arXiv subjects

Mathias Vast

Publications and source records attributed to Mathias Vast.

8 recordsLinked to original sources

arXiv2026

How BERT-based Cross-Encoders Model Relevance in Information Retrieval?

Cross-encoders are the most effective but also the least efficient models in Information Retrieval. Despite their success, the mechanisms underlying their relevance predictions remain poorly understood, particularly how they process to match queries and documents. This lack of interpretability limits their optimization, can lead to suboptimal retrieval, especially out-of-domain, and raises efficiency concerns for over-parameterized models. To address these challenges, this paper investigates two under-explored aspects of cross-encoder behavior. First, to understand the relevance prediction process, we trace the flow of information at the different stages of the process and between different input segments. We study the nature of the signals that compose this flow and that allow cross-encoders to predict the relevance of a passage to a query. Our work particularly focuses on the matching performed by the cross-encoders between queries and documents. Second, we explore how cross-encoders behave with out-of-domain documents and/or queries. Through a series of interpretability experiments, we evaluate three BERT-based cross-encoder models, varying in size and pre-training strategies, across both in-domain and out-of-domain datasets. Overall, this study both consolidates what is known about how cross-encoders model relevance and extends it by exploring the role of matching signals.

cs.IR↗

arXiv2026

MICE to Meet You: From Re-ranking Only to End-to-End Retrieval

Late-interaction models offer state-of-the-art efficiency-effectiveness trade-offs for neural retrieval by enabling document representations to be pre-computed and efficiently indexed. Yet, their relatively simple interaction mechanisms, typically MaxSim, can limit their effectiveness compared to cross-encoders. MICE, a recently introduced re-ranking architecture, aims to combine cross-encoder-like effectiveness through expressive transformer-based scoring with reduced inference-time computation through document-side pre-computation. However, no practical way to store these document representations was explored. We address this limitation here and introduce MICE-P, the first end-to-end retrieval pipeline with MICE, based on PLAID to index the contextualized document token representations. We conduct a controlled comparison of MICE, ColBERT and cross-encoders by training comparable models using the Ettin Family under matched experimental conditions. Our results show that MICE-P defines a new class of retrieval architecture with an efficiency-effectiveness trade-off of its own: it matches late-interaction retrieval throughput while retaining the effectiveness of cross-encoders.

cs.IR↗

arXiv2026

MICE: Minimal Interaction Cross-Encoders for efficient Re-ranking

In Information Retrieval (IR), cross-encoders deliver state-of-the-art ranking effectiveness but have a high inference cost, limiting their use to second-stage re-rankers. Prior work has addressed this bottleneck from two largely separate directions: accelerating cross-encoder inference through attention sparsification, or improving first-stage retrieval effectiveness to alleviate the need of a re-ranker, using more complex models, e.g. late-interactions. In this work, we bridge these two directions through an in-depth analysis of cross-encoder internal mechanisms. By identifying and removing superfluous interactions, we derive MICE (Minimal Interaction Cross-Encoders), a new cross-encoder architecture that retains effectiveness while reducing computational overhead. Extensive evaluations show MICE retains most of the performances of its cross-encoder counterparts in-domain and matches or even exceeds it in out-of-domain, while reducing FLOPs down to 2.5 times.

cs.IR↗

arXiv2026

Building Better Encoder-only Cross-Encoders: A Controlled Study of Training Strategies for Neural Re-ranking

Cross-encoders fine-tuned from Transformer backbones remain the standard for second-stage re-ranking, and recent knowledge-distillation strategies have closed much of the gap with LLM re-rankers. However, these strategies have not been compared under controlled conditions. In particular, it remains unclear how distillation from LLM rankers compares to distillation from strong cross-encoder teachers, or to purely supervised objectives. It is also unclear how much newer backbones (RoBERTa, ELECTRA, DeBERTaV3, ModernBERT) contribute compared to the original BERT. We run 162 controlled training runs (9 backbones x 6 objectives x 3 seeds), spanning pointwise, pairwise, and listwise losses with both human labels and two distillation signals, and evaluate on TREC-DL, MSMARCO dev, BEIR, LoTTE, and Robust04. We find that objectives emphasizing relative comparisons - pairwise MarginMSE and listwise InfoNCE - consistently outperform alternative objectives, including more complex listwise LLM distillation, across all backbones, and switching objective yields gains comparable to moving up one backbone size tier. A controlled disentanglement further shows that, once the negative-sampling pool is matched, even a simple pairwise Hinge loss with ColBERTv2 hard negatives matches - and on out of domain beats - listwise LLM distillation, indicating that the quality of the negatives is at least as important as the choice of loss. We release all 162 trained models on HuggingFace (https://huggingface.co/collections/xpmir/reproducing-cross-encoders) and a unified training codebase. (https://github.com/xpmir/cross-encoders)

cs.IR↗

arXiv2026

From Tokens to Concepts: Leveraging SAE for SPLADE

Learned Sparse IR models, such as SPLADE, offer an excellent efficiency-effectiveness tradeoff. However, they rely on the underlying backbone vocabulary, which might hinder performance (polysemicity and synonymy) and pose a challenge for multi-lingual and multi-modal usages. To solve this limitation, we propose to replace the backbone vocabulary with a latent space of semantic concepts learned using Sparse Auto-Encoders (SAE). Throughout this paper, we study the compatibility of these 2 concepts, explore training approaches, and analyze the differences between our SAE-SPLADE model and traditional SPLADE models. Our experiments demonstrate that SAE-SPLADE achieves retrieval performance comparable to SPLADE on both in-domain and out-of-domain tasks while offering improved efficiency.

cs.IR↗

arXiv2025

Understanding Matching Mechanisms in Cross-Encoders

Neural IR architectures, particularly cross-encoders, are highly effective models whose internal mechanisms are mostly unknown. Most works trying to explain their behavior focused on high-level processes (e.g., what in the input influences the prediction, does the model adhere to known IR axioms) but fall short of describing the matching process. Instead of Mechanistic Interpretability approaches which specifically aim at explaining the hidden mechanisms of neural models, we demonstrate that more straightforward methods can already provide valuable insights. In this paper, we first focus on the attention process and extract causal insights highlighting the crucial roles of some attention heads in this process. Second, we provide an interpretation of the mechanism underlying matching detection.

cs.IR↗

arXiv2024

Simple Domain Adaptation for Sparse Retrievers

In Information Retrieval, and more generally in Natural Language Processing, adapting models to specific domains is conducted through fine-tuning. Despite the successes achieved by this method and its versatility, the need for human-curated and labeled data makes it impractical to transfer to new tasks, domains, and/or languages when training data doesn't exist. Using the model without training (zero-shot) is another option that however suffers an effectiveness cost, especially in the case of first-stage retrievers. Numerous research directions have emerged to tackle these issues, most of them in the context of adapting to a task or a language. However, the literature is scarcer for domain (or topic) adaptation. In this paper, we address this issue of cross-topic discrepancy for a sparse first-stage retriever by transposing a method initially designed for language adaptation. By leveraging pre-training on the target data to learn domain-specific knowledge, this technique alleviates the need for annotated data and expands the scope of domain adaptation. Despite their relatively good generalization ability, we show that even sparse retrievers can benefit from our simple domain adaptation method.

cs.IR↗

arXiv2024

Which Neurons Matter in IR? Applying Integrated Gradients-based Methods to Understand Cross-Encoders

With the recent addition of Retrieval-Augmented Generation (RAG), the scope and importance of Information Retrieval (IR) has expanded. As a result, the importance of a deeper understanding of IR models also increases. However, interpretability in IR remains under-explored, especially when it comes to the models' inner mechanisms. In this paper, we explore the possibility of adapting Integrated Gradient-based methods in an IR context to identify the role of individual neurons within the model. In particular, we provide new insights into the role of what we call "relevance" neurons, as well as how they deal with unseen data. Finally, we carry out an in-depth pruning study to validate our findings.

cs.IR↗