Shengyao Zhuang

Shengyao Zhuang, Honglei Zhuang, Bevan Koopman et al. · deepmind

We propose a novel zero-shot document ranking approach based on Large Language Models (LLMs): the Setwise prompting approach. Our approach complements existing prompting approaches for LLM-based zero-shot ranking: Pointwise, Pairwise, and Listwise. Through the first-of-its-kind comparative evaluation within a consistent experimental framework and considering factors like model size, token consumption, latency, among others, we show that existing approaches are inherently characterised by trade-offs between effectiveness and efficiency. We find that while Pointwise approaches score high on efficiency, they suffer from poor effectiveness. Conversely, Pairwise approaches demonstrate superior effectiveness but incur high computational overhead. Our Setwise approach, instead, reduces the number of LLM inferences and the amount of prompt token consumption during the ranking procedure, compared to previous methods. This significantly improves the efficiency of LLM-based zero-shot ranking, while also retaining high zero-shot ranking effectiveness. We make our code and results publicly available at \url{https://github.com/ielab/llm-rankers}.

Shengyao Zhuang, Bing Liu, Bevan Koopman et al.

In the field of information retrieval, Query Likelihood Models (QLMs) rank documents based on the probability of generating the query given the content of a document. Recently, advanced large language models (LLMs) have emerged as effective QLMs, showcasing promising ranking capabilities. This paper focuses on investigating the genuine zero-shot ranking effectiveness of recent LLMs, which are solely pre-trained on unstructured text data without supervised instruction fine-tuning. Our findings reveal the robust zero-shot ranking ability of such LLMs, highlighting that additional instruction fine-tuning may hinder effectiveness unless a question generation task is present in the fine-tuning dataset. Furthermore, we introduce a novel state-of-the-art ranking system that integrates LLM-based QLMs with a hybrid zero-shot retriever, demonstrating exceptional effectiveness in both zero-shot and few-shot scenarios. We make our codebase publicly available at https://github.com/ielab/llm-qlm.

Shengyao Zhuang, Guido Zuccon

Current dense retrievers are not robust to out-of-domain and outlier queries, i.e. their effectiveness on these queries is much poorer than what one would expect. In this paper, we consider a specific instance of such queries: queries that contain typos. We show that a small character level perturbation in queries (as caused by typos) highly impacts the effectiveness of dense retrievers. We then demonstrate that the root cause of this resides in the input tokenization strategy employed by BERT. In BERT, tokenization is performed using the BERT's WordPiece tokenizer and we show that a token with a typo will significantly change the token distributions obtained after tokenization. This distribution change translates to changes in the input embeddings passed to the BERT-based query encoder of dense retrievers. We then turn our attention to devising dense retriever methods that are robust to such queries with typos, while still being as performant as previous methods on queries without typos. For this, we use CharacterBERT as the backbone encoder and an efficient yet effective training method, called Self-Teaching (ST), that distills knowledge from queries without typos into the queries with typos. Experimental results show that CharacterBERT in combination with ST achieves significantly higher effectiveness on queries with typos compared to previous methods. Along with these results and the open-sourced implementation of the methods, we also provide a new passage retrieval dataset consisting of real-world queries with typos and associated relevance assessments on the MS MARCO corpus, thus supporting the research community in the investigation of effective and robust dense retrievers. Code, experimental results and dataset are made available at https://github.com/ielab/CharacterBERT-DR.

Haodong Chen, Shengyao Zhuang, Zheng Yao et al.

Zero-shot document re-ranking with Large Language Models (LLMs) has evolved from Pointwise methods to Listwise and Setwise approaches that optimize computational efficiency. Despite their success, these methods predominantly rely on generative scoring or output logits, which face bottlenecks in inference latency and result consistency. In-Context Re-ranking (ICR) has recently been proposed as an O(1) alternative method. ICR extracts internal attention signals directly, avoiding the overhead of text generation. However, existing ICR methods simply aggregate signals across all layers; layer-wise contributions and their consistency across architectures have been left unexplored. Furthermore, no unified study has compared internal attention with traditional generative and likelihood-based mechanisms across diverse ranking frameworks under consistent conditions. In this paper, we conduct an orthogonal evaluation of generation, likelihood, and internal attention mechanisms across multiple ranking frameworks. We further identify a universal "bell-curve" distribution of relevance signals across transformer layers, which motivates the proposed Selective-ICR strategy that reduces inference latency by 30%-50% without compromising effectiveness. Finally, evaluation on the reasoning-intensive BRIGHT benchmark shows that precisely capturing high-quality in-context attention signals fundamentally reduces the need for model scaling and reinforcement learning: a zero-shot 8B model matches the performance of 14B reinforcement-learned re-rankers, while even a 0.6B model outperforms state-of-the-art generation-based approaches. These findings redefine the efficiency-effectiveness frontier for LLM-based re-ranking and highlight the latent potential of internal signals for complex reasoning ranking tasks. Our code and results are publicly available at https://github.com/ielab/Selective-ICR.

Shengyao Zhuang, Houxing Ren, Linjun Shou et al.

The Differentiable Search Index (DSI) is an emerging paradigm for information retrieval. Unlike traditional retrieval architectures where index and retrieval are two different and separate components, DSI uses a single transformer model to perform both indexing and retrieval. In this paper, we identify and tackle an important issue of current DSI models: the data distribution mismatch that occurs between the DSI indexing and retrieval processes. Specifically, we argue that, at indexing, current DSI methods learn to build connections between the text of long documents and the identifier of the documents, but then retrieval of document identifiers is based on queries that are commonly much shorter than the indexed documents. This problem is further exacerbated when using DSI for cross-lingual retrieval, where document text and query text are in different languages. To address this fundamental problem of current DSI models, we propose a simple yet effective indexing framework for DSI, called DSI-QG. When indexing, DSI-QG represents documents with a number of potentially relevant queries generated by a query generation model and re-ranked and filtered by a cross-encoder ranker. The presence of these queries at indexing allows the DSI models to connect a document identifier to a set of queries, hence mitigating data distribution mismatches present between the indexing and the retrieval phases. Empirical results on popular mono-lingual and cross-lingual passage retrieval datasets show that DSI-QG significantly outperforms the original DSI model.

Ekaterina Khramtsova, Shengyao Zhuang, Mahsa Baktashmotlagh et al.

We propose the new problem of choosing which dense retrieval model to use when searching on a new collection for which no labels are available, i.e. in a zero-shot setting. Many dense retrieval models are readily available. Each model however is characterized by very differing search effectiveness -- not just on the test portion of the datasets in which the dense representations have been learned but, importantly, also across different datasets for which data was not used to learn the dense representations. This is because dense retrievers typically require training on a large amount of labeled data to achieve satisfactory search effectiveness in a specific dataset or domain. Moreover, effectiveness gains obtained by dense retrievers on datasets for which they are able to observe labels during training, do not necessarily generalise to datasets that have not been observed during training. This is however a hard problem: through empirical experimentation we show that methods inspired by recent work in unsupervised performance evaluation with the presence of domain shift in the area of computer vision and machine learning are not effective for choosing highly performing dense retrievers in our setup. The availability of reliable methods for the selection of dense retrieval models in zero-shot settings that do not require the collection of labels for evaluation would allow to streamline the widespread adoption of dense retrieval. This is therefore an important new problem we believe the information retrieval community should consider. Implementation of methods, along with raw result files and analysis scripts are made publicly available at https://www.github.com/anonymized.

Zijian Chen, Xueguang Ma, Shengyao Zhuang et al.

Deep Research agents are rapidly emerging as primary consumers of modern retrieval systems. Unlike human users who issue and refine queries without documenting their intermediate thought processes, Deep Research agents generate explicit natural language reasoning before each search call, revealing rich intent and contextual information that existing retrievers entirely ignore. To exploit this overlooked signal, we introduce: (1) Reasoning-Aware Retrieval, a retrieval paradigm that jointly embeds the agent's reasoning trace alongside its query; and (2) DR-Synth, a data synthesis method that generates Deep Research retriever training data from standard QA datasets. We demonstrate that both components are independently effective, and their combination yields a trained embedding model, AgentIR-4B, with substantial gains. On the challenging BrowseComp-Plus benchmark, AgentIR-4B achieves 68\% accuracy with the open-weight agent Tongyi-DeepResearch, compared to 50\% with conventional embedding models twice its size, and 37\% with BM25. Code and data are available at: https://texttron.github.io/AgentIR/.

Shuai Wang, Yin Yu, Shengyao Zhuang et al.

PromptReps showed that an autoregressive language model can be used directly as a retriever by prompting it to generate dense and sparse representations of a query or passage. Extending this to multiple representatives is inefficient for autoregressive models, since tokens must be generated sequentially, and prior multi-token variants did not reliably improve over single-token decoding. We show that the bottleneck is sequential generation, not the multi-token idea itself. DiffRetriever is a representative-token retriever for diffusion language models: it appends K masked positions to the prompt and reads all K in a single bidirectional forward pass. Across in-domain and out-of-domain evaluation, multi-token DiffRetriever substantially improves over single-token on every diffusion backbone we test, while autoregressive multi-token is flat or negative and pays a latency cost that scales with K where diffusion does not. After supervised fine-tuning, DiffRetriever on Dream is the strongest BEIR-7 retriever in our comparison, ahead of PromptReps, the encoder-style DiffEmbed baseline on the same diffusion backbones, and the contrastively fine-tuned single-vector RepLLaMA. A per-query oracle on the frozen base model exceeds contrastive fine-tuning at the same fixed budget, pointing to adaptive budget selection as future work. Code is available at https://github.com/ielab/diffretriever.

Shengyao Zhuang, zhichao Xu, Ivano Lauriola

Transformer-based document cross-encoder rerankers are a central component of modern information retrieval systems. Despite their success, these models suffer from high computational costs due to processing long query-document sequences at inference time. A known approach to improve efficiency is token compression, which consists of aggregating groups of tokens together in the initial embedding layer, reducing the effective number of tokens, and making the computation faster. While token compression has proven to be successful for bi-encoder retrievers, we empirically observed that this approach may be ineffective for cross-encoder rerankers. In this paper, we propose Layer-wise Token Compression (LTC), which applies adaptive token pooling at intermediate transformer layers. Through extensive ablation studies on MS MARCO passage and document ranking tasks, we demonstrate that compression at middle layers preserves ranking quality while increasing inference QPS by up to 25% for passage ranking and up to 116% for document ranking. We also extend LTC to listwise LLM rerankers and show that the same approach can be easily applied to long-context listwise reranking, where the QPS improvements are even greater. More surprisingly, when applying rerankers trained on short passages to long-document ranking tasks, models trained with compression outperform their uncompressed counterparts, suggesting that compression may act as a beneficial regularizer that encourages length-invariant representations.

Zijian Chen, Xueguang Ma, Shengyao Zhuang et al.

Deep-Research agents, which integrate large language models (LLMs) with search tools, have shown success in improving the effectiveness of handling complex queries that require iterative search planning and reasoning over search results. Evaluations on current benchmarks like BrowseComp relies on black-box live web search APIs, have notable limitations in (1) fairness: dynamic and opaque web APIs hinder fair comparisons and reproducibility of deep research methods; (2) transparency: lack of control over the document corpus makes it difficult to isolate retriever contributions. In other words, the current evaluations may compare a complete deep research system at a given time, but they do not foster well-controlled experiments to provide insights into the capability of underlying deep research LLMs. To address these challenges, we introduce BrowseComp-Plus, a benchmark derived from BrowseComp, employing a fixed, carefully curated corpus. Each query in BrowseComp-Plus includes human-verified supporting documents and mined challenging negatives, enabling controlled experimentation. The benchmark is shown to be effective in distinguishing the performance of deep research systems. For instance, the open-source model Search-R1, when paired with the BM25 retriever, achieves 3.86% accuracy, whereas the GPT-5 achieves 55.9%. Integrating the GPT-5 with the Qwen3-Embedding-8B retriever further enhances its accuracy to 70.1% with fewer search calls. This benchmark allows comprehensive evaluation and disentangled analysis of deep research agents and retrieval methods, fostering insights into retrieval effectiveness, citation accuracy, and context engineering in Deep-Research system.

Zhichao Xu, Zhiqi Huang, Shengyao Zhuang et al.

Training effective text rerankers is crucial for information retrieval. Two strategies are widely used: contrastive learning (optimizing directly on ground-truth labels) and knowledge distillation (transferring knowledge from a larger reranker). While both have been studied extensively, a clear comparison of their effectiveness for training cross-encoder rerankers under practical conditions is needed. This paper empirically compares these strategies by training rerankers of different sizes (0.5B, 1.5B, 3B, 7B) and architectures (Transformer, Recurrent) using both methods on the same data, with a strong contrastive learning model acting as the distillation teacher. Our results show that knowledge distillation generally yields better in-domain and out-of-domain ranking performance than contrastive learning when distilling from a more performant teacher model. This finding is consistent across student model sizes and architectures. However, distilling from a teacher of the same capacity does not provide the same advantage, particularly for out-of-domain tasks. These findings offer practical guidance for choosing a training strategy based on available teacher models. We recommend using knowledge distillation to train smaller rerankers if a larger, more performant teacher is accessible; in its absence, contrastive learning remains a robust baseline. Our code implementation is made available to facilitate reproducbility.

Shuai Wang, Shengyao Zhuang, Bevan Koopman et al.

2D Matryoshka Training is an advanced embedding representation training approach designed to train an encoder model simultaneously across various layer-dimension setups. This method has demonstrated higher effectiveness in Semantic Text Similarity (STS) tasks over traditional training approaches when using sub-layers for embeddings. Despite its success, discrepancies exist between two published implementations, leading to varied comparative results with baseline models. In this reproducibility study, we implement and evaluate both versions of 2D Matryoshka Training on STS tasks and extend our analysis to retrieval tasks. Our findings indicate that while both versions achieve higher effectiveness than traditional Matryoshka training on sub-dimensions, and traditional full-sized model training approaches, they do not outperform models trained separately on specific sub-layer and sub-dimension setups. Moreover, these results generalize well to retrieval tasks, both in supervised (MSMARCO) and zero-shot (BEIR) settings. Further explorations of different loss computations reveals more suitable implementations for retrieval tasks, such as incorporating full-dimension loss and training on a broader range of target dimensions. Conversely, some intuitive approaches, such as fixing document encoders to full model outputs, do not yield improvements. Our reproduction code is available at https://github.com/ielab/2DMSE-Reproduce.

Shengyao Zhuang, Guido Zuccon

The process of model checkpoint validation refers to the evaluation of the performance of a model checkpoint executed on a held-out portion of the training data while learning the hyperparameters of the model, and is used to avoid over-fitting and determine when the model has converged so as to stop training. A simple and efficient strategy to validate deep learning checkpoints is the addition of validation loops to execute during training. However, the validation of dense retrievers (DR) checkpoints is not as trivial -- and the addition of validation loops is not efficient. This is because, in order to accurately evaluate the performance of a DR checkpoint, the whole document corpus needs to be encoded into vectors using the current checkpoint before any actual retrieval operation for checkpoint validation can be performed. This corpus encoding process can be very time-consuming if the document corpus contains millions of documents (e.g., 8.8m for MS MARCO and 21m for Natural Questions). Thus, a naive use of validation loops during training will significantly increase training time. To address this issue, in this demo paper, we propose Asyncval: a Python-based toolkit for efficiently validating DR checkpoints during training. Instead of pausing the training loop for validating DR checkpoints, Asyncval decouples the validation loop from the training loop, uses another GPU to automatically validate new DR checkpoints and thus permits to perform validation asynchronously from training. Asyncval also implements a range of different corpus subset sampling strategies for validating DR checkpoints; these strategies allow to further speed up the validation process. We provide an investigation of these methods in terms of their impact on validation time and validation fidelity. Asyncval is made available as an open-source project at https://github.com/ielab/asyncval.

Shengyao Zhuang, Zhihao Qiao, Guido Zuccon

Online learning to rank (OLTR) aims to learn a ranker directly from implicit feedback derived from users' interactions, such as clicks. Clicks however are a biased signal: specifically, top-ranked documents are likely to attract more clicks than documents down the ranking (position bias). In this paper, we propose a novel learning algorithm for OLTR that uses reinforcement learning to optimize rankers: Reinforcement Online Learning to Rank (ROLTR). In ROLTR, the gradients of the ranker are estimated based on the rewards assigned to clicked and unclicked documents. In order to de-bias the users' position bias contained in the reward signals, we introduce unbiased reward shaping functions that exploit inverse propensity scoring for clicked and unclicked documents. The fact that our method can also model unclicked documents provides a further advantage in that less users interactions are required to effectively train a ranker, thus providing gains in efficiency. Empirical evaluation on standard OLTR datasets shows that ROLTR achieves state-of-the-art performance, and provides significantly better user experience than other OLTR approaches. To facilitate the reproducibility of our experiments, we make all experiment code available at https://github.com/ielab/OLTR.

Shengyao Zhuang, Xueguang Ma, Bevan Koopman et al.

In this paper, we introduce Rank-R1, a novel LLM-based reranker that performs reasoning over both the user query and candidate documents before performing the ranking task. Existing document reranking methods based on large language models (LLMs) typically rely on prompting or fine-tuning LLMs to order or label candidate documents according to their relevance to a query. For Rank-R1, we use a reinforcement learning algorithm along with only a small set of relevance labels (without any reasoning supervision) to enhance the reasoning ability of LLM-based rerankers. Our hypothesis is that adding reasoning capabilities to the rerankers can improve their relevance assessement and ranking capabilities. Our experiments on the TREC DL and BRIGHT datasets show that Rank-R1 is highly effective, especially for complex queries. In particular, we find that Rank-R1 achieves effectiveness on in-domain datasets at par with that of supervised fine-tuning methods, but utilizing only 18\% of the training data used by the fine-tuning methods. We also find that the model largely outperforms zero-shot and supervised fine-tuning when applied to out-of-domain datasets featuring complex queries, especially when a 14B-size model is used. Finally, we qualitatively observe that Rank-R1's reasoning process improves the explainability of the ranking results, opening new opportunities for search engine results presentation and fruition.

Shuai Wang, Harrisen Scells, Shengyao Zhuang et al.

Systematic reviews are crucial for evidence-based medicine as they comprehensively analyse published research findings on specific questions. Conducting such reviews is often resource- and time-intensive, especially in the screening phase, where abstracts of publications are assessed for inclusion in a review. This study investigates the effectiveness of using zero-shot large language models~(LLMs) for automatic screening. We evaluate the effectiveness of eight different LLMs and investigate a calibration technique that uses a predefined recall threshold to determine whether a publication should be included in a systematic review. Our comprehensive evaluation using five standard test collections shows that instruction fine-tuning plays an important role in screening, that calibration renders LLMs practical for achieving a targeted recall, and that combining both with an ensemble of zero-shot models saves significant screening time compared to state-of-the-art approaches.

Shuai Wang, Ekaterina Khramtsova, Shengyao Zhuang et al.

Federated search systems aggregate results from multiple search engines, selecting appropriate sources to enhance result quality and align with user intent. With the increasing uptake of Retrieval-Augmented Generation (RAG) pipelines, federated search can play a pivotal role in sourcing relevant information across heterogeneous data sources to generate informed responses. However, existing datasets, such as those developed in the past TREC FedWeb tracks, predate the RAG paradigm shift and lack representation of modern information retrieval challenges. To bridge this gap, we present FeB4RAG, a novel dataset specifically designed for federated search within RAG frameworks. This dataset, derived from 16 sub-collections of the widely used \beir benchmarking collection, includes 790 information requests (akin to conversational queries) tailored for chatbot applications, along with top results returned by each resource and associated LLM-derived relevance judgements. Additionally, to support the need for this collection, we demonstrate the impact on response generation of a high quality federated search system for RAG compared to a naive approach to federated search. We do so by comparing answers generated through the RAG pipeline through a qualitative side-by-side comparison. Our collection fosters and supports the development and evaluation of new federated search methods, especially in the context of RAG pipelines.

Shuai Wang, Shengyao Zhuang, Bevan Koopman et al.

Federated search, which involves integrating results from multiple independent search engines, will become increasingly pivotal in the context of Retrieval-Augmented Generation pipelines empowering LLM-based applications such as chatbots. These systems often distribute queries among various search engines, ranging from specialized (e.g., PubMed) to general (e.g., Google), based on the nature of user utterances. A critical aspect of federated search is resource selection - the selection of appropriate resources prior to issuing the query to ensure high-quality and rapid responses, and contain costs associated with calling the external search engines. However, current SOTA resource selection methodologies primarily rely on feature-based learning approaches. These methods often involve the labour intensive and expensive creation of training labels for each resource. In contrast, LLMs have exhibited strong effectiveness as zero-shot methods across NLP and IR tasks. We hypothesise that in the context of federated search LLMs can assess the relevance of resources without the need for extensive predefined labels or features. In this paper, we propose ReSLLM. Our ReSLLM method exploits LLMs to drive the selection of resources in federated search in a zero-shot setting. In addition, we devise an unsupervised fine tuning protocol, the Synthetic Label Augmentation Tuning (SLAT), where the relevance of previously logged queries and snippets from resources is predicted using an off-the-shelf LLM and then in turn used to fine-tune ReSLLM with respect to resource selection. Our empirical evaluation and analysis details the factors influencing the effectiveness of LLMs in this context. The results showcase the merits of ReSLLM for resource selection: not only competitive effectiveness in the zero-shot setting, but also obtaining large when fine-tuned using SLAT-protocol.

Shengyao Zhuang, Bevan Koopman, Xiaoran Chu et al.

The emergence of Vec2Text -- a method for text embedding inversion -- has raised serious privacy concerns for dense retrieval systems which use text embeddings, such as those offered by OpenAI and Cohere. This threat comes from the ability for a malicious attacker with access to embeddings to reconstruct the original text. In this paper, we investigate various factors related to embedding models that may impact text recoverability via Vec2Text. We explore factors such as distance metrics, pooling functions, bottleneck pre-training, training with noise addition, embedding quantization, and embedding dimensions, which were not considered in the original Vec2Text paper. Through a comprehensive analysis of these factors, our objective is to gain a deeper understanding of the key elements that affect the trade-offs between the text recoverability and retrieval effectiveness of dense retrieval systems, offering insights for practitioners designing privacy-aware dense retrieval systems. We also propose a simple embedding transformation fix that guarantees equal ranking effectiveness while mitigating the recoverability risk. Overall, this study reveals that Vec2Text could pose a threat to current dense retrieval systems, but there are some effective methods to patch such systems.

Hang Li, Shengyao Zhuang, Bevan Koopman et al.

Vector Pseudo Relevance Feedback (VPRF) has shown promising results in improving BERT-based dense retrieval systems through iterative refinement of query representations. This paper investigates the generalizability of VPRF to Large Language Model (LLM) based dense retrievers. We introduce LLM-VPRF and evaluate its effectiveness across multiple benchmark datasets, analyzing how different LLMs impact the feedback mechanism. Our results demonstrate that VPRF's benefits successfully extend to LLM architectures, establishing it as a robust technique for enhancing dense retrieval performance regardless of the underlying models. This work bridges the gap between VPRF with traditional BERT-based dense retrievers and modern LLMs, while providing insights into their future directions.

Shuai Wang, Shengyao Zhuang, Guido Zuccon

Text stemming is a natural language processing technique that is used to reduce words to their base form, also known as the root form. The use of stemming in IR has been shown to often improve the effectiveness of keyword-matching models such as BM25. However, traditional stemming methods, focusing solely on individual terms, overlook the richness of contextual information. Recognizing this gap, in this paper, we investigate the promising idea of using large language models (LLMs) to stem words by leveraging its capability of context understanding. With this respect, we identify three avenues, each characterised by different trade-offs in terms of computational cost, effectiveness and robustness : (1) use LLMs to stem the vocabulary for a collection, i.e., the set of unique words that appear in the collection (vocabulary stemming), (2) use LLMs to stem each document separately (contextual stemming), and (3) use LLMs to extract from each document entities that should not be stemmed, then use vocabulary stemming to stem the rest of the terms (entity-based contextual stemming). Through a series of empirical experiments, we compare the use of LLMs for stemming with that of traditional lexical stemmers such as Porter and Krovetz for English text. We find that while vocabulary stemming and contextual stemming fail to achieve higher effectiveness than traditional stemmers, entity-based contextual stemming can achieve a higher effectiveness than using Porter stemmer alone, under specific conditions.

Zhichao Xu, Shengyao Zhuang, Crystina Zhang et al.

While dense retrieval models have become the standard for state-of-the-art information retrieval, their deployment is often constrained by high memory requirements and reliance on GPU accelerators for vector similarity search. Learned sparse retrieval offers a compelling alternative by enabling efficient search via inverted indices, yet it has historically received less attention than dense approaches. In this report, we introduce LACONIC, a family of learned sparse retrievers based on the Llama-3 architecture (1B, 3B, and 8B). We propose a streamlined two-phase training curriculum consisting of (1) weakly supervised pre-finetuning to adapt causal LLMs for bidirectional contextualization and (2) high-signal finetuning using curated hard negatives. Our results demonstrate that LACONIC effectively bridges the performance gap with dense models: the 8B variant achieves a state-of-the-art 60.2 nDCG on the MTEB Retrieval benchmark, ranking 15th on the leaderboard as of January 1, 2026, while utilizing 71\% less index memory than an equivalent dense model. By delivering high retrieval effectiveness on commodity CPU hardware with a fraction of the compute budget required by competing models, LACONIC provides a scalable and efficient solution for real-world search applications.

Zhichao Xu, Shengyao Zhuang, Xueguang Ma et al.

Recent advances in large language models (LLMs) have led to a surge of interest in query augmentation for information retrieval (IR). Two main approaches have emerged. The first prompts LLMs to generate answers or pseudo-documents that serve as new queries, relying purely on the model's parametric knowledge or contextual information. The second applies reinforcement learning (RL) to fine-tune LLMs for query rewriting, directly optimizing retrieval metrics. While having respective advantages and limitations, the two approaches have not been compared under consistent experimental conditions. In this work, we present the first systematic comparison of prompting-based and RL-based query augmentation across diverse benchmarks, including evidence-seeking, ad hoc, and tool retrieval. Our key finding is that simple, training-free query augmentation often performs on par with, or even surpasses, more expensive RL-based counterparts, especially when using powerful LLMs. Motivated by this discovery, we introduce a novel hybrid method, On-policy Pseudo-document Query Expansion (OPQE), which, instead of rewriting a query, the LLM policy learns to generate a pseudo-document that maximizes retrieval performance, thus merging the flexibility and generative structure of prompting with the targeted optimization of RL. We show OPQE outperforms both standalone prompting and RL-based rewriting, demonstrating that a synergistic approach yields the best results. Our implementation is made available to facilitate reproducibility.

Shuoqi Sun, Shengyao Zhuang, Shuai Wang et al.

We provide a systematic understanding of the impact of specific components and wordings used in prompts on the effectiveness of rankers based on zero-shot Large Language Models (LLMs). Several zero-shot ranking methods based on LLMs have recently been proposed. Among many aspects, methods differ across (1) the ranking algorithm they implement, e.g., pointwise vs. listwise, (2) the backbone LLMs used, e.g., GPT3.5 vs. FLAN-T5, (3) the components and wording used in prompts, e.g., the use or not of role-definition (role-playing) and the actual words used to express this. It is currently unclear whether performance differences are due to the underlying ranking algorithm, or because of spurious factors such as better choice of words used in prompts. This confusion risks to undermine future research. Through our large-scale experimentation and analysis, we find that ranking algorithms do contribute to differences between methods for zero-shot LLM ranking. However, so do the LLM backbones -- but even more importantly, the choice of prompt components and wordings affect the ranking. In fact, in our experiments, we find that, at times, these latter elements have more impact on the ranker's effectiveness than the actual ranking algorithms, and that differences among ranking methods become more blurred when prompt variations are considered.

Aaron Nicolson, Shengyao Zhuang, Jason Dowling et al.

This study investigates the integration of diverse patient data sources into multimodal language models for automated chest X-ray (CXR) report generation. Traditionally, CXR report generation relies solely on CXR images and limited radiology data, overlooking valuable information from patient health records, particularly from emergency departments. Utilising the MIMIC-CXR and MIMIC-IV-ED datasets, we incorporate detailed patient information such as vital signs, medicines, and clinical history to enhance diagnostic accuracy. We introduce a novel approach to transform these heterogeneous data sources into embeddings that prompt a multimodal language model; this significantly enhances the diagnostic accuracy of generated radiology reports. Our comprehensive evaluation demonstrates the benefits of using a broader set of patient data, underscoring the potential for enhanced diagnostic capabilities and better patient outcomes through the integration of multimodal data in CXR report generation.

Hang Li, Shengyao Zhuang, Ahmed Mourad et al.

Pseudo-Relevance Feedback (PRF) utilises the relevance signals from the top-k passages from the first round of retrieval to perform a second round of retrieval aiming to improve search effectiveness. A recent research direction has been the study and development of PRF methods for deep language models based rankers, and in particular in the context of dense retrievers. Dense retrievers, compared to more complex neural rankers, provide a trade-off between effectiveness, which is often reduced compared to more complex neural rankers, and query latency, which also is reduced making the retrieval pipeline more efficient. The introduction of PRF methods for dense retrievers has been motivated as an attempt to further improve their effectiveness. In this paper, we reproduce and study a recent method for PRF with dense retrievers, called ANCE-PRF. This method concatenates the query text and that of the top-k feedback passages to form a new query input, which is then encoded into a dense representation using a newly trained query encoder based on the original dense retriever used for the first round of retrieval. While the method can potentially be applied to any of the existing dense retrievers, prior work has studied it only in the context of the ANCE dense retriever. We study the reproducibility of ANCE-PRF in terms of both its training (encoding of the PRF signal) and inference (ranking) steps. We further extend the empirical analysis provided in the original work to investigate the effect of the hyper-parameters that govern the training process and the robustness of the method across these different settings. Finally, we contribute a study of the generalisability of the ANCE-PRF method when dense retrievers other than ANCE are used for the first round of retrieval and for encoding the PRF signal.

Shengyao Zhuang, Guido Zuccon

Passage retrieval and ranking is a key task in open-domain question answering and information retrieval. Current effective approaches mostly rely on pre-trained deep language model-based retrievers and rankers. These methods have been shown to effectively model the semantic matching between queries and passages, also in presence of keyword mismatch, i.e. passages that are relevant to a query but do not contain important query keywords. In this paper we consider the Dense Retriever (DR), a passage retrieval method, and the BERT re-ranker, a popular passage re-ranking method. In this context, we formally investigate how these models respond and adapt to a specific type of keyword mismatch -- that caused by keyword typos occurring in queries. Through empirical investigation, we find that typos can lead to a significant drop in retrieval and ranking effectiveness. We then propose a simple typos-aware training framework for DR and BERT re-ranker to address this issue. Our experimental results on the MS MARCO passage ranking dataset show that, with our proposed typos-aware training, DR and BERT re-ranker can become robust to typos in queries, resulting in significantly improved effectiveness compared to models trained without appropriately accounting for typos.

Hang Li, Ahmed Mourad, Shengyao Zhuang et al.

Pseudo Relevance Feedback (PRF) is known to improve the effectiveness of bag-of-words retrievers. At the same time, deep language models have been shown to outperform traditional bag-of-words rerankers. However, it is unclear how to integrate PRF directly with emergent deep language models. In this article, we address this gap by investigating methods for integrating PRF signals into rerankers and dense retrievers based on deep language models. We consider text-based and vector-based PRF approaches, and investigate different ways of combining and scoring relevance signals. An extensive empirical evaluation was conducted across four different datasets and two task settings (retrieval and ranking). Text-based PRF results show that the use of PRF had a mixed effect on deep rerankers across different datasets. We found that the best effectiveness was achieved when (i) directly concatenating each PRF passage with the query, searching with the new set of queries, and then aggregating the scores; (ii) using Borda to aggregate scores from PRF runs. Vector-based PRF results show that the use of PRF enhanced the effectiveness of deep rerankers and dense retrievers over several evaluation metrics. We found that higher effectiveness was achieved when (i) the query retains either the majority or the same weight within the PRF mechanism, and (ii) a shallower PRF signal (i.e., a smaller number of top-ranked passages) was employed, rather than a deeper signal. Our vector-based PRF method is computationally efficient; thus this represents a general PRF method others can use with deep rerankers and dense retrievers.

Shengyao Zhuang, Guido Zuccon

BERT-based information retrieval models are expensive, in both time (query latency) and computational resources (energy, hardware cost), making many of these models impractical especially under resource constraints. The reliance on a query encoder that only performs tokenization and on the pre-processing of passage representations at indexing, has allowed the recently proposed TILDE method to overcome the high query latency issue typical of BERT-based models. This however is at the expense of a lower effectiveness compared to other BERT-based re-rankers and dense retrievers. In addition, the original TILDE method is characterised by indexes with a very high memory footprint, as it expands each passage into the size of the BERT vocabulary. In this paper, we propose TILDEv2, a new model that stems from the original TILDE but that addresses its limitations. TILDEv2 relies on contextualized exact term matching with expanded passages. This requires to only store in the index the score of tokens that appear in the expanded passages (rather than all the vocabulary), thus producing indexes that are 99% smaller than those of TILDE. This matching mechanism also improves ranking effectiveness by 24%, without adding to the query latency. This makes TILDEv2 the state-of-the-art passage re-ranking method for CPU-only environments, capable of maintaining query latency below 100ms on commodity hardware.