Reward models (RMs) play a pivotal role in aligning large language models (LLMs) with human preferences. However, traditional RM training, which relies on response pairs tied to specific prompts, struggles to disentangle prompt-driven preferences from prompt-independent artifacts, such as response length and format. In this work, we expose a fundamental limitation of current RM training methods, where RMs fail to effectively distinguish between contextual signals and irrelevant artifacts when determining preferences. To address this, we introduce a causal framework that learns preferences independent of these artifacts and propose a novel data augmentation technique designed to eliminate them. Extensive experiments show that our approach successfully filters out undesirable artifacts, yielding a more robust reward model (RRM). Our RRM improves the performance of a pairwise reward model trained on Gemma-2-9b-it, on Reward-Bench, increasing accuracy from 80.61% to 84.15%. Additionally, we train two DPO policies using both the RM and RRM, demonstrating that the RRM significantly enhances DPO-aligned policies, improving MT-Bench scores from 7.27 to 8.31 and length-controlled win-rates in AlpacaEval-2 from 33.46% to 52.49%.

Recent studies have shown that large language models' (LLMs) mathematical problem-solving capabilities can be enhanced by integrating external tools, such as code interpreters, and employing multi-turn Chain-of-Thought (CoT) reasoning. While current methods focus on synthetic data generation and Supervised Fine-Tuning (SFT), this paper studies the complementary direct preference learning approach to further improve model performance. However, existing direct preference learning algorithms are originally designed for the single-turn chat task, and do not fully address the complexities of multi-turn reasoning and external tool integration required for tool-integrated mathematical reasoning tasks. To fill in this gap, we introduce a multi-turn direct preference learning framework, tailored for this context, that leverages feedback from code interpreters and optimizes trajectory-level preferences. This framework includes multi-turn DPO and multi-turn KTO as specific implementations. The effectiveness of our framework is validated through training of various language models using an augmented prompt set from the GSM8K and MATH datasets. Our results demonstrate substantial improvements: a supervised fine-tuned Gemma-1.1-it-7B model's performance increased from 77.5% to 83.9% on GSM8K and from 46.1% to 51.2% on MATH. Similarly, a Gemma-2-it-9B model improved from 84.1% to 86.3% on GSM8K and from 51.0% to 54.5% on MATH.

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.

Large Language Models (LLMs) have revolutionized artificial intelligence with capabilities in reasoning, coding, and communication, driving innovation across industries. Their true potential depends on effective alignment to ensure correct, trustworthy and ethical behavior, addressing challenges like misinformation, hallucinations, bias and misuse. While existing Reinforcement Learning (RL)-based alignment methods are notoriously complex, direct optimization approaches offer a simpler alternative. In this work, we introduce a novel direct optimization approach for LLM alignment by drawing on established Information Retrieval (IR) principles. We present a systematic framework that bridges LLM alignment and IR methodologies, mapping LLM generation and reward models to IR’s retriever-reranker paradigm. Building on this foundation, we propose LLM Alignment as Retriever Preference Optimization (LarPO), a new alignment method that enhances overall alignment quality. Extensive experiments validate LarPO’s effectiveness with 38.9 % and 13.7 % averaged improvement on AlpacaEval2 and MixEval-Hard respectively. Our work opens new avenues for advancing LLM alignment by integrating IR foundations, offering a promising direction for future research.

Tianqi Liu, Zhen Qin, Junru Wu, Jiaming Shen, Misha Khalman, Rishabh Joshi, Yao Zhao, Mohammad Saleh, Simon Baumgartner, Jialu Liu, Peter J Liu, Xuanhui Wang. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

Recent advances in large language models (LLMs) have introduced latent reasoning as a promising alternative to autoregressive reasoning. By performing internal computation with hidden states from previous steps, latent reasoning benefit from more informative features rather than sampling a discrete chain-of-thought (CoT) path. Yet latent reasoning approaches are often incompatible with LLMs, as their continuous paradigm conflicts with the discrete nature of autoregressive generation. Moreover, these methods rely on CoT traces for training and thus fail to exploit the inherent reasoning patterns of LLMs. In this work, we explore latent reasoning by leveraging the intrinsic capabilities of LLMs via reinforcement learning (RL). To this end, we introduce hybrid reasoning policy optimization (HRPO), an RL-based hybrid latent reasoning approach that (1) integrates prior hidden states into sampled tokens with a learnable gating mechanism, and (2) initializes training with predominantly token embeddings while progressively incorporating more hidden features. This design maintains LLMs' generative capabilities and incentivizes hybrid reasoning using both discrete and continuous representations. In addition, the hybrid HRPO introduces stochasticity into latent reasoning via token sampling, thereby enabling RL-based optimization without requiring CoT trajectories. Extensive evaluations across diverse benchmarks show that HRPO outperforms prior methods in both knowledge- and reasoning-intensive tasks. Furthermore, HRPO-trained LLMs remain interpretable and exhibit intriguing behaviors like cross-lingual patterns and shorter completion lengths, highlighting the potential of our RL-based approach and offer insights for future work in latent reasoning.

Large language models (LLMs) are increasingly integral to information retrieval (IR), powering ranking, evaluation, and AI-assisted content creation. This widespread adoption necessitates a critical examination of potential biases arising from the interplay between these LLM-based components. This paper synthesizes existing research and presents novel experiment designs that explore how LLM-based rankers and assistants influence LLM-based judges. We provide the first empirical evidence of LLM judges exhibiting significant bias towards LLM-based rankers. Furthermore, we observe limitations in LLM judges' ability to discern subtle system performance differences. Contrary to some previous findings, our preliminary study does not find evidence of bias against AI-generated content. These results highlight the need for a more holistic view of the LLM-driven information ecosystem. To this end, we offer initial guidelines and a research agenda to ensure the reliable use of LLMs in IR evaluation.

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.

Jing Nathan Yan, Tianqi Liu, Justin Chiu, Jiaming Shen, Zhen Qin, Yue Yu, Charumathi Lakshmanan, Yair Kurzion, Alexander Rush, Jialu Liu, Michael Bendersky. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024.

Yue Yu, Jiaming Shen, Tianqi Liu, Zhen Qin, Jing Nathan Yan, Jialu Liu, Chao Zhang, Michael Bendersky. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024.

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.

Unbiased learning to rank (ULTR) studies the problem of mitigating various biases from implicit user feedback data such as clicks, and has been receiving considerable attention recently. A popular ULTR approach for real-world applications uses a two-tower architecture, where click modeling is factorized into a relevance tower with regular input features, and a bias tower with bias-relevant inputs such as the position of a document. A successful factorization will allow the relevance tower to be exempt from biases. In this work, we identify a critical issue that existing ULTR methods ignored - the bias tower can be confounded with the relevance tower via the underlying true relevance. In particular, the positions were determined by the logging policy, i.e., the previous production model, which would possess relevance information. We give both theoretical analysis and empirical results to show the negative effects on relevance tower due to such a correlation. We then propose two methods to mitigate the negative confounding effects by better disentangling relevance and bias. Offline empirical results on both controlled public datasets and a large-scale industry dataset show the effectiveness of the proposed approaches. We conduct a live experiment on a popular web store for four weeks, and find a significant improvement in user clicks over the baseline, which ignores the negative confounding effect.

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.

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.

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.

Learning-to-Rank (LTR) systems are ubiquitous in web applications nowadays. The existing literature mainly focuses on improving ranking performance by trying to generate the optimal order of candidate items. However, virtually all advanced ranking functions are not scale calibrated. For example, rankers have the freedom to add a constant to all item scores without changing their relative order. This property has resulted in several limitations in deploying advanced ranking methods in practice. On the one hand, it limits the use of effective ranking functions in important applications. For example, in ads ranking, predicted Click-Through Rate (pCTR) is used for ranking and is required to be calibrated for the downstream ads auction. This is a major reason that existing ads ranking methods use scale calibrated pointwise loss functions that may sacrifice ranking performance. On the other hand, popular ranking losses are translation-invariant. We rigorously show that, both theoretically and empirically, this property leads to training instability that may cause severe practical issues.

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.

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.

Two-tower architecture is commonly used in real-world systems for Unbiased Learning to Rank (ULTR), where a Deep Neural Network (DNN) tower models unbiased relevance predictions, while another tower models observation biases inherent in the training data like user clicks. This two-tower architecture introduces inductive biases to allow more efficient use of limited observational logs and better generalization during deployment than single-tower architecture that may learn spurious correlations between relevance predictions and biases. However, despite their popularity, it is largely neglected in the literature that existing two-tower models assume that the joint distribution of relevance prediction and observation probabilities are completely factorizable. In this work, we revisit two-tower models for ULTR. We rigorously show that the factorization assumption can be too strong for real-world user behaviors, and existing methods may easily fail under slightly milder assumptions. We then propose several novel ideas that consider a wider spectrum of user behaviors while still under the two-tower framework to maintain simplicity and generalizability. Our concerns of existing two-tower models and the effectiveness of our proposed methods are validated on both controlled synthetic and large-scale real-world datasets.

Yi Tay, Mostafa Dehghani, Jai Prakash Gupta, Vamsi Aribandi, Dara Bahri, Zhen Qin, Donald Metzler. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 2021.

This paper proposes Omnidirectional Representations from Transformers (OMNINET). In OmniNet, instead of maintaining a strictly horizon-tal receptive field, each token is allowed to attend to all tokens in the entire network. This process can also be interpreted as a form of extreme or intensive attention mechanism that has the receptive field of the entire width and depth of the network. To this end, the omnidirectional attention is learned via a meta-learner, which is essentially another self-attention based model. In order to mitigate the computationally expensive costs of full receptive field attention, we leverage efficient self-attention models such as kernel-based, low-rank attention and/or Big Bird as the meta-learner. Extensive experiments are conducted on autoregressive language modeling(LM1B, C4), Machine Translation, Long Range Arena (LRA), and Image Recognition.The experiments show that OmniNet achieves considerable improvements across these tasks, including achieving state-of-the-art performance on LM1B,WMT’14 En-De/En-Fr, and Long Range Arena.Moreover, using omnidirectional representation in Vision Transformers leads to significant improvements on image recognition tasks on both few-shot learning and fine-tuning setups.

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.

Recent advances in unbiased learning to rank (LTR) count on Inverse Propensity Scoring (IPS) to eliminate bias in implicit feedback. Though theoretically sound in correcting the bias introduced by treating clicked documents as relevant, IPS ignores the bias caused by (implicitly) treating non-clicked ones as irrelevant. In this work, we first rigorously prove that such use of click data leads to unnecessary pairwise comparisons between relevant documents, which prevent unbiased ranker optimization.

Existing work on search result diversification typically falls into the “next document” paradigm, that is, selecting the next document based on the ones already chosen. A sequential process of selecting documents one-by-one is naturally modeled in learning-based approaches. However, such a process makes the learning difficult because there are an exponential number of ranking lists to consider. Sampling is usually used to reduce the computational complexity but this makes the learning less effective. In this paper, we propose a soft version of the “next document” paradigm in which we associate each document with an approximate rank, and thus the subtopics covered prior to a document can also be estimated. We show that we can derive differentiable diversification-aware losses, which are smooth approximation of diversity metrics like α-NDCG, based on these estimates. We further propose to optimize the losses in the learning-to-rank setting using neural distributed representations of queries and documents. Experiments are conducted on the public benchmark TREC datasets. By comparing with an extensive list of baseline methods, we show that our Diversification-Aware LEarning-TO-Rank (DALETOR) approaches outperform them by a large margin, while being much simpler during learning and inference.

A well-known challenge in leveraging implicit user feedback like clicks to improve real-world search services and recommender systems is its inherent bias. Most existing click models are based on the examination hypothesis in user behaviors and differ in how to model such an examination bias. However, they are constrained by assuming a simple position-based bias or enforcing a sequential order in user examination behaviors. These assumptions are insufficient to capture complex real-world user behaviors and hardly generalize to modern user interfaces (UI) in web applications (e.g., results shown in a grid view). In this work, we propose a fully data-driven neural model for the examination bias, Cross-Positional Attention (XPA), which is more flexible in fitting complex user behaviors. Our model leverages the attention mechanism to effectively capture cross-positional interactions among displayed items and is applicable to arbitrary UIs. We employ XPA in a novel neural click model that can both predict clicks and estimate relevance. Our experiments on offline synthetic data sets show that XPA is robust among different click generation processes. We further apply XPA to a large-scale real-world recommender system, showing significantly better results than baselines in online A/B experiments that involve millions of users. This validates the necessity to model more complex user behaviors than those proposed in the literature.

Quick Access is a machine-learned system in Google Drive that predicts which files a user wants to open. Adding Quick Access recommendations to the Drive homepage cut the amount of time that users spend locating their files in half. Aggregated over the ~1 billion users of Drive, the time saved up adds up to ~1000 work weeks every day. In this paper, we discuss both the challenges of iteratively improving the quality of a personal recommendation system as well as the variety of approaches that we took in order to improve this feature. We explored different deep network architectures, novel modeling techniques, additional data sources, and the effects of latency and biases in the UX. We share both pitfalls as well as successes in our attempts to improve this product, and also discuss how we scaled and managed the complexity of the system. We believe that these insights will be especially useful to those who are working with private corpora as well as those who are building a large-scale production recommendation system.

Many modern recommender systems train their models based on a large amount of implicit user feedback data. Due to the inherent bias in this data (e.g., position bias), learning from it directly can lead to suboptimal models. Recently, unbiased learning was proposed to address such problems by leveraging counterfactual techniques like inverse propensity weighting (IPW). In these methods, propensity scores estimation is usually limited to item's display position in a single user interface (UI).

It is often desirable to model multiple objectives in real-world web applications, such as user satisfaction and user engagement in recommender systems. Multi-task learning has become the standard approach for such applications recently.

Retrieval effectiveness in information retrieval systems is heavily dependent on how various parameters are tuned. One option to find these parameters is to run multiple online experiments and using a parameter sweep approach in order to optimize the search system. There are multiple downsides of this approach, mainly that it may lead to a poor experience for users. Another option is to do offline evaluation, which can act as a safeguard against potential quality issues. Offline evaluation requires a validation set of data that can be benchmarked against different parameter settings. However, for search over personal corpora, e.g. email and file search, it is impractical and often impossible to get a complete representative validation set, due to the inability to save raw queries and document information. In this work, we show how to do offline parameter tuning with only a partial validation set. In addition, we demonstrate how to do parameter tuning in the cases when we have complete knowledge of the internal implementation of the search system (white-box tuning), as well as the case where we have only partial knowledge (grey-box tuning). This has allowed us to do offline parameter tuning in a privacy-sensitive manner.

Neural search ranking models have been not only actively studied in the information retrieval community, but also widely adopted in real-world industrial applications. However, due to the non-convexity and stochastic training of neural model formulations, the obtained models are unstable in the sense that model predictions can vary a lot for two models trained with the same configuration. In practice, new features are continuously introduced and new model architectures are explored to improve model effectiveness. In these cases, the instability of neural models leads to unnecessary document ranking changes for a large portion of queries. Such changes not only lead to inconsistent user experience, but also add noise to online experimentation and can slow down model improvement cycles. How to stabilize neural search ranking models during model update is an important but largely unexplored problem. Motivated by trigger analysis, we suggest balancing the trade-off between performance improvement and the number of affected queries. Concretely, we formulate it as an optimization problem with the objective as maximizing the average effect over the affected queries. We propose two heuristics and one theory-guided stabilization method to solve the optimization problem. Our proposed methods are evaluated on two of the world’s largest personal search services: Gmail search and Google Drive search. Empirical results show that our proposed methods are very effective in optimizing the proposed objective and are applicable to different model update scenarios.

Recent neural ranking algorithms focus on learning semantic matching between query and document terms. However, practical learning to rank systems typically rely on a wide range of side information beyond query and document textual features, like location, user context, etc. It is common practice to concatenate all of these features and rely on deep models to learn a complex representation.

Decision Trees (DTs) like LambdaMART have been one of the most effective types of learning-to-rank algorithms in the past decade. They typically work well with hand-crafted dense features (e.g., BM25 scores). Recently, Neural Networks (NNs) have shown impressive results in leveraging sparse and complex features (e.g., query and document keywords) directly when a large amount of training data is available. While there is a large body of work on how to use NNs for semantic matching between queries and documents, relatively less work has been conducted to compare NNs with DTs for general learning-to-rank tasks, where dense features are also available and DTs can achieve state-of-the-art performance. In this paper, we study how to combine DTs and NNs to effectively bring the benefits from both sides in the learning-to-rank setting. Specifically, we focus our study on personal search where clicks are used as the primary labels with unbiased learning-to-rank algorithms and a significantly large amount of training data is easily available. Our combination methods are based on ensemble learning. We design 12 variants and compare them based on two aspects, ranking effectiveness and ease-of-deployment, using two of the largest personal search services: Gmail search and Google Drive search. We show that direct application of existing ensemble methods can not achieve both aspects. We thus design a novel method that uses NNs to compensate DTs via boosting. We show that such a method is not only easier to deploy, but also gives comparable or better ranking accuracy.

Spell correction is a must-have feature for any modern search engine in applications such as web or e-commerce search. Typical spell correction solutions used in production systems consist of large indexed lookup tables based on a global model trained across many users over a large scale web corpus or a query log.

A multivariate Hawkes process is a class of marked point processes: A sample consists of a finite set of events of unbounded random size; each event has a real-valued time and a discrete-valued label (mark). It is self-excitatory: Each event causes an increase in the rate of other events (of either the same or a different label) in the (near) future. Prior work has developed methods for parameter estimation from complete samples. However, just as unobserved variables can increase the modeling power of other probabilistic models, allowing unobserved events can increase the modeling power of point processes. In this paper we develop a method to sample over the posterior distribution of unobserved events in a multivariate Hawkes process. We demonstrate the efficacy of our approach, and its utility in improving predictive power and identifying latent structure in real-world data.

User information needs vary significantly across different tasks, and therefore their queries will also differ considerably in their expressiveness and semantics. Many studies have been proposed to model such query diversity by obtaining query types and building query-dependent ranking models. These studies typically require either a labeled query dataset or clicks from multiple users aggregated over the same document. These techniques, however, are not applicable when manual query labeling is not viable, and aggregated clicks are unavailable due to the private nature of the document collection, e.g., in email search scenarios. In this paper, we study how to obtain query type in an unsupervised fashion and how to incorporate this information into query-dependent ranking models. We first develop a hierarchical clustering algorithm based on truncated SVD and varimax rotation to obtain coarse-to-fine query types. Then, we study three query-dependent ranking models, including two neural models that leverage query type information as additional features, and one novel multi-task neural model that views query type as the label for the auxiliary query cluster prediction task. This multi-task model is trained to simultaneously rank documents and predict query types. Our experiments on tens of millions of real-world email search queries demonstrate that the proposed multi-task model can significantly outperform the baseline neural ranking models, which either do not incorporate query type information or just simply feed query type as an additional feature.

Matching people in different cameras, commonly referred to as person re-identification, is a challenging task. The challenges come from the drastic appearance variation across different camera views caused by changes in pose, lighting condition, occlusion, background, and so on. Instead of matching images in the original feature space, many existing methods learn distance metrics or feature transformations to improve the matching accuracy. For example, data from different camera views can be projected onto their common space to mitigate the feature gap caused by view discrepancy. However, a single-step mapping may not be sufficient to close this gap as the view difference can be significant between different cameras. To overcome this limitation, we propose a multi-level common space learning framework, which can gradually minimize the data discrepancy between different views in an iterative manner. At each intermediate level, synthetic data are generated using the automatically discovered grouping information, and these synthetic data can be viewed as a transitional state between the original camera views. The matching is performed by mapping the probe and gallery onto their common space in multiple steps. We evaluate the proposed method on two widely used person re-identification data sets. The results show that the proposed multi-level scheme significantly outperforms single-level mapping. In addition, competitive results are achieved as compared with over 20 state-of-the-art techniques.

We propose a novel approach toward event detection in real-world continuous video sequences. The method: 1) is able to model arbitrary-order non-Markovian dependences in videos to mitigate local visual ambiguities; 2) conducts simultaneous event segmentation and labeling; and 3) is time-window free. The idea is to represent a video as an event stream of both high-level semantic events and low-level video observations. In training, we learn a point process model called a piecewise-constant conditional intensity model (PCIM) that is able to capture complex non-Markovian dependences in the event streams. In testing, event detection can be modeled as the inference of high-level semantic events, given low-level image observations. We develop the first inference algorithm for PCIM and show it samples exactly from the posterior distribution. We then evaluate the video event detection task on real-world video sequences. Our model not only provides competitive results on the video event segmentation and labeling task, but also provides benefits, including being interpretable and efficient.

Many computer vision tasks are more difficult when tackled without contextual information. For example, in multi-camera tracking, pedestrians may look very different in different cameras with varying pose and lighting conditions. Similarly, head direction estimation in high-angle surveillance video in which human head images are low resolution is challenging. Even humans can have trouble without contextual information. In this work, we couple novel contextual information, social grouping, with two important computer vision tasks: multi-target tracking and head pose/direction estimation in surveillance video. These three components are modeled in a probabilistic formulation and we provide effective solvers.We show that social grouping effectively helps to mitigate visual ambiguities in multi-camera tracking and head pose estimation. We further notice that in single-camera multi-target tracking, social grouping provides a natural high-order association cue that avoids existing complex algorithms for high-order track association. In experiments, we demonstrate improvements with our model over models without social grouping context and several state-of-art approaches on a number of publicly available datasets on tracking, head pose estimation, and group discovery.

An online approach to learn elementary groups containing only two targets, i.e., pedestrians, for inferring high-level context is introduced to improve multiperson tracking. In most existing data association-based tracking approaches, only low-level information (e.g., time, appearance, and motion) is used to build the affinity model, and each target is considered as an independent agent. Unlike those previous methods, in this paper, an online learned social grouping behavior model is used to provide more robust tracklet affinities. A disjoint grouping graph is used to encode social grouping behavior of pairwise targets, where each node represents an elementary group of two targets, and two nodes are connected if they share a common target. Probabilities of the uncertain target in two connected nodes being the same person are inferred from each edge of the grouping graph. Relationships between elementary groups are discovered by group tracking, and a nonlinear motion map is used for explaining nonlinear motion pattern between elementary groups. The proposed method is efficient, able to handle group split and merge, and can be easily integrated into any basic affinity model. The approach is evaluated on four data sets, and it shows significant improvements compared with state-of-the-art methods.

Image annotation is a fundamental and challenging task in the field of semantic image retrieval. In this paper, we deal with image annotation via matrix completion. Concretely, we formulate the problem of annotating the tags of an image into a constrained optimization problem, in which the constraint is to keep the consistency with the given initial labels and the objective is to minimize the discrepancy between the correlation in visual content and the correlation in semantic tags. We solve the optimization problem with the linearized alternating direction method. Experimental results on benchmark data demonstrate the effectiveness of our proposals.

We introduce an online approach to learn possible elementary groups (groups that contain only two targets) for inferring high level context that can be used to improve multi-target tracking in a data-association based framework. Unlike most existing association-based tracking approaches that use only low level information (e.g., time, appearance, and motion) to build the affinity model and consider each target as an independent agent, we online learn social grouping behavior to provide additional information for producing more robust tracklets affinities. Social grouping behavior of pairwise targets is first learned from confident tracklets and encoded in a disjoint grouping graph. The grouping graph is further completed with the help of group tracking. The proposed method is efficient, handles group merge and split, and can be easily integrated into any basic affinity model. We evaluate our approach on two public datasets, and show significant improvements compared with state-of-the-art methods.

Extracting product aspects and their associated sentiments is one of the key tasks in sentiment analysis. Estimating the confidences of extracted aspects is important to ensure the performance. To tackle the issue, this paper proposes a two-step estimation method. Collocations of product features and opinion words are initially extracted through pattern bootstrapping. A criterion synthesizing two measurements, Popularity and Reliability, is novelly exploited to assess both patterns and features. Then the features are further clustered into aspects based on path similarities in the Word Net. Each cluster is assigned a weight based on its Compactness and Texture, and the light ones are filtered out. In addition, this paper also captures global aspect reputations by aggregating sentiment strengths through opinion collocations. Experimental results on a benchmark data set with 5 products demonstrate the effectiveness and reliability of our proposed method.

In this paper, we enhance the Shape Context-based descriptor, shapemes, by introducing an ordered bag-of-features model and dynamic programming. The proposed descriptor consists of a series of sub-histograms of shapemes, each of which represents a subset of sampled points. The division of the sampled points is based on their sequential positions on the contour of the shape, so the representation has intrinsic order and is therefore named ordered histogram of shapemes. Then dynamic programming is utilized for descriptor matching. The framework is effective and efficient owing to the following properties: 1) points division approach together with dynamic programming for invariance under the change of starting point, 2) Earth Mover's Distance for discriminative power, and 3) pre-caculated shapemes dissimilarity matrix for fast descriptor distance calculation. Experiments on standard shape database and real world application scenario demonstrate the effectiveness and efficiency of the descriptor and the matching framework. We make our code and experimental data publicly available for future reference.

Reference-based image classification approach introduces a reference-set for both image representation and dictionary learning. It significantly reduces the dimensionality of represented images and shows outstanding performance even with randomly selected reference images and simple distance measure. In this paper, we improve upon existing work with two major contributions. First, we show that a more representative reference-set contributes to better classification accuracy. To this end, we carefully adapt the K-means clustering algorithm in the feature space to select a distinguished reference-set. Second, in the image classification process, we propose to represent each image by measuring its betweenness centrality in a social network composed of the representative reference-set in each class, leading to a more coherent distance measure that considers the overall connectivity between the probe image and the reference-set. Extensive experiment results demonstrate that our proposed scheme achieves better performance than existing methods.

This paper uses a social grouping model to improve target handover across multiple non-overlapping cameras to enable wide-area video understanding. Prior work focuses on modeling appearance and spatial-temporal cues for target handover. In cameras with different conditions, these cues are weak, at best. We provide a complete generative social grouping model which generalizes a recent single-camera case. Our extension requires strengthening the probabilistic interpretations and the resulting optimization over track handovers and social groupings can be formulated in terms of standard fast algorithms. We demonstrate the effectiveness of the method over existing techniques on challenging real-world multi-camera video.

We address the problem of multi-person data-association-based tracking (DAT) in semi-crowded environments from a single camera. Existing tracklet-association-based methods using purely visual cues (like appearance and motion information) show impressive results but rely on heavy training, a number of tuned parameters, and sophisticated detectors to cope with visual ambiguities within the video and low-level processing errors. In this work, we consider clustering dynamics to mitigate such ambiguities. This leads to a general optimization framework that adds social grouping behavior (SGB) to any basic affinity model. We formulate this as a nonlinear global optimization problem to maximize the consistency of visual and grouping cues for trajectories in both tracklet-tracklet linking space and tracklet-grouping assignment space. We formulate the Lagrange dual and solve it using a two-stage iterative algorithm, employing the Hungarian algorithm and K-means clustering. We build SGB upon a simple affinity model and show very promising performance on two publicly available real-world datasets with different tracklet extraction methods.

With the prevalence of E-Commerce sites such as eBay, Content-based Commercial Product Image Retrieval (CBCPIR) has become an emerging application-oriented field of Content-based Image Retrieval (CBIR). Though a number of traditional CBIR techniques and evaluation criterions have been applied directly or with minor modifications, they tend to neglect one critical factor that greatly affects user experience: users usually care about the exact ranks of the results, especially few top ones, which should share very high similarity with the query image. In this work, we propose a novel two-stage retrieval framework that uses a compression-based re-ranking method and a new subjective retrieval evaluation criterion to address such a problem. More specifically, we extend the state-of-art texture descriptor Campana-Keogh (CK) method from data mining in several aspects and validate the superiority of our framework via extensive experiments and real-world user feedback. We also make our code and CBCPIR dataset publicly available. The number of images of the latter is much larger than current freely accessible ones and better represents real-world commercial product images.