When a language model has seen benchmark questions during training, its scores overstate what it can do. The standard way to catch this is to check whether the model is unusually confident on the test items, for example by measuring low loss on exact strings. Recent work shows that this check stops working once the model goes through ordinary reinforcement-learning fine-tuning: the training squeezes the confidence gap between seen and unseen questions until detectors are close to coin-flips. We argue that this failure resembles a known failure in adversarial disentanglement, the line of work that trains representations to hide an attribute from a classifier. In that literature, hiding an attribute from one fixed classifier usually leaves it recoverable by a freshly trained one. Our hypothesis is that reinforcement learning hides contamination only from output-level detectors, and that the evidence survives in the model's internal activations, where a fresh probe can find it.
The proposal has two parts. The first is a detector that compares the internal geometry of a benchmark question with that of meaning-preserving "twins" of the same question, and that also uses probes trained on activations. The second, and the main methodological contribution, is a hider–finder stress test. A motivated adversary fine-tunes the contaminated model with gradient reversal against our detector. We then measure whether newly trained probes recover the signal, and how much of the inflated score the hider has to give up to defeat them. The result tells auditors either that internal-geometry auditing is robust or, just as usefully, that contamination can be concealed even from the inside.
Core Claim
On The Fragility of Benchmark Contamination Detection in Reasoning Models motivates the work. It contaminates Qwen2.5-7B-Instruct through supervised fine-tuning on olympiad math and GPQA items, which inflates pass@1 by 8.8% on average. It then shows that one epoch of clean GRPO (group relative policy optimization, a PPO-style reinforcement-learning method with clipped updates) degrades LiRA from 89.1% AUROC by about 14%. When contamination enters through chain-of-thought training of DeepSeek-R1-Distill-Llama-8B, all ten detectors average 50–58% AUROC. That paper attributes the collapse to clipping shrinking the loss variance of unseen items, and calls for new detection paradigms.
A Narrowing Geometry in Contaminated Reasoning supplies our candidate signal. On GSM8K and MSCalc with Qwen2.5 and Llama3.1, leaked problems drive activations into a low-dimensional subspace, measured by an Eigenspectrum Concentration Index (ECI), and decouple activations from loss gradients. Projecting out the top subspace (Top-Subspace Suppression, TSS) restores reasoning on about half of leaked samples. That work uses the geometry to repair models. It never tests whether the geometry survives reinforcement learning, and it never uses the geometry as a detector against an adaptive opponent.
Esser et al. supply the conceptual warning. Encoders can fool an adversarial discriminator by hiding attributes in low-amplitude patterns that a fresh nonlinear probe still reads. DECO-Bench supplies the audit methodology. It separates privacy measured by the training-time adversary from privacy measured by a dedicated classifier trained afterward, and finds that the second consistently reveals leakage the first missed.
Our claim is twofold. First, reinforcement-learning cloaking is accidental invariance at the output level, and contamination remains decodable from intermediate activations of RL-tuned reasoning models well above the 50–58% AUROC of output detectors. Second, the honest test of any internal detector is the hider–finder protocol: train a concealer against it, then train fresh finders.
Research Program
The detector has an unsupervised score and a supervised probe. For each benchmark item $q$, we generate $K$ twins $\tilde q_1,\dots,\tilde q_K$. These are paraphrases with the same answer, and for multiple choice they also have shuffled options. We define the twin-contrast score at layer $\ell$ as
$$s_\ell(q) = \frac{1}{K}\sum_{k=1}^{K}\mathrm{ECI}_\ell(\tilde q_k) - \mathrm{ECI}_\ell(q),$$
where $\mathrm{ECI}_\ell$ is the effective-rank-based concentration index of the activations over the model's own reasoning trace. We define it so that larger values mean a broader, less concentrated spectrum. A positive $s_\ell$ therefore means the original question is processed more narrowly than its unseen twins.
The twins serve as a per-item reference that the model cannot have memorized. This matters because a clipped reinforcement-learning update shrinks loss variance, which is a statistic over outputs. There is no obvious reason it should equalize the activation spectrum of an exact memorized string relative to its paraphrase.
The supervised finder is a linear probe, plus a small multilayer perceptron, trained on activations from twin–original pairs. Its labels come from a controlled contamination split, and it is evaluated on held-out items from a different benchmark. This rules out the probe simply learning topic.
The hider starts from the contaminated, GRPO-tuned model. It fine-tunes the model with a gradient-reversal layer against the frozen finder and a differentiable surrogate of $s_\ell$, following the encoder–adversary game of Fader Networks. It is constrained to keep its inflated pass@1 on the leaked items. We then train fresh finders on the hidden model at all layers. We report three quantities: the fixed-finder AUROC, the fresh-finder AUROC, and the fraction of the contamination boost the hider retains. A concealment is only meaningful if the boost survives, because a hider that destroys the memorized answers has effectively unlearned them.
Baselines
The naive reference is chance, 50% AUROC, together with the plain loss-threshold detector. The fragility paper reports that loss thresholding drops by about 16% after one GRPO epoch, so this shows where an untuned auditor stands.
The strongest existing method is LiRA, a membership attack calibrated with reference models. On The Fragility of Benchmark Contamination Detection in Reasoning Models reports 89.1% AUROC for LiRA before reinforcement learning, a drop of about 14% after one GRPO epoch, and 50–58% averaged across detectors under chain-of-thought contamination. We also run Min-K% and verbatim matching. We reproduce the paper's contamination pipeline exactly: the same models, the same benchmarks, a random half of items leaked, and one clean GRPO epoch. All detectors are scored on identical items, and reference-model budgets are matched in GPU-hours to our probe training. Beating LiRA means at least 70% AUROC in the chain-of-thought contamination setting, with non-overlapping bootstrap intervals, where output detectors sit near 50–58%.
The second strong competitor is the raw ECI of A Narrowing Geometry in Contaminated Reasoning, used directly as a detection score without twins or probes. We found no reported detection result on Lacuna for ECI used as a contamination detector, and none after reinforcement learning.
We run three ablations, each isolating one ingredient. The first removes the twins. The second replaces the learned probe with the fixed score. The third evaluates each detector only against the hider it was trained on, without fresh finders. This last ablation shows how much a single-adversary evaluation overstates robustness.
Why not just…?
Why not just use raw ECI as the detector? This is the real threat. If narrowing geometry alone separates leaked from clean items after GRPO, the twin contrast adds little to detection. That would still be a new and useful finding, since it is untested. The proposal's distinct value would then rest on the hider–finder test. A single unsupervised statistic is exactly the kind of fixed adversary that Esser et al. show can be gamed. The raw-ECI baseline and the hider experiment together settle this: we expect raw ECI to fall under concealment faster than fresh probes do.
Why not just improve output-level calibration with more reference models? The fragility paper shows the collapse is structural. Clipping contracts the variance of unseen items, and chain-of-thought training makes the model confident on unseen neighbors, so more reference models calibrate a gap that no longer exists. We test this directly by tripling LiRA's reference budget in the chain-of-thought contamination setting.
Why not just compare accuracy on originals versus paraphrases, as EvoEval does? The survey Benchmark Data Contamination of Large Language Models: A Survey reports a 39.4% average drop from HumanEval to EvoEval. However, an accuracy gap is aggregate rather than per-item. It is also confounded by changes in difficulty: Disentangling Text and Math in Word Problems shows that rephrasing shifts accuracy by factors of 2 to 7 on ENEM. In addition, reasoning models that have internalized templates may solve twins equally well. We include per-item correctness gap as a baseline using the same twins, so the comparison isolates internal geometry from behavior.
Why not just move to live benchmarks such as LiveBench? Live benchmarks prevent future leaks. They cannot audit whether an already released model inflated its scores on the static benchmarks that the field keeps reporting, and that audit is the question here.
Proposed Experiments
The main testbed reproduces both contamination stages of the fragility paper on AIME, AMC, OlympiadBench, Minerva Math and GPQA Diamond. Leakage labels are known by construction. A secondary testbed contaminates Qwen2.5-1.5B, Qwen2.5-7B and Llama3.1-8B on GSM8K, following the TSS setup, and then adds GRPO, so we can check whether the narrowing geometry reported there survives reinforcement learning. Probes are trained on GSM8K leaks and tested on olympiad leaks, and the reverse.
There are three decisive checks. The first asks whether twin-contrast or probe AUROC after GRPO exceeds the 50–58% output-detector band. The second asks whether, under the hider, fresh-finder AUROC stays well above fixed-finder AUROC while the retained pass@1 boost stays high. The third asks whether concealment that defeats fresh finders requires giving up most of the 8.8–11.7% boost. If concealment is cheap and complete, we will report internal auditing as broken too. That outcome is a clear negative result that would redirect contamination auditing toward cryptographic or live evaluation.
References
- On The Fragility of Benchmark Contamination Detection in Reasoning Models
- A Narrowing Geometry in Contaminated Reasoning
- Unsupervised Robust Disentangling of Latent Characteristics for Image Synthesis
- DECO-Bench: Unified Benchmark for Decoupled Task-Agnostic Synthetic Data Release
- Fader Networks: Manipulating Images by Sliding Attributes
- Benchmark Data Contamination of Large Language Models: A Survey
- Disentangling Text and Math in Word Problems: Evidence for the Bidimensional Structure of Large Language Models' Reasoning
- LiveBench: A Challenging, Contamination-Limited LLM Benchmark