretraining Data from Large Language Models. It averages the log-likelihood of the lowest-probability tokens and reported a 12.2% gain over earlier baselines on deliberately leaked benchmarks. LiRA, which compares against shadow reference models, is the strongest reference-based detector. On contaminated reasoning models, the reported number to beat for this whole class is 50–58% AUROC across ten detectors including Min-K%, LiRA, and loss (Fragility paper). Both baselines are rerun on the proposal's own checkpoints, with the same items and with $k$ and the LiRA reference-model count tuned on a held-out injected split. The tuning budget in GPU-hours matches the budget used to choose the proposal's $\sigma$ grid and $K$.

EPGS is adapted to reasoning traces by applying embedding noise to all answer tokens, since entity extraction does not suit math. Its reported 0.7373 AUROC is on a hallucination task, and no reported result was found on Lacuna for EPGS on contamination. It is run at matched forward-and-backward compute.

A black-box paraphrase-consistency baseline measures the drop in accuracy when each question is reworded by an external model. No reported result was found on Lacuna for paraphrase consistency on post-RL reasoning models.

Three ablations isolate the key ingredient. The first replaces neuron-normalized noise with isotropic noise, testing whether scale invariance matters as Tsuzuku et al. argue. The second replaces weight noise with input-embedding noise at matched output-change magnitude, testing whether the input–parameter equivalence makes weight access unnecessary. The third removes the difficulty residualization, testing whether the raw gap is just a confidence proxy.

The method counts as beating the field if it holds at least 0.70 item-level AUROC after two GRPO epochs on chain-of-thought contamination, where the baselines sit near 0.55. It must also beat paraphrase consistency and embedding noise by a margin outside bootstrap confidence intervals.

Why not just…?

Why not just run EPGS on benchmark items? EPGS is the nearest competitor and is run as a baseline. It has three relevant weaknesses. It relies on named entities, which math traces lack. It needs two backward passes per item. And it was never tested on RL-cloaked models. If adapted EPGS matches weight noise, the conclusion would be that the input–parameter equivalence carries over to contamination. That would still be a useful result, and the embedding-noise ablation is designed to show it.

Why not just paraphrase the questions and see whether accuracy drops? This is the most serious threat, because it is cheap and works on black-box models. Its weakness is that chain-of-thought contamination teaches transferable solution templates, so a reworded question may still be solved; the Fragility paper describes exactly this generalization to similar unseen items. The paraphrase arm and the head-to-head comparison settle the question directly. If paraphrase consistency matches $s(x)$, the honest recommendation is to use paraphrase consistency for black-box models and reserve weight noise for open-weight audits.

Why not just evaluate on fresh competitions like MathArena? Fresh problems are the right long-term answer, but MathArena gathered only 162 problems in 2025. Fresh evaluation also cannot retroactively audit scores already claimed on GPQA or OlympiadBench. The proposal uses AIME 2025 as a clean anchor rather than a competitor.

Why not just use a reference model with LiRA? LiRA is included. It is already documented as degrading under GRPO, and it requires shadow models whose training data is known, which is unrealistic for a third-party auditor.

Decisive Checks

Three outcomes would decide the question. The first is whether $s(x)$ AUROC stays roughly flat across RL epochs while loss and Min-K% decline. A matching decline would falsify the claim that clipping leaves the neighborhood untouched. The second is whether the advantage survives difficulty matching: on clean and contaminated items with equal unperturbed accuracy, the gap must still separate them. Otherwise the method measures hardness, not leakage. The third is whether the estimated separability $\rho$ is large enough that the $\alpha\rho\sqrt{m}$ calculation makes a 20% contamination fraction detectable at 80% power on GPQA Diamond. If it is, item-level gains become audits that work on frontier-scale benchmarks. If $s(x)$ also flags many unseen items that are similar to the injected ones in the chain-of-thought arm, the result would support the Fragility paper's claim that the model genuinely generalized. That is a substantive finding about what contamination does, not just a failure of the detector.