arXiv:2605. 24818v2 Announce Type: replace-cross Abstract: The literature on test set contamination largely focuses on detection, but the correction of contaminated test scores is underexplored.
The paper investigates how benchmark contamination—leakage of test items into training data—affects large language model (LLM) leaderboards. By comparing original test items with semantically equivalent paraphrases, the authors measure contamination as a violation of anchor-item invariance and find that it inflates absolute scores but rarely changes model rankings. Across 47 public models and 74 finetuned models on four benchmarks, the rank correlation between standard and paraphrase-controlled leaderboards is 0.997, with only a handful of cases showing differential contamination that could alter rankings.
arXiv:2606. 31208v1 Announce Type: new Abstract: Large tabular models (LTMs), i.
The paper addresses the problem of evaluating deployed machine learning models when their input data distribution shifts, which can arise from changes in the underlying population or from data integrity issues in the feature pipeline. It proposes an online model‑testing framework that reproduces the detected shift on a controlled baseline with known labels, distinguishing between the two mechanisms of shift to estimate current performance and potential gains from repairing data integrity problems. Experiments on real‑world datasets with simulated faults demonstrate that separating the two sources of shift generally improves performance estimates, and a churn‑prediction case study shows how these tests can guide monitoring and repair prioritization.
arXiv:2609.37687v1 Announce Type: new Abstract: Active test-time adaptation (ATTA) improves robustness under distribution shift by updating a deployed model during inference while selectively queryin...
arXiv:2610.16152v1 Announce Type: cross Abstract: Although post-training significantly enhances model performance, the use of massive undisclosed data raises concerns about benchmark contamination an...
arXiv:2510. 20351v3 Announce Type: replace-cross Abstract: Large language models (LLMs) are increasingly exposed to data contamination, i.
arXiv:2608. 02985v1 Announce Type: new Abstract: The standard check for contamination in LLM backtests is simple: compare scores before and after the training cutoff.
arXiv:2605.01699v4 Announce Type: replace Abstract: Recent attacks show that behavioural unlearning of large language models leaves internal traces recoverable by adversarial probes. We characterise...
The paper introduces Counterfactual Test-Time Augmentation (CTTA) to mitigate shortcut exploitation in large language models. CTTA generates counterfactual edits of an input using an auxiliary model, then averages the LLM’s predictions across these edits, with theoretically optimal weights derived from the training distribution. The authors propose Select-CTTA, which selects among CTTA variants based on tests for assumption violations, and demonstrate that it outperforms other test-time methods and rivals models fine‑tuned on shortcut‑balanced data.
The paper uncovers a hidden flaw in evaluating machine unlearning on BatchNorm-based models: a single forward pass over retained data can alter the model’s normalization state without changing weights, misleadingly restoring performance. This effect, formalized as a weight‑preserving fixed‑point operator, shows that apparent forgetting can be entirely due to BatchNorm running statistics, not to the unlearning method itself. Experiments demonstrate that the artifact can inflate forget accuracy by up to 78 percentage points across nine methods, and that switching to GroupNorm eliminates the problem.
arXiv:2609.37076v1 Announce Type: new Abstract: Large language models trained on vast corpora inherently risk memorizing harmful content that may later re-emerge in their outputs. To mitigate this is...