cs.LGFeb 12, 2026

Soft Contamination Means Benchmarks Test Shallow Generalization

Authors: Ari Spiesberger, Juan J. Vazquez, Nicky Pochinkov, Tomáš Gavenčiak, Peli Grietzer, Gavin Leech, Nandi Schoots

Organizations: Arb Research · Charles University, Prague · University of Cambridge · University of Oxford

Abstract

Training on benchmark test items is a classic way to overestimate generalization to unseen problems. Standard decontamination methods miss semantic duplicates (SemDs): training items equivalent in content to a test item, but distant in string space. We call training on such items soft contamination; when the resulting gains extend to held-out items of the same benchmark (but not to related benchmarks), we call the result shallow generalization. We first embed Olmo 3's entire post-training dataset and search it for SemDs of benchmark items; we then finetune Olmo 3 and Qwen models on natural and synthetic semantic duplicates. We find (1) that soft contamination is widespread: the Olmo corpus contains SemDs of 76% of the CodeForces test set and 100% of MBPP's; (2) that finetuning on SemDs of a benchmark raises accuracy on its held-out items by up to 22pp in controlled experiments and 10-11pp at realistic contamination rates, though the effect varies across benchmarks; (3) that the gains from training on SemDs mostly do not transfer to conceptually similar benchmarks, i.e. we find shallow generalization; (4) that finetuning on items with similar embeddings that are not SemDs produces no such gain, indicating that the mechanism is not mere topical similarity; and (5) that at realistic contamination rates, soft contamination can raise held-out performance about as much as exact duplicates (or more). Benchmark gains here are therefore confounded - in a way that standard decontamination does not address.

Figures & tables

Appendix figures & tables18 assets

Supplementary material from the paper’s appendix.

Appendix

Explore similar work

Nov 16, 2026

Benchleaking Is Accelerating: Benchmark contamination has worsened sharply in the past month -- from post-training down to pre-training base models

Benchmark contamination is not new, but recent model releases show it in a sharper form. We use the term benchleaking for the case in which a benchmark's own test examples enter training data, so that a score reflects recall rather than capability. This differs from benchmaxxing, which optimizes toward a benchmark without seeing its test set. We document two escalations over the past month. First, in post-training, the traces have become explicit: recent audits show an agent stating hidden test requirements it was never given in 23% of its attempts, where its predecessor did so in none, and continuing to follow them after the task is changed to contradict them (Schaeffer and Räuker, 2026; Lo and Lo, 2026). Second, the problem has reached pre-training. Using few-shot continuation on 20 benchmarks across four public base models, we find that some base models write out other test examples---complete questions, options, and answers---that appear nowhere in their input. For the recently released ERNIE-5.5-Flash-Base, 96% of HumanEval problems leak with their reference implementations and 85% of OpenBookQA questions leak with their answers, against 9% and none for DeepSeek-V4-Flash-Base. To quantify what a reported score owes to such examples, we define an Anti-Hack score: the benchmark score recomputed with every leaked example marked wrong and the number of examples unchanged. Averaged over the suite, it costs DeepSeek-V4-Flash-Base 1.5 points and Hy3-preview-Base 1.3, against 5.0 for MiMo-V2.5-Base and 16.6 for ERNIE-5.5-Flash-Base. Open base models are the foundation of much downstream research, so benchleaking now affects more than leaderboard rankings. We call for decontamination reports and contamination-adjusted scores to become standard practice.
May 19, 2026

LLM Benchmark Datasets Should Be Contamination-Resistant

Benchmark datasets are critical for reproducible, reliable, and discriminative evaluation of LLMs. However, recent studies reveal that many benchmark datasets are included in pretraining corpora, i.e., contaminated\textit{contaminated}, which diminishes their value as reliable measures of model generalization. In this paper, we argue that benchmark datasets should be contamination-resistant\textit{contamination-resistant}, i.e., unlearnable\textit{unlearnable}, but support inference\textit{inference}. To accomplish this, we first highlight the wide prevalence of benchmark dataset contamination and outline the properties of contamination-resistant datasets. Second, we highlight how the asymmetry between the inference and training pipelines in the Transformer architecture can be leveraged to support contamination-resistance. Third, we outline mathematical advancements to make these datasets interoperable across various LLM architectures. Based on the above, we call on the community to ensure the reliability of LLM benchmarking by: (i) advancing novel contamination-resistant methodologies, (ii) developing supporting methods and platforms, and (iii) adopting contamination-resistant benchmarks into existing evaluation pipelines.
Jan 7, 2026

Quantifying the Effect of Test Set Contamination on Generative Evaluations

As frontier AI systems are pretrained on web-scale data, test set contamination has become a critical concern for accurately assessing their capabilities. While research has thoroughly investigated the impact of test set contamination on discriminative evaluations like multiple-choice question-answering, comparatively little research has studied the impact of test set contamination on generative evaluations. In this work, we quantitatively assess the effect of test set contamination on generative evaluations through the language model lifecycle. We pretrain language models on mixtures of web data and the MATH benchmark, sweeping model sizes and number of test set replicas contaminating the pretraining corpus; performance improves with contamination and model size. Using scaling laws, we make a surprising discovery: including even a single test set replica enables models to achieve lower loss than the irreducible error of training on the uncontaminated corpus. We then study further training: overtraining with fresh data reduces the effects of contamination, whereas supervised finetuning on the training set can either increase or decrease performance on test data, depending on the amount of pretraining contamination. Finally, at inference, we identify factors that modulate memorization: high sampling temperatures mitigate contamination effects, and longer solutions are exponentially more difficult to memorize than shorter ones, presenting a contrast with discriminative evaluations, where solutions are only a few tokens in length. By characterizing how generation and memorization interact, we highlight a new layer of complexity for trustworthy evaluation of AI systems.