LLM-based chatbots are increasingly augmented with retrieval mechanisms operating over web-scale corpora. Evaluating the effectiveness of these retrieval components is challenging, as explicit relevance judgments are often unavailable. Query performance prediction (QPP) addresses this limitation by providing unsupervised estimates of retrieval effectiveness. However, existing QPP methods assume a static corpus and do not account for the impact of corpus growth on query performance. In this work, we extend the QPP paradigm by studying query performance degradation under corpus inflation in dense retrieval systems. Using tiered corpora with fixed relevance judgments, we analyze how query effectiveness evolves as the corpus (index) size increases and evaluate the ability of established score-based and embedding-based post-retrieval QPP methods to predict such degradation. Our findings show that the reliability of these predictors is dependent on the dataset. We propose simple adaptations to established QPP measures, most notably a top-k vs background Wasserstein distance measure, which yield more consistent associations with degradation and outperform their original counterparts. These findings highlight limitations of several QPP approaches in large-scale, continuously expanding retrieval environments and motivate the development of corpus-growth-aware QPP measures.

Embedding models are central to dense retrieval, semantic search, and recommendation systems, but their size often makes them impractical to deploy in resource-constrained environments such as browsers or edge devices. While smaller embedding models offer practical advantages, they typically underperform compared to their larger counterparts. To bridge this gap, we demonstrate that concatenating the raw embedding vectors of multiple small models can outperform a single larger baseline on standard retrieval benchmarks. To overcome the resulting high dimensionality of naive concatenation, we introduce a lightweight unified decoder trained with a Matryoshka Representation Learning (MRL) loss. This decoder maps the high-dimensional joint representation to a low-dimensional space, preserving most of the original performance without fine-tuning the base models. We also show that while concatenating more base models yields diminishing gains, the robustness of the decoder's representation under compression and quantization improves. Our experiments show that, on a subset of MTEB retrieval tasks, our concat-encode-quantize pipeline recovers 89% of the original performance with a 48× compression factor when the pipeline is applied to a concatenation of four small embedding models.

We present WebFAQ, a large-scale collection of open-domain question answering datasets derived from FAQ-style schema.org annotations. In total, the data collection consists of 96 million natural question-answer (QA) pairs across 75 languages, including 47 million (49%) non-English samples. WebFAQ further serves as the foundation for 49 monolingual retrieval benchmarks with a total size of 11.2 million QA pairs (5.9 million non-English). These datasets are carefully curated through refined filtering and near-duplicate detection, yielding high-quality resources for training and evaluating multilingual dense retrieval models. To empirically confirm WebFAQ's efficacy, we use the collected QAs to fine-tune an in-domain pretrained XLM-RoBERTa model. Through this process of dataset-specific fine-tuning, the model achieves significant retrieval performance gains, which generalize - beyond WebFAQ - to other multilingual retrieval benchmarks evaluated in zero-shot setting. Last but not least, we utilize WebFAQ to construct a set of QA-aligned bilingual corpora spanning over 1000 language pairs using state-of-the-art bitext mining and automated LLM-assessed translation evaluation. Due to our advanced, automated method of bitext dataset generation, the resulting bilingual corpora demonstrate higher translation quality compared to similar datasets. WebFAQ and all associated resources are publicly available on GitHub and HuggingFace.

Fraud detection systems (FDS) mainly perform two tasks: (i) real-time detection while the payment is being processed and (ii) posterior detection to block the card retrospectively and avoid further frauds. Since human verification is often necessary and the payment processing time is limited, the second task manages the largest volume of transactions. In the literature, fraud detection challenges and algorithms performance are widely studied but the very formulation of the problem is never disrupted: it aims at predicting if a transaction is fraudulent based on its characteristics and the past transactions of the cardholder. Yet, in posterior detection, verification often takes days, so new payments on the card become available before a decision is taken. This is our motivation to propose a new paradigm: posterior fraud detection with "future" information. We start by providing evidence of the on-time availability of subsequent transactions, usable as extra context to improve detection. We then design a Bidirectional LSTM to make use of these transactions. On a real-world dataset with over 30 million transactions, it achieves higher performance than a regular LSTM, which is the state-of-the-art classifier for fraud detection that only uses the past context. We also introduce new metrics to show that the proposal catches more frauds, more compromised cards, and based on their earliest frauds. We believe that future works on this new paradigm will have a significant impact on the detection of compromised cards.

Abstract The state-of-the-art feature-engineering method for fraud classification of electronic payments uses manually engineered feature aggregates, i.e., descriptive statistics of the transaction history. However, this approach has limitations, primarily that of being dependent on expensive human expert knowledge. There have been attempts to replace manual aggregation through automatic feature extraction approaches. They, however, do not consider the specific structure of the manual aggregates. In this paper, we define the novel Neural Aggregate Generator (NAG), a neural network-based feature extraction module that learns feature aggregates end-to-end on the fraud classification task. In contrast to other automatic feature extraction approaches, the network architecture of the NAG closely mimics the structure of feature aggregates. Furthermore, the NAG extends learnable aggregates over traditional ones through soft feature value matching and relative weighting of the importance of different feature constraints. We provide a proof to show the modeling capabilities of the NAG. We compare the performance of the NAG to the state-of-the-art approaches on a real-world dataset with millions of transactions. More precisely, we show that features generated with the NAG lead to improved results over manual aggregates for fraud classification, thus demonstrating its viability to replace them. Moreover, we compare the NAG to other end-to-end approaches such as the LSTM or a generic CNN. Here we also observe improved results. We perform a robust evaluation of the NAG through a parameter budget study, an analysis of the impact of different sequence lengths and also the predictions across days. Unlike the LSTM or the CNN, our approach also provides further interpretability through the inspection of its parameters.

The state-of-the-art feature-engineering method for fraud classification of electronic payments uses manually engineered feature aggregates, i.e. descriptive statistics of the transaction history. However, this approach has limitations, primarily that of being dependent on expensive human expert knowledge. There have been attempts to replace manual aggregation through automatic feature extraction approaches. They, however, do not consider the specific structure of the manual aggregates. In this paper, we define the novel Neural Aggregate Generator (NAG), a neural network based feature extraction module that learns feature aggregates end-to-end on the fraud classification task. In contrast to other automatic feature extraction approaches, the network architecture of the NAG closely mimics the structure of feature aggregates. Furthermore, the NAG extends learnable aggregates over traditional ones through soft feature value matching, and relative weighting of the importance of different feature constraints. We compare the performance of the NAG to the state-of-the-art approaches on a real-world dataset with millions of transactions. More precisely, we show that features generated with the NAG lead to improved results over manual aggregates for fraud classification, thus demonstrating its viability to replace them. Moreover, we compare the NAG to other automatic approaches such as the LSTM or a generic CNN. Here we also observe improved results, with the NAG requiring far less parameters. Unlike the LSTM or the CNN, our approach also provides further interpretability through the inspection of its parameters.