Hampering the interpretation of benchmark scores, evaluation data contamination has become a growing concern in the evaluation of LLMs, and an active area of research studies its effects. While evaluation data contamination is easily understood intuitively, it is surprisingly difficult to define precisely which samples should be considered contaminated and, consequently, how it impacts benchmark scores. We propose that these questions should be addressed together and that contamination metrics can be assessed based on whether models benefit from the examples they mark contaminated. We propose a novel analysis method called ConTAM, and show with a large scale survey of existing and novel n-gram based contamination metrics across 13 benchmarks and 7 models from 2 different families that ConTAM can be used to better understand evaluation data contamination and its effects. We find that contamination may have a much larger effect than reported in recent LLM releases and benefits models differently at different scales. We also find that considering only the longest contaminated substring provides a better signal than considering a union of all contaminated substrings, and that doing model and benchmark specific threshold analysis greatly increases the specificity of the results. Lastly, we investigate the impact of hyperparameter choices, finding that, among other things, both using larger values of n and disregarding matches that are infrequent in the pre-training data lead to many false negatives. With ConTAM, we provide a method to empirically ground evaluation data contamination metrics in downstream effects. With our exploration, we shed light on how evaluation data contamination can impact LLMs and provide insight into the considerations important when doing contamination analysis. We end our paper by discussing these in more detail and providing concrete suggestions for future work.

Smart metering networks are increasingly susceptible to cyber threats, where false data injection (FDI) appears as a critical attack. Data-driven-based machine learning (ML) methods have shown immense benefits in detecting FDI attacks via data learning and prediction abilities. Literature works have mostly focused on centralized learning and deploying FDI attack detection models at the control center, which requires data collection from local utilities like meters and transformers. However, this data sharing may raise privacy concerns due to the potential disclosure of household information like energy usage patterns. This paper proposes a new privacy-preserved FDI attack detection by developing an efficient federated learning (FL) framework in the smart meter network with edge computing. Distributed edge servers located at the network edge run an ML-based FDI attack detection model and share the trained model with the grid operator, aiming to build a strong FDI attack detection model without data sharing. Simulation results demonstrate the efficiency of our proposed FL method over the conventional method without collaboration.

Benchmarking can include comparing an ICU's actual performance with predicted performance. The increased interoperability of medical devices, electronic health records (EHRs) and information systems has improved the acquisition and presentation of data to healthcare professionals. This data has enabled the training of predictive models. However, thi s plethora of data sources has also introduced new risks that societal bias will lead to machine learning systems with fairness issues for patient groups. In addition, when variations in data documentation are non-random, significant bias can be introduced, improving, or worsening measured performance for an institution relative to peers. This work focuses on ICU mortality benchmarking. In particular, we analyze the fairness of a model based on Generalised Additiv e Models (GAM) [ 3 ] that predicts mortality in the ICU. This model is used to compare actual versus predicted outcom es to assess ICU performance.

We examine diverging preferences in human-labeled preference datasets. We develop a taxonomy of disagreement sources spanning 10 categories across four high-level classes--task underspecification, response style, refusals, and annotation errors. We find that the majority of disagreements are in opposition with standard reward modeling approaches, which are designed with the assumption that annotator disagreement is noise. We then explore how these findings impact two areas of LLM development: reward modeling and evaluation. In our experiments, we demonstrate how standard reward modeling methods, like the Bradley-Terry model, fail to differentiate whether a given preference judgment is the result of unanimous agreement among annotators or the majority opinion among diverging user preferences. We also find that these tendencies are also echoed by popular LLM-as-Judge evaluation methods, which consistently identify a winning response in cases of diverging preferences. These findings highlight remaining challenges in LLM evaluations, which are greatly influenced by divisive features like response style, and in developing pluralistically aligned LLMs. To address these issues, we develop methods for identifying diverging preferences to mitigate their influence on evaluation and training. As large language models (LLMs) continue to rise in prominence and to serve millions of people on a daily basis, there is an increasing need to ensure that systems are pluralistically aligned (Sorensen et al., 2024). Learning from human preferences has emerged as the standard method for adapting LLMs to facilitate user-assistant interactions with much success. Despite these advances, however, the field continues to struggle with the challenge of handing diverging preferences, where users disagree on the ideal response to a prompt. Prior works on developing pluralistically aligned LLMs have focused on the development of synthetic preference datasets, where disagreements are simulated based on author-defined features and frequencies (Poddar et al., 2024; Chen et al., 2024). In this work, we take a step back to ask the foundational question when and why do human annotators disagree in their preferences? To make this research possible, we introduce MultiPref-Disagreements and HelpSteer2-Disagreements.

The goal of this paper is to develop a multilingual classifier and conditional probability estimator of occupation codes for online job advertisements according in accordance with the International Standard Classification of Occupations (ISCO) extended with the Polish Classification of Occupations and Specializations (KZiS), which is analogous to the European Classification of Occupations. In this paper, we utilise a range of data sources, including a novel one, namely the Central Job Offers Database, which is a register of all vacancies submitted to Public Employment Offices. Their staff members code the vacancies according to the ISCO and KZiS. A hierarchical multi-class classifier has been developed based on the transformer architecture. The classifier begins by encoding the jobs found in advertisements to the widest 1-digit occupational group, and then narrows the assignment to a 6-digit occupation code. We show that incorporation of the hierarchical structure of occupations improves prediction accuracy by 1-2 percentage points, particularly for the hand-coded online job advertisements. Finally, a bilingual (Polish and English) and multilingual (24 languages) model is developed based on data translated using closed and open-source software. The open-source software is provided for the benefit of the official statistics community, with a particular focus on international comparability.

The propensity of Large Language Models (LLMs) to generate hallucinations and non-factual content undermines their reliability in high-stakes domains, where rigorous control over Type I errors (the conditional probability of incorrectly classifying hallucinations as truthful content) is essential. Despite its importance, formal verification of LLM factuality with such guarantees remains largely unexplored. In this paper, we introduce FactTest, a novel framework that statistically assesses whether a LLM can confidently provide correct answers to given questions with high-probability correctness guarantees. We formulate factuality testing as hypothesis testing problem to enforce an upper bound of Type I errors at user-specified significance levels. Notably, we prove that our framework also ensures strong Type II error control under mild conditions and can be extended to maintain its effectiveness when covariate shifts exist. Our approach is distribution-free and works for any number of human-annotated samples. It is model-agnostic and applies to any black-box or white-box LM. Extensive experiments on question-answering (QA) and multiple-choice benchmarks demonstrate that FactTest effectively detects hallucinations and improves the model's ability to abstain from answering unknown questions, leading to an over 40% accuracy improvement.

Despite recent advances in understanding visual and auditory perception, olfactory perception remains an under-explored topic in the machine learning community due to the lack of large-scale datasets annotated with labels of human olfactory perception. In this work, we ask the question of whether pre-trained transformer models of chemical structures encode representations that are aligned with human olfactory perception, i.e., can transformers smell like humans? We demonstrate that representations encoded from transformers pre-trained on general chemical structures are highly aligned with human olfactory perception. We use multiple datasets and different types of perceptual representations to show that the representations encoded by transformer models are able to predict: (i) labels associated with odorants provided by experts; (ii) continuous ratings provided by human participants with respect to pre-defined descriptors; and (iii) similarity ratings between odorants provided by human participants. Finally, we evaluate the extent to which this alignment is associated with physicochemical features of odorants known to be relevant for olfactory decoding.

Metric-based methods rely on the analysis of certain statistical properties of a target model's output, such as confidence scores, prediction probabilities, or loss values, to distinguish between training data and non-training data. Specifically, Metric-loss (Yeom et al., 2018) is the first metricbased detection method, predicting that data points with a loss below a certain threshold are part of the training data for the target model. Similarly, other works have proposed using the maximum confidence of the target model output (denoted as Metric-conf (Song et al., 2019)), the correctness of the target model output (denoted as Metric-corr (Leino & Fredrikson, 2020)), the entropy of prediction probability distributions (denoted as Metric-ent (Shokri et al., 2017; Song & Mittal, 2021)), and modified entropy of the prediction (denoted as Metric-ment (Song & Mittal, 2021)). Learning-based methods involve training an auxiliary classifier (meta-classifier) to distinguish between training data and non-training data. In the literature, neural networks (NNs) are often employed as the auxiliary classifier. The primary differences between learning-based TDD methods lie in the choice of input features for the auxiliary classifier. Earlier work (Shokri et al., 2017) has proposed using the original prediction vector of the target model (denoted as Learn-original). Other works have suggested using the top-3 prediction confidences (denoted as Learn-top3 (Salem et al., 2019)), the sorted prediction vector (denoted as Learn-sorted (Salem et al., 2019)), the true label of the example combined with the prediction vector (denoted as Learn-label

We perform a detailed evaluation of two important techniques: In-Context Learning (ICL) and Supervised Fine-Tuning (SFT), showing how they can improve detection. SFT increases performance by using domain-specific data, whereas ICL helps the detection model to quickly adapt to new threats without requiring much retraining. We use Meta's Llama3 8B model, on a custom dataset with 68 malware families and normal domains, covering several hard-to-detect schemes, including recent word-based DGAs. Results proved that LLM-based methods can achieve competitive results in DGA detection. In particular, the SFT-based LLM DGA detector outperforms state-of-the-art models using attention layers, achieving 94% accuracy with a 4% false positive rate (FPR) and excelling at detecting word-based DGA domains.

Diabetic Retinopathy (DR) is a primary cause of blindness, necessitating early detection and diagnosis. This paper focuses on referable DR classification to enhance the applicability of the proposed method in clinical practice. We develop an advanced cross-learning DR classification method leveraging transfer learning and cross-attention mechanisms. The proposed method employs the Swin U-Net architecture to segment lesion maps from DR fundus images. The Swin U-Net segmentation model, enriched with DR lesion insights, is transferred to generate a lesion map. Both the fundus image and its segmented lesion map are used as complementary inputs for the classification model. A cross-attention mechanism is deployed to improve the model's ability to capture fine-grained details from the input pairs. Our experiments, utilizing two public datasets, FGADR and EyePACS, demonstrate a superior accuracy of 94.6%, surpassing current state-of-the-art methods by 4.4%. To this end, we aim for the proposed method to be seamlessly integrated into clinical workflows, enhancing accuracy and efficiency in identifying referable DR.