Benchmark Contamination: A Taxonomy Organized by Defeated Mitigation

A benchmark score is a joint property of the model, the evaluation harness, the elicitation budget, the sampled population, and contamination status. Leaderboards publish the model and the score, so capability and leakage stay observationally equivalent. Existing taxonomies classify contamination for automated detection, not the question a reporter faces at publication: given the mitigations already applied, which validity threats remain open? We introduce a taxonomy organized by the mitigation each type defeats -- direct, derivative, temporal, distributional, and acquired -- spanning training-time and evaluation-time leakage. Holding out a private test set closes the first alone. The fifth is acquired during the evaluation itself; because it is a property of one run, it must be recorded with the reported score rather than with the benchmark release. We operationalize it as a four-field disclosure protocol in which "unknown" is a valid entry, released under CC BY 4.0 with a JSON Schema, a validator, and worked examples. Two coders external to the design team applied a pre-registered instrument to 41 documents. Per-variable linear-weighted $κ$ runs from 0.00 to 0.35 (median 0.21) over 29 main-pass documents against a single-coder test-retest ceiling of 0.84, collapsing under the class skew the registration anticipated; pooling raises it to 0.46 through chance correction rather than better agreement. Two variables fall below the prevalence-robust threshold registered in advance: strata reporting and the acquired type introduced here. Disagreement concentrates on when a variable applies rather than on what a document states. Elicitation budgets are reported in 13% of documents, and no document addresses all five types. The contribution is the taxonomy, the score-side artifact that follows from it, and a pre-registered measurement of instrument reliability and current disclosure.

Structure, Association, and Decision Value: Representation-Based Difficulty Estimation for Adaptive Inference in African-Language NLI

We ask whether internal representation statistics can provide useful example-level difficulty signals for adaptive inference in multilingual African NLP, and find that they cannot in this setting. Studying natural language inference across 15 African languages with frozen off-the-shelf checkpoints, we report four results. First, AfriXNLI's English configuration shares 1,047 of its 1,050 examples verbatim with XNLI evaluation data, and one widely used NLI checkpoint scores 1.000 on that test split, consistent with XNLI test exposure. Because AfriXNLI is derived from XNLI, its English, French and Swahili configurations cannot serve as clean evaluations for XNLI-trained models. Second, parameter count does not reliably order capability across African languages: our larger checkpoint is better in seven languages and worse in eight, with no significant aggregate difference. Third, across three multilingual representation spaces, angular dispersion is consistently more language-determined than effective rank, so pooled correlations can inflate one and mask the other. Fourth, the association that survives language control depends on the target: effective rank predicts probability gain from escalation but not whether escalation changes the prediction, while cheap-model confidence shows the opposite pattern; the two targets correlate at only 0.655. Under the tested models, signals, and compute budgets, no evaluated signal makes adaptive routing preferable to always-expensive inference, although an oracle exceeds it by 11 accuracy points at 60% of the compute. Our central methodological finding is that a representation statistic can be statistically significant for one notion of computational benefit while being irrelevant to another, and therefore be a poor decision variable.

Excess Separability: Nuisance-Controlled Residual-Stream Probing for Benchmark Contamination Detection

Benchmark contamination is diagnosed with n-gram overlap, likelihood-based membership inference, or canary strings, and each needs something usually unavailable: the training corpus, a well-chosen test statistic, or foresight at release. A recent alternative reads it off a linear probe on internal activations. We show the natural way to do this does not work, specify one that survives measurement, then find that the correction making it work carries more variance than the null it is tested against. The protocol reports a zero-sum contrast on the depth profile of probe accuracy, recentred on a level-matched placebo baseline, tested against a label-permutation null, with the reference set twice the size of the suspect set. Each choice replaces a simpler alternative we rejected on measurement. Reporting the level of excess separability rather than its shape makes the false positive rate track the size of the analyst's own control set, 0.03 to 0.99 under a true null. Contrasting against a flat depth profile rejects a true null 0.72 of the time when surface decodability rises with depth, and loses all power when it falls. On real transformers the protocol fails a test the simulations did not pose. The recentring subtracts an estimate, and the permutation null holds it fixed. Re-estimated across split seeds on four audits of contaminated checkpoints, its standard deviation is 1.30 to 1.56 times the null's own in every arm: what is subtracted to remove a bias is more variable than what it corrects. The one nominally significant result, p = 0.0075, becomes 0.0745 once that variance is propagated, and no verdict is issued. The simulations missed this because their surface key is the covariate driving item variation; on real text it is a proxy, and degrading key quality in simulation reproduces it. We add a companion measurement and a widened null. No arm shows contamination.

When Is Benchmark Contamination Detectable? Information Limits and Power-Calibrated Audits

Behavioral contamination detectors can return "no evidence" either because a benchmark is clean or because the audit has little power. We formalize this distinction for a benchmark in which an unknown fraction alpha of items was seen during training. With matched clean and seen controls, the behavioral channel is the sparse mixture Q_alpha = (1 - alpha) P_0 + alpha P_1, and an exact second-moment argument shows that detectability is governed by alpha * rho * sqrt(m), where rho^2 = chi^2(P_1 || P_0) measures behavioral separability. Any scalar detector reduces to its efficacy, ef = |E_1 f - E_0 f| / sqrt(Var_0(f)) <= rho, which can be estimated from controls before the audit is run. A separate sample-split certificate lower-bounds alpha distribution-free, without requiring an orientation assumption. Our empirical finding is two-sided. Frozen calibration efficacy predicts held-out power curves, with R^2 = 0.83-0.98 across six exact-permutation channels, but the efficacy-only Gaussian budget is miscalibrated at the small sample sizes it prescribes, failing in 9/9 gate-passing channels even though efficacy itself transports. The failure is in the inversion, not the calibration. A predeclared two-stage planner that simulates the deployed test repairs the budgets, is uniformly conservative, and abstains when its probe does not transport. The certificate is valid but vacuous at audit scale, and a five-seed paired injection study recovers the mechanism ordering verbatim > paraphrase > surface, in which the apparent answer-only signal is explained by baseline drift. We report the audit contract and its failures together: a non-rejection is interpretable only alongside the efficacy, budget, and validity gates that produced it.

Zero Gap Is Not Restoration: Stratified Per-Question Probability Evaluation and Step-wise Mitigation of Benchmark Contamination

Test data from public benchmarks inevitably leaks into pretraining corpora, inflating evaluation scores once memorized. \textbf{Contamination mitigation evaluation} intervenes in the decoding process to suppress memorization and restore a contaminated model's genuine capability, but its prevailing metric, the \textbf{G-AP} (\textbf{G}ap of \textbf{A}ggregate \textbf{P}erformance), is flawed. Discrete correct/incorrect readouts cannot characterize per-question performance, averaging before differencing lets over- and under-suppression cancel out, and uniform per-question weighting invites strategies to push solve probabilities onto the clean model's high-frequency values. We propose \textbf{SA-PPG} (\textbf{S}tratified \textbf{A}ggregate of \textbf{P}er-question \textbf{P}robability \textbf{G}aps): estimate each question's solve probability by sampling, difference it against the clean model per question, and aggregate within groups defined by the clean model's solve probability. Existing mitigation strategies first estimate where contamination lies and then operate on the estimate, so they are only as correct as the estimate. \textbf{RailCap} instead judges contamination during generation: whenever a sample falls back onto the greedy trajectory, the next trajectory token is capped to the runner-up, accumulating suppression until the response distribution becomes sufficiently dispersed. Across multiple contaminated models and benchmarks, SA-PPG reveals that prior strategies' restoration is substantially overestimated, while RailCap attains the lowest SA-PPG.

Contamination Inflates Scores but Rarely Reorders Large Language Model Leaderboards

Benchmark contamination, the leakage of test items into training data, is widely described as a threat to the reliability of large language model (LLM) leaderboards. We argue that this concern conflates two distinct questions: whether contamination inflates absolute scores, and whether it reorders the ranking of models. We recast contamination as a violation of anchor-item invariance and measure it through the differential functioning of original versus semantically equivalent paraphrased items, a within-item contrast that holds the measured skill fixed and isolates memorization from capability. Using per-instance responses from 47 publicly released models and 74 models finetuned with a known dose of contamination, across four benchmarks (ARC, GSM8K, HellaSwag, MMLU), we first calibrate the measure against ground truth: it recovers injected contamination dose-responsively (a corrected effect of +0.187 accuracy points for test-set leakage) and never flags a negative-control model trained only on the legitimate training split (-0.012). We then quantify leaderboard impact: the rank correlation between a standard leaderboard and a paraphrase-controlled leaderboard is 0.997, and a sensitivity analysis shows that the observed differential contamination is far below the level needed to move rankings, with only 3 of 188 model-by-benchmark cases showing differential contamination corroborated across two references. Contamination among these public models is therefore largely uniform: it inflates absolute scores without reordering the leaderboard, and ranking distortion requires the rare case of differential contamination. We provide a calibrated invariance audit, released as a reference implementation, and recommend that leaderboards report paraphrase-controlled rankings alongside confidence intervals.

Search-Time Contamination in Deep Research Agents: Measuring Performance Inflation in Public Benchmark Evaluation

Public benchmarks enable fair and reproducible evaluation of LLM reasoning, but they become fragile for deep research agents that actively search the web during inference. Such agents may retrieve public benchmark metadata, question context, or even ground-truth answers via web search. This gives rise to Search-Time Contamination (STC), where external retrieval bypasses intended reasoning and inflates measured performance. We systematically study STC in deep research agent evaluation. We define three contamination types with increasing severity, namely Benchmark Metadata Leakage, Question-Context Leakage, and Explicit Answer Leakage, and develop detection algorithms to identify them and quantify their impact on agent performance. Evaluating modern deep research agents on six public benchmarks, we find that STC is widespread and can inflate performance by up to 4%. Our findings show that existing evaluations may overestimate true reasoning ability. We therefore advocate contamination-aware practices, including isolated sandboxes, transparent search trajectories, and controlled benchmark access.

The Reliability Gap in Benchmark Auditing: Distribution Shift and Scale as Failure Modes of Contamination Detection

Benchmark contamination, where evaluation examples appear in a model's training data, threatens the validity of LLM assessment. Statistical tools for detecting training-data membership exist, but have been validated almost exclusively in controlled academic regimes: large, homogeneous pre-training corpora and transparent, single-stage training pipelines. Whether these methods remain reliable in realistic auditing scenarios remains unclear. We identify two under-studied failure modes: distribution shift, which arises when suspect and validation sets violate the IID assumption, and scale constraints, which arise because benchmarks are orders of magnitude smaller than pre-training corpora. We systematically evaluate three leading paradigms, LLM Dataset Inference, Post-Hoc Dataset Inference, and CoDeC, across 25 models from multiple families (including Pythia, OLMo 2, and specialised cultural and medical LLMs) and scales (up to 27B). We then further extend our analysis to frontier industry models. Across 335 evaluations, only 201 yield correct outcomes. LLM Dataset Inference results in false positives under distribution shift, Post-Hoc Dataset Inference is underpowered at benchmark scale, and CoDeC provides only coarse provenance signals that are insufficient to verify individual benchmark splits. Our results reveal a systematic reliability gap between controlled validation and practical benchmark auditing, and show that statistical detection cannot yet replace transparent data provenance. We open-source our benchmark for further research.