Kai Hui

Zhen Qin, Rolf Jagerman, Kai Hui et al. · deepmind

Ranking documents using Large Language Models (LLMs) by directly feeding the query and candidate documents into the prompt is an interesting and practical problem. However, researchers have found it difficult to outperform fine-tuned baseline rankers on benchmark datasets. We analyze pointwise and listwise ranking prompts used by existing methods and argue that off-the-shelf LLMs do not fully understand these challenging ranking formulations. In this paper, we propose to significantly reduce the burden on LLMs by using a new technique called Pairwise Ranking Prompting (PRP). Our results are the first in the literature to achieve state-of-the-art ranking performance on standard benchmarks using moderate-sized open-sourced LLMs. On TREC-DL 2019&2020, PRP based on the Flan-UL2 model with 20B parameters performs favorably with the previous best approach in the literature, which is based on the blackbox commercial GPT-4 that has 50x (estimated) model size, while outperforming other LLM-based solutions, such as InstructGPT which has 175B parameters, by over 10% for all ranking metrics. By using the same prompt template on seven BEIR tasks, PRP outperforms supervised baselines and outperforms the blackbox commercial ChatGPT solution by 4.2% and pointwise LLM-based solutions by more than 10% on average NDCG@10. Furthermore, we propose several variants of PRP to improve efficiency and show that it is possible to achieve competitive results even with linear complexity.

Bernd Bohnet, Vinh Q. Tran, Pat Verga et al. · deepmind, mit

Large language models (LLMs) have shown impressive results while requiring little or no direct supervision. Further, there is mounting evidence that LLMs may have potential in information-seeking scenarios. We believe the ability of an LLM to attribute the text that it generates is likely to be crucial in this setting. We formulate and study Attributed QA as a key first step in the development of attributed LLMs. We propose a reproducible evaluation framework for the task and benchmark a broad set of architectures. We take human annotations as a gold standard and show that a correlated automatic metric is suitable for development. Our experimental work gives concrete answers to two key questions (How to measure attribution?, and How well do current state-of-the-art methods perform on attribution?), and give some hints as to how to address a third (How to build LLMs with attribution?).

Honglei Zhuang, Zhen Qin, Rolf Jagerman et al. · deepmind

Recently, substantial progress has been made in text ranking based on pretrained language models such as BERT. However, there are limited studies on how to leverage more powerful sequence-to-sequence models such as T5. Existing attempts usually formulate text ranking as classification and rely on postprocessing to obtain a ranked list. In this paper, we propose RankT5 and study two T5-based ranking model structures, an encoder-decoder and an encoder-only one, so that they not only can directly output ranking scores for each query-document pair, but also can be fine-tuned with "pairwise" or "listwise" ranking losses to optimize ranking performances. Our experiments show that the proposed models with ranking losses can achieve substantial ranking performance gains on different public text ranking data sets. Moreover, when fine-tuned with listwise ranking losses, the ranking model appears to have better zero-shot ranking performance on out-of-domain data sets compared to the model fine-tuned with classification losses.

Kai Hui, Honglei Zhuang, Tao Chen et al. · deepmind

State-of-the-art neural models typically encode document-query pairs using cross-attention for re-ranking. To this end, models generally utilize an encoder-only (like BERT) paradigm or an encoder-decoder (like T5) approach. These paradigms, however, are not without flaws, i.e., running the model on all query-document pairs at inference-time incurs a significant computational cost. This paper proposes a new training and inference paradigm for re-ranking. We propose to finetune a pretrained encoder-decoder model using in the form of document to query generation. Subsequently, we show that this encoder-decoder architecture can be decomposed into a decoder-only language model during inference. This results in significant inference time speedups since the decoder-only architecture only needs to learn to interpret static encoder embeddings during inference. Our experiments show that this new paradigm achieves results that are comparable to the more expensive cross-attention ranking approaches while being up to 6.8X faster. We believe this work paves the way for more efficient neural rankers that leverage large pretrained models.

Minghan Li, Honglei Zhuang, Kai Hui et al. · deepmind

Query expansion has been widely used to improve the search results of first-stage retrievers, yet its influence on second-stage, cross-encoder rankers remains under-explored. A recent work of Weller et al. [44] shows that current expansion techniques benefit weaker models such as DPR and BM25 but harm stronger rankers such as MonoT5. In this paper, we re-examine this conclusion and raise the following question: Can query expansion improve generalization of strong cross-encoder rankers? To answer this question, we first apply popular query expansion methods to state-of-the-art cross-encoder rankers and verify the deteriorated zero-shot performance. We identify two vital steps for cross-encoders in the experiment: high-quality keyword generation and minimal-disruptive query modification. We show that it is possible to improve the generalization of a strong neural ranker, by prompt engineering and aggregating the ranking results of each expanded query via fusion. Specifically, we first call an instruction-following language model to generate keywords through a reasoning chain. Leveraging self-consistency and reciprocal rank weighting, we further combine the ranking results of each expanded query dynamically. Experiments on BEIR and TREC Deep Learning 2019/2020 show that the nDCG@10 scores of both MonoT5 and RankT5 following these steps are improved, which points out a direction for applying query expansion to strong cross-encoder rankers.

Ruicheng Xian, Honglei Zhuang, Zhen Qin et al. · deepmind

Domain adaptation aims to transfer the knowledge learned on (data-rich) source domains to (low-resource) target domains, and a popular method is invariant representation learning, which matches and aligns the data distributions on the feature space. Although this method is studied extensively and applied on classification and regression problems, its adoption on ranking problems is sporadic, and the few existing implementations lack theoretical justifications. This paper revisits invariant representation learning for ranking. Upon reviewing prior work, we found that they implement what we call item-level alignment, which aligns the distributions of the items being ranked from all lists in aggregate but ignores their list structure. However, the list structure should be leveraged, because it is intrinsic to ranking problems where the data and the metrics are defined and computed on lists, not the items by themselves. To close this discrepancy, we propose list-level alignment -- learning domain-invariant representations at the higher level of lists. The benefits are twofold: it leads to the first domain adaptation generalization bound for ranking, in turn providing theoretical support for the proposed method, and it achieves better empirical transfer performance for unsupervised domain adaptation on ranking tasks, including passage reranking.

Kai Hui, Tao Chen, Zhen Qin et al. · deepmind

Retrieval augmentation has shown promising improvements in different tasks. However, whether such augmentation can assist a large language model based re-ranker remains unclear. We investigate how to augment T5-based re-rankers using high-quality information retrieved from two external corpora -- a commercial web search engine and Wikipedia. We empirically demonstrate how retrieval augmentation can substantially improve the effectiveness of T5-based re-rankers for both in-domain and zero-shot out-of-domain re-ranking tasks.

Gheorghe Comanici, Eric Bieber, Mike Schaekermann et al. · amazon-science, baidu

In this report, we introduce the Gemini 2.X model family: Gemini 2.5 Pro and Gemini 2.5 Flash, as well as our earlier Gemini 2.0 Flash and Flash-Lite models. Gemini 2.5 Pro is our most capable model yet, achieving SoTA performance on frontier coding and reasoning benchmarks. In addition to its incredible coding and reasoning skills, Gemini 2.5 Pro is a thinking model that excels at multimodal understanding and it is now able to process up to 3 hours of video content. Its unique combination of long context, multimodal and reasoning capabilities can be combined to unlock new agentic workflows. Gemini 2.5 Flash provides excellent reasoning abilities at a fraction of the compute and latency requirements and Gemini 2.0 Flash and Flash-Lite provide high performance at low latency and cost. Taken together, the Gemini 2.X model generation spans the full Pareto frontier of model capability vs cost, allowing users to explore the boundaries of what is possible with complex agentic problem solving.

Xiaoyang Chen, Kai Hui, Ben He et al.

BERT-based text ranking models have dramatically advanced the state-of-the-art in ad-hoc retrieval, wherein most models tend to consider individual query-document pairs independently. In the mean time, the importance and usefulness to consider the cross-documents interactions and the query-specific characteristics in a ranking model have been repeatedly confirmed, mostly in the context of learning to rank. The BERT-based ranking model, however, has not been able to fully incorporate these two types of ranking context, thereby ignoring the inter-document relationships from the ranking and the differences among queries. To mitigate this gap, in this work, an end-to-end transformer-based ranking model, named Co-BERT, has been proposed to exploit several BERT architectures to calibrate the query-document representations using pseudo relevance feedback before modeling the relevance of a group of documents jointly. Extensive experiments on two standard test collections confirm the effectiveness of the proposed model in improving the performance of text re-ranking over strong fine-tuned BERT-Base baselines. We plan to make our implementation open source to enable further comparisons.

Jinhyuk Lee, Zhuyun Dai, Xiaoqi Ren et al. · uw

We present Gecko, a compact and versatile text embedding model. Gecko achieves strong retrieval performance by leveraging a key idea: distilling knowledge from large language models (LLMs) into a retriever. Our two-step distillation process begins with generating diverse, synthetic paired data using an LLM. Next, we further refine the data quality by retrieving a set of candidate passages for each query, and relabeling the positive and hard negative passages using the same LLM. The effectiveness of our approach is demonstrated by the compactness of the Gecko. On the Massive Text Embedding Benchmark (MTEB), Gecko with 256 embedding dimensions outperforms all existing entries with 768 embedding size. Gecko with 768 embedding dimensions achieves an average score of 66.31, competing with 7x larger models and 5x higher dimensional embeddings.

Hansi Zeng, Kai Hui, Honglei Zhuang et al. · deepmind

While metrics available during pre-training, such as perplexity, correlate well with model performance at scaling-laws studies, their predictive capacities at a fixed model size remain unclear, hindering effective model selection and development. To address this gap, we formulate the task of selecting pre-training checkpoints to maximize downstream fine-tuning performance as a pairwise classification problem: predicting which of two LLMs, differing in their pre-training, will perform better after supervised fine-tuning (SFT). We construct a dataset using 50 1B parameter LLM variants with systematically varied pre-training configurations, e.g., objectives or data, and evaluate them on diverse downstream tasks after SFT. We first conduct a study and demonstrate that the conventional perplexity is a misleading indicator. As such, we introduce novel unsupervised and supervised proxy metrics derived from pre-training that successfully reduce the relative performance prediction error rate by over 50%. Despite the inherent complexity of this task, we demonstrate the practical utility of our proposed proxies in specific scenarios, paving the way for more efficient design of pre-training schemes optimized for various downstream tasks.

Ronak Pradeep, Kai Hui, Jai Gupta et al.

Popularized by the Differentiable Search Index, the emerging paradigm of generative retrieval re-frames the classic information retrieval problem into a sequence-to-sequence modeling task, forgoing external indices and encoding an entire document corpus within a single Transformer. Although many different approaches have been proposed to improve the effectiveness of generative retrieval, they have only been evaluated on document corpora on the order of 100k in size. We conduct the first empirical study of generative retrieval techniques across various corpus scales, ultimately scaling up to the entire MS MARCO passage ranking task with a corpus of 8.8M passages and evaluating model sizes up to 11B parameters. We uncover several findings about scaling generative retrieval to millions of passages; notably, the central importance of using synthetic queries as document representations during indexing, the ineffectiveness of existing proposed architecture modifications when accounting for compute cost, and the limits of naively scaling model parameters with respect to retrieval performance. While we find that generative retrieval is competitive with state-of-the-art dual encoders on small corpora, scaling to millions of passages remains an important and unsolved challenge. We believe these findings will be valuable for the community to clarify the current state of generative retrieval, highlight the unique challenges, and inspire new research directions.

Yi Tay, Vinh Q. Tran, Mostafa Dehghani et al.

In this paper, we demonstrate that information retrieval can be accomplished with a single Transformer, in which all information about the corpus is encoded in the parameters of the model. To this end, we introduce the Differentiable Search Index (DSI), a new paradigm that learns a text-to-text model that maps string queries directly to relevant docids; in other words, a DSI model answers queries directly using only its parameters, dramatically simplifying the whole retrieval process. We study variations in how documents and their identifiers are represented, variations in training procedures, and the interplay between models and corpus sizes. Experiments demonstrate that given appropriate design choices, DSI significantly outperforms strong baselines such as dual encoder models. Moreover, DSI demonstrates strong generalization capabilities, outperforming a BM25 baseline in a zero-shot setup.

Vamsi Aribandi, Yi Tay, Tal Schuster et al.

Despite the recent success of multi-task learning and transfer learning for natural language processing (NLP), few works have systematically studied the effect of scaling up the number of tasks during pre-training. Towards this goal, this paper introduces ExMix (Extreme Mixture): a massive collection of 107 supervised NLP tasks across diverse domains and task-families. Using ExMix, we study the effect of multi-task pre-training at the largest scale to date, and analyze co-training transfer amongst common families of tasks. Through this analysis, we show that manually curating an ideal set of tasks for multi-task pre-training is not straightforward, and that multi-task scaling can vastly improve models on its own. Finally, we propose ExT5: a model pre-trained using a multi-task objective of self-supervised span denoising and supervised ExMix. Via extensive experiments, we show that ExT5 outperforms strong T5 baselines on SuperGLUE, GEM, Rainbow, Closed-Book QA tasks, and several tasks outside of ExMix. ExT5 also significantly improves sample efficiency while pre-training.

Kai Hui, Klaus Berberich

Preference judgments have been demonstrated as a better alternative to graded judgments to assess the relevance of documents relative to queries. Existing work has verified transitivity among preference judgments when collected from trained judges, which reduced the number of judgments dramatically. Moreover, strict preference judgments and weak preference judgments, where the latter additionally allow judges to state that two documents are equally relevant for a given query, are both widely used in literature. However, whether transitivity still holds when collected from crowdsourcing, i.e., whether the two kinds of preference judgments behave similarly remains unclear. In this work, we collect judgments from multiple judges using a crowdsourcing platform and aggregate them to compare the two kinds of preference judgments in terms of transitivity, time consumption, and quality. That is, we look into whether aggregated judgments are transitive, how long it takes judges to make them, and whether judges agree with each other and with judgments from TREC. Our key findings are that only strict preference judgments are transitive. Meanwhile, weak preference judgments behave differently in terms of transitivity, time consumption, as well as of quality of judgment.

Xuanang Chen, Ben He, Kai Hui et al.

Despite the effectiveness of utilizing the BERT model for document ranking, the high computational cost of such approaches limits their uses. To this end, this paper first empirically investigates the effectiveness of two knowledge distillation models on the document ranking task. In addition, on top of the recently proposed TinyBERT model, two simplifications are proposed. Evaluations on two different and widely-used benchmarks demonstrate that Simplified TinyBERT with the proposed simplifications not only boosts TinyBERT, but also significantly outperforms BERT-Base when providing 15$\times$ speedup.

Zhi Zheng, Kai Hui, Ben He et al.

Query expansion aims to mitigate the mismatch between the language used in a query and in a document. However, query expansion methods can suffer from introducing non-relevant information when expanding the query. To bridge this gap, inspired by recent advances in applying contextualized models like BERT to the document retrieval task, this paper proposes a novel query expansion model that leverages the strength of the BERT model to select relevant document chunks for expansion. In evaluation on the standard TREC Robust04 and GOV2 test collections, the proposed BERT-QE model significantly outperforms BERT-Large models.

Sean MacAvaney, Andrew Yates, Arman Cohan et al.

Many questions cannot be answered simply; their answers must include numerous nuanced details and additional context. Complex Answer Retrieval (CAR) is the retrieval of answers to such questions. In their simplest form, these questions are constructed from a topic entity (e.g., `cheese') and a facet (e.g., `health effects'). While topic matching has been thoroughly explored, we observe that some facets use general language that is unlikely to appear verbatim in answers. We call these low-utility facets. In this work, we present an approach to CAR that identifies and addresses low-utility facets. We propose two estimators of facet utility. These include exploiting the hierarchical structure of CAR queries and using facet frequency information from training data. To improve the retrieval performance on low-utility headings, we also include entity similarity scores using knowledge graph embeddings. We apply our approaches to a leading neural ranking technique, and evaluate using the TREC CAR dataset. We find that our approach perform significantly better than the unmodified neural ranker and other leading CAR techniques. We also provide a detailed analysis of our results, and verify that low-utility facets are indeed more difficult to match, and that our approach improves the performance for these difficult queries.

Canjia Li, Yingfei Sun, Ben He et al.

Pseudo-relevance feedback (PRF) is commonly used to boost the performance of traditional information retrieval (IR) models by using top-ranked documents to identify and weight new query terms, thereby reducing the effect of query-document vocabulary mismatches. While neural retrieval models have recently demonstrated strong results for ad-hoc retrieval, combining them with PRF is not straightforward due to incompatibilities between existing PRF approaches and neural architectures. To bridge this gap, we propose an end-to-end neural PRF framework that can be used with existing neural IR models by embedding different neural models as building blocks. Extensive experiments on two standard test collections confirm the effectiveness of the proposed NPRF framework in improving the performance of two state-of-the-art neural IR models.

Sean MacAvaney, Andrew Yates, Arman Cohan et al.

Complex answer retrieval (CAR) is the process of retrieving answers to questions that have multifaceted or nuanced answers. In this work, we present two novel approaches for CAR based on the observation that question facets can vary in utility: from structural (facets that can apply to many similar topics, such as 'History') to topical (facets that are specific to the question's topic, such as the 'Westward expansion' of the United States). We first explore a way to incorporate facet utility into ranking models during query term score combination. We then explore a general approach to reform the structure of ranking models to aid in learning of facet utility in the query-document term matching phase. When we use our techniques with a leading neural ranker on the TREC CAR dataset, our methods rank first in the 2017 TREC CAR benchmark, and yield up to 26% higher performance than the next best method.

Sean MacAvaney, Andrew Yates, Kai Hui et al.

One challenge with neural ranking is the need for a large amount of manually-labeled relevance judgments for training. In contrast with prior work, we examine the use of weak supervision sources for training that yield pseudo query-document pairs that already exhibit relevance (e.g., newswire headline-content pairs and encyclopedic heading-paragraph pairs). We also propose filtering techniques to eliminate training samples that are too far out of domain using two techniques: a heuristic-based approach and novel supervised filter that re-purposes a neural ranker. Using several leading neural ranking architectures and multiple weak supervision datasets, we show that these sources of training pairs are effective on their own (outperforming prior weak supervision techniques), and that filtering can further improve performance.

Kai Hui, Andrew Yates, Klaus Berberich et al.

Neural IR models, such as DRMM and PACRR, have achieved strong results by successfully capturing relevance matching signals. We argue that the context of these matching signals is also important. Intuitively, when extracting, modeling, and combining matching signals, one would like to consider the surrounding text (local context) as well as other signals from the same document that can contribute to the overall relevance score. In this work, we highlight three potential shortcomings caused by not considering context information and propose three neural ingredients to address them: a disambiguation component, cascade k-max pooling, and a shuffling combination layer. Incorporating these components into the PACRR model yields Co-PACRR, a novel context-aware neural IR model. Extensive comparisons with established models on Trec Web Track data confirm that the proposed model can achieve superior search results. In addition, an ablation analysis is conducted to gain insights into the impact of and interactions between different components. We release our code to enable future comparisons.

Andrew Yates, Kai Hui

Recent neural IR models have demonstrated deep learning's utility in ad-hoc information retrieval. However, deep models have a reputation for being black boxes, and the roles of a neural IR model's components may not be obvious at first glance. In this work, we attempt to shed light on the inner workings of a recently proposed neural IR model, namely the PACRR model, by visualizing the output of intermediate layers and by investigating the relationship between intermediate weights and the ultimate relevance score produced. We highlight several insights, hoping that such insights will be generally applicable.

Kai Hui, Andrew Yates, Klaus Berberich et al.

In order to adopt deep learning for information retrieval, models are needed that can capture all relevant information required to assess the relevance of a document to a given user query. While previous works have successfully captured unigram term matches, how to fully employ position-dependent information such as proximity and term dependencies has been insufficiently explored. In this work, we propose a novel neural IR model named PACRR aiming at better modeling position-dependent interactions between a query and a document. Extensive experiments on six years' TREC Web Track data confirm that the proposed model yields better results under multiple benchmarks.