The scaling of inference computation has unlocked the potential of long-context large language models (LLMs) across diverse settings. For knowledge-intensive tasks, the increased compute is often allocated to incorporate more external knowledge. However, without effectively utilizing such knowledge, solely expanding context does not always enhance performance. In this work, we investigate inference scaling for retrieval augmented generation (RAG), exploring the combination of multiple strategies beyond simply increasing the quantity of knowledge, including in-context learning and iterative prompting. These strategies provide additional flexibility to scale test-time computation (e.g., by increasing retrieved documents or generation steps), thereby enhancing LLMs’ ability to effectively acquire and utilize contextual information. We address two key questions: (1) How does RAG performance benefit from the scaling of inference computation when optimally configured? (2) Can we predict the optimal test-time compute allocation for a given budget by modeling the relationship between RAG performance and inference parameters? Our observations reveal that increasing inference computation leads to nearly linear gains in RAG performance when optimally allocated, a relationship we describe as the inference scaling laws for RAG. Building on this, we further develop the computation allocation model to estimate RAG performance across different inference configurations. The model predicts optimal inference parameters under various computation constraints, which align closely with the experimental results. By applying these optimal configurations, we demonstrate that scaling inference compute on long-context LLMs achieves up to 58.9% gains on benchmark datasets compared to standard RAG.

Modern retrieval systems do not rely on a single ranking model to construct their rankings. Instead, they generally take a cascading approach where a sequence of ranking models are applied in multiple re-ranking stages. Thereby, they balance the quality of the top-K ranking with computational costs by limiting the number of documents each model re-ranks. However, the cascading approach is not the only way models can interact to form a retrieval system.

The powerful generative abilities of large language models (LLMs) show potential in generating relevance labels for search applications.Previous work has found that directly asking about relevancy, such as "How relevant is document A to query Q?", results in sub-optimal ranking.Instead, the pairwise-ranking prompting (PRP) approach produces promising ranking performance through asking about pairwise comparisons, e.g., "Is document A more relevant than document B to query Q?".Thus, while LLMs are effective at their ranking ability, this is not reflected in their relevance label generation.In this work, we propose a post-processing method to consolidate the relevance labels generated by an LLM with its powerful ranking abilities.Our method takes both LLM generated relevance labels and pairwise preferences.The labels are then altered to satisfy the pairwise preferences of the LLM, while staying as close to the original values as possible.Our experimental results indicate that our approach effectively balances label accuracy and ranking performance.Thereby, our work shows it is possible to combine both the ranking and labeling abilities of LLMs through post-processing.

The traditional evaluation of information retrieval (IR) systems is generally very costly as it requires manual relevance annotation from human experts. Recent advancements in generative artificial intelligence -specifically large language models (LLMs)- can generate relevance annotations at an enormous scale with relatively small computational costs. Potentially, this could alleviate the costs traditionally associated with IR evaluation and make it applicable to numerous low-resource applications. However, generated relevance annotations are not immune to (systematic) errors, and as a result, directly using them for evaluation produces unreliable results.

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 study shows that current expansion techniques benefit weaker models but harm stronger rankers. 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 different cross-encoder rankers and verify the deteriorated zero-shot effectiveness. We identify two vital steps in the experiment: high-quality keyword generation and minimally-disruptive query modification. We show that it is possible to improve the generalization of a strong neural ranker, by generating keywords through a reasoning chain and aggregating the ranking results of each expanded query via self-consistency, reciprocal rank weighting, and fusion. 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.

As Learning-to-Rank (LTR) approaches primarily seek to improve ranking quality, their output scores are not scale-calibrated by design. This fundamentally limits LTR usage in score-sensitive applications. Though a simple multi-objective approach that combines a regression and a ranking objective can effectively learn scale-calibrated scores, we argue that the two objectives are not necessarily compatible, which makes the trade-off less ideal for either of them. In this paper, we propose a practical regression compatible ranking (RCR) approach that achieves a better trade-off, where the two ranking and regression components are proved to be mutually aligned. Although the same idea applies to ranking with both binary and graded relevance, we mainly focus on binary labels in this paper. We evaluate the proposed approach on several public LTR benchmarks and show that it consistently achieves either best or competitive result in terms of both regression and ranking metrics, and significantly improves the Pareto frontiers in the context of multi-objective optimization. Furthermore, we evaluated the proposed approach on YouTube Search and found that it not only improved the ranking quality of the production pCTR model, but also brought gains to the click prediction accuracy. The proposed approach has been successfully deployed in the YouTube production system.

The distillation of ranking models has become an important topic in both academia and industry. In recent years, several advanced methods have been proposed to tackle this problem, often leveraging ranking information from teacher rankers that is absent in traditional classification settings. To date, there is no well-established consensus on how to evaluate this class of models. Moreover, inconsistent benchmarking on a wide range of tasks and datasets make it difficult to assess or invigorate advances in this field. This paper first examines representative prior arts on ranking distillation, and raises three questions to be answered around methodology and reproducibility. To that end, we propose a systematic and unified benchmark, Ranking Distillation Suite (RD-Suite), which is a suite of tasks with 4 large real-world datasets, encompassing two major modalities (textual and numeric) and two applications (standard distillation and distillation transfer). RD-Suite consists of benchmark results that challenge some of the common wisdom in the field, and the release of datasets with teacher scores and evaluation scripts for future research. RD-Suite paves the way towards better understanding of ranking distillation, facilities more research in this direction, and presents new challenges.

Pretrained language models such as BERT have been shown to be exceptionally effective for text ranking. 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 a classification problem 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 performance. Our experiments show that the proposed models with ranking losses can achieve substantial ranking performance gains on different public text ranking data sets. Moreover, ranking models fine-tuned with listwise ranking losses have better zero-shot ranking performance on out-of-domain data than models fine-tuned with classification losses.

Rax is a library for composable Learning-to-Rank (LTR) written entirely in JAX. The goal of Rax is to facilitate easy prototyping of LTR systems by leveraging the flexibility and simplicity of JAX. Rax provides a diverse set of popular ranking metrics and losses that integrate well with the rest of the JAX ecosystem. Furthermore, Rax implements a system of ranking-specific function transformations which allows fine-grained customization of ranking losses and metrics. Most notably Rax provides approx_t12n: a function transformation (t12n) that can transform any of our ranking metrics into an approximate and differentiable form that can be optimized. This provides a systematic way to directly optimize neural ranking models for ranking metrics that are not easily optimizable in other libraries. We empirically demonstrate the effectiveness of Rax by benchmarking neural models implemented using Flax and trained using Rax on two popular LTR benchmarks: WEB30K and Istella. Furthermore, we show that integrating ranking losses with T5, a large language model, can improve overall ranking performance on the MS MARCO passage ranking task. We are sharing the Rax library with the open source community as part of the larger JAX ecosystem at https://github.com/google/rax.

Top-K metrics such as [email protected] are frequently used to evaluate ranking performance. The traditional tree-based models such as LambdaMART, which are based on Gradient Boosted Decision Trees (GBDT), are designed to optimize [email protected] using the LambdaRank losses. Recently, there is a good amount of research interest on neural ranking models for learning-to-rank tasks. These models are fundamentally different from the decision tree models and behave differently with respect to different loss functions. For example, the most popular ranking losses used in neural models are the Softmax loss and the GumbelApproxNDCG loss. These losses do not connect to top-K metrics such as [email protected] naturally. It remains a question on how to effectively optimize [email protected] for neural ranking models. In this paper, we follow the LambdaLoss framework and design novel and theoretically sound losses for [email protected] metrics, while the original LambdaLoss paper can only do so using an unsound heuristic. We study the new losses on the LETOR benchmark datasets and show that the new losses work better than other losses for neural ranking models.

Google Chrome, one of the world's most popular web browsers, features an extension framework allowing third-party developers to enhance Chrome's functionality. Chrome extensions are distributed through the Chrome Web Store (CWS), a Google-operated online marketplace. In this paper, we describe how we developed and deployed three recommender systems for discovering relevant extensions in CWS, namely non-personalized recommendations, related extension recommendations, and personalized recommendations. Unlike most existing papers that focus on novel algorithms, this paper focuses on sharing practical experiences when building large-scale recommender systems under various real-world constraints, such as privacy constraints, data sparsity and skewness issues, and product design choices (e.g., user interface). We show how these constraints make standard approaches difficult to succeed in practice. We share success stories that turn negative live metrics to positive ones, including: 1) how we use interpretable neural models to bootstrap the systems, help identifying pipeline issues, and pave the way for more advanced models; 2) a new item-item based algorithm for related recommendations that works under highly skewed data distributions; and 3) how the previous two techniques can help bootstrapping the personalized recommendations, which significantly reduces development cycles and bypasses various real-world difficulties. All the explorations in this work are verified in live traffic on millions of users. We believe that the findings in this paper can help practitioners to build better large-scale recommender systems.

Cloud-based file storage platforms such as Google Drive are widely used as a means for storing, editing and sharing personal and organizational documents. In this paper, we improve search ranking quality for cloud storage platforms by utilizing user activity logs. Different from search logs, activity logs capture general document usage activity beyond search, such as opening, editing and sharing documents. We propose to automatically learn text embeddings that are effective for search ranking from activity logs. We develop a novel co-access signal, i.e., whether two documents were accessed by a user around the same time, to train deep semantic matching models that are useful for improving the search ranking quality. We confirm that activity-trained semantic matching models can improve ranking by conducting extensive offline experimentation using Google Drive search and activity logs. To the best of our knowledge, this is the first work to examine the benefits of leveraging document usage activity at large scale for cloud storage search; as such it can shed light on using such activity in scenarios where direct collection of search-specific interactions (e.g., query and click logs) may be expensive or infeasible.

Counterfactual Learning To Rank (LTR) algorithms learn a ranking model from logged user interactions, often collected using a production system. Employing such an offline learning approach has many benefits compared to an online one, but it is challenging as user feedback often contains high levels of bias. Unbiased LTR uses Inverse Propensity Scoring (IPS) to enable unbiased learning from logged user interactions. One of the major difficulties in applying Stochastic Gradient Descent (SGD) approaches to counterfactual learning problems is the large variance introduced by the propensity weights. In this paper we show that the convergence rate of SGD approaches with IPS-weighted gradients suffers from the large variance introduced by the IPS weights: convergence is slow, especially when there are large IPS weights.

Contextual bandit problems are a natural fit for many information retrieval tasks, such as learning to rank, text classification, recommendation, and so on. However, existing learning methods for contextual bandit problems have one of two drawbacks: They either do not explore the space of all possible document rankings (i.e., actions) and, thus, may miss the optimal ranking, or they present suboptimal rankings to a user and, thus, may harm the user experience. We introduce a new learning method for contextual bandit problems, Safe Exploration Algorithm (SEA), which overcomes the above drawbacks. SEA starts by using a baseline (or production) ranking system (i.e., policy), which does not harm the user experience and, thus, is safe to execute but has suboptimal performance and, thus, needs to be improved. Then SEA uses counterfactual learning to learn a new policy based on the behavior of the baseline policy. SEA also uses high-confidence off-policy evaluation to estimate the performance of the newly learned policy. Once the performance of the newly learned policy is at least as good as the performance of the baseline policy, SEA starts using the new policy to execute new actions, allowing it to actively explore favorable regions of the action space. This way, SEA never performs worse than the baseline policy and, thus, does not harm the user experience, while still exploring the action space and, thus, being able to find an optimal policy. Our experiments using text classification and document retrieval confirm the above by comparing SEA (and a boundless variant called BSEA) to online and offline learning methods for contextual bandit problems.

Learning to Rank (LTR) from user interactions is challenging as user feedback often contains high levels of bias and noise. At the moment, two methodologies for dealing with bias prevail in the field of LTR: counterfactual methods that learn from historical data and model user behavior to deal with biases; and online methods that perform interventions to deal with bias but use no explicit user models. For practitioners the decision between either methodology is very important because of its direct impact on end users. Nevertheless, there has never been a direct comparison between these two approaches to unbiased LTR. In this study we provide the first benchmarking of both counterfactual and online LTR methods under different experimental conditions. Our results show that the choice between the methodologies is consequential and depends on the presence of selection bias, and the degree of position bias and interaction noise. In settings with little bias or noise counterfactual methods can obtain the highest ranking performance; however, in other circumstances their optimization can be detrimental to the user experience. Conversely, online methods are very robust to bias and noise but require control over the displayed rankings. Our findings confirm and contradict existing expectations on the impact of model-based and intervention-based methods in LTR, and allow practitioners to make an informed decision between the two methodologies.

This tutorial aims to weave together diverse strands of modern Learning to Rank (LtR) research, and present them in a unified full-day tutorial. First, we will introduce the fundamentals of LtR, and an overview of its various sub-fields. Then, we will discuss some recent advances in gradient boosting methods such as LambdaMART by focusing on their efficiency/effectiveness trade-offs and optimizations. Subsequently, we will then present TF-Ranking, a new open source TensorFlow package for neural LtR models, and how it can be used for modeling sparse textual features. Finally, we will conclude the tutorial by covering unbiased LtR -- a new research field aiming at learning from biased implicit user feedback. The tutorial will consist of three two-hour sessions, each focusing on one of the topics described above. It will provide a mix of theoretical and hands-on sessions, and should benefit both academics interested in learning more about the current state-of-the-art in LtR, as well as practitioners who want to use LtR techniques in their applications.

We consider the novel problem of evaluating a recommendation policy offline in environments where the reward signal is non-stationary. Non-stationarity appears in many Information Retrieval (IR) applications such as recommendation and advertising, but its effect on off-policy evaluation has not been studied at all. We are the first to address this issue. First, we analyze standard off-policy estimators in non-stationary environments and show both theoretically and experimentally that their bias grows with time. Then, we propose new off-policy estimators with moving averages and show that their bias is independent of time and can be bounded. Furthermore, we provide a method to trade-off bias and variance in a principled way to get an off-policy estimator that works well in both non-stationary and stationary environments. We experiment on publicly available recommendation datasets and show that our newly proposed moving average estimators accurately capture changes in non-stationary environments, while standard off-policy estimators fail to do so.

Topic models such as Latent Dirichlet Allocation (LDA) have been widely used in information retrieval for tasks ranging from smoothing and feedback methods to tools for exploratory search and discovery. However, classical methods for inferring topic models do not scale up to the massive size of today's publicly available Web-scale data sets. The state-of-the-art approaches rely on custom strategies, implementations and hardware to facilitate their asynchronous, communication-intensive workloads. We present APS-LDA, which integrates state-of-the-art topic modeling with cluster computing frameworks such as Spark using a novel asynchronous parameter server. Advantages of this integration include convenient usage of existing data processing pipelines and eliminating the need for disk writes as data can be kept in memory from start to finish. Our goal is not to outperform highly customized implementations, but to propose a general high-performance topic modeling framework that can easily be used in today's data processing pipelines. We compare APS-LDA to the existing Spark LDA implementations and show that our system can, on a 480-core cluster, process up to 135× more data and 10× more topics without sacricing model quality.