Large Language Models (LLMs) have demonstrated exceptional proficiency in text understanding and embedding tasks. However, their potential in multimodal representation, particularly for item-to-item (I2I) recommendations, remains underexplored. While leveraging existing Multimodal Large Language Models (MLLMs) for such tasks is promising, challenges arise due to their delayed release compared to corresponding LLMs and the inefficiency in representation tasks. To address these issues, we propose an end-to-end fine-tuning method that customizes the integration of any existing LLMs and vision encoders for efficient multimodal representation. Preliminary experiments revealed that fine-tuned LLMs often neglect image content. To counteract this, we propose NoteLLM-2, a novel framework that enhances visual information. Specifically, we propose two approaches: first, a prompt-based method that segregates visual and textual content, employing a multimodal In-Context Learning strategy to balance focus across modalities; second, a late fusion technique that directly integrates visual information into the final representations. Extensive experiments, both online and offline, demonstrate the effectiveness of our approach. Code is available at https://github.com/Applied-Machine-Learning-Lab/NoteLLM.
论文检索
输入标题、作者或关键词，从 844 篇学术成果中精准定位
As GPUs are widely deployed in High Performance Computing systems, it is critical to ensure that these systems can perform reliably. To improve system reliability, researchers estimate the error resilience of GPU programs by understanding resilience characteristics or modeling error propagation. However, features indicative of resilience rely on manual extraction from simulations of numerous faults, and error propagation analysis cannot target fine-grained bit-level faults. To address those problems, this paper introduces a novel paradigm, namely InstrDGM, for efficiently predicting GPU error resilience. Specifically, InstrDGM first fine-tunes a large language model using extensive sequences of GPU assembly instructions for extracting the semantic representation of instructions automatically. Meanwhile, we consider the propagation of bit-level faults during instruction execution and data transfer processes, and leverage graph neural networks to capture their distinct error propagation patterns. Then, the fault embeddings extracted from these error propagation patterns are integrated for error resilience prediction. Additionally, this paper releases a new dataset for GPU error resilience assessment, containing 1.2 million fault samples. Finally, extensive experiments show that InstrDGM significantly outperforms existing methods.
Over the past few years, the emergence of backdoor attacks has presented significant challenges to deep learning systems, allowing attackers to insert backdoors into neural networks. When data with a trigger is processed by a backdoor model, it can lead to mispredictions targeted by attackers, whereas normal data yields regular results. The scope of backdoor attacks is expanding beyond computer vision and encroaching into areas such as natural language processing and speech recognition. Nevertheless, existing backdoor defense methods are typically tailored to specific data modalities, restricting their application in multimodal contexts. While multimodal learning proves highly applicable in facial recognition, sentiment analysis, action recognition, visual question answering, the security of these models remains a crucial concern. Specifically, there are no existing backdoor benchmarks targeting multimodal applications or related tasks. In order to facilitate the research in multimodal backdoor, we introduce BackdoorMBTI, the first backdoor learning toolkit and benchmark designed for multimodal evaluation across three representative modalities from eleven commonly used datasets. BackdoorMBTI provides a systematic backdoor learning pipeline, encompassing data processing, data poisoning, backdoor training, and evaluation. The generated poison datasets and backdoor models enable detailed evaluation of backdoor defenses. Given the diversity of modalities, BackdoorMBTI facilitates systematic evaluation across different data types. Furthermore, BackdoorMBTI offers a standardized approach to handling practical factors in backdoor learning, such as issues related to data quality and erroneous labels. We anticipate that BackdoorMBTI will expedite future research in backdoor defense methods within a multimodal context. Code is available at https://github.com/SJTUHaiyangYu/BackdoorMBTI.
Sepsis is an organ dysfunction caused by a deregulated immune response to an infection. Early sepsis prediction and identification allow for timely intervention, leading to improved clinical outcomes. Clinical calculators (e.g., the six-organ dysfunction assessment of SOFA) play a vital role in sepsis identification within clinicians' workflow, providing evidence-based risk assessments essential for sepsis diagnosis. However, artificial intelligence (AI) sepsis prediction models typically generate a single sepsis risk score without incorporating clinical calculators for assessing organ dysfunctions, making the models less convincing and transparent to clinicians. To bridge the gap, we propose to mimic clinicians' workflow with a novel framework SepsisCalc to integrate clinical calculators into the predictive model, yielding a clinically transparent and precise model for utilization in clinical settings. Practically, clinical calculators usually combine information from multiple component variables in Electronic Health Records (EHR), and might not be applicable when the variables are (partially) missing. We mitigate this issue by representing EHRs as temporal graphs and integrating a learning module to dynamically add the accurately estimated calculator to the graphs. Experimental results on real-world datasets show that the proposed model outperforms state-of-the-art methods on sepsis prediction tasks. Moreover, we developed a system to identify organ dysfunctions and potential sepsis risks, providing a human-AI interaction tool for deployment, which can help clinicians understand the prediction outputs and prepare timely interventions for the corresponding dysfunctions, paving the way for actionable clinical decision-making support for early intervention.
The significance of estimating video watch time has been highlighted by the rising importance of (short) video recommendation, which has become a core product of mainstream social media platforms. Modeling video watch time, however, has been challenged by the complexity of user-video interaction, such as different user behavior modes in watching the recommended videos and varying watching probability over the video progress bar. Despite the importance and challenges, existing literature on modeling video watch time mostly focuses on relatively black-box mechanical enhancement of the classical regression/classification losses, without factoring in user behavior in a principled manner. In this paper, we for the first time take on a user-centric perspective to model video watch time, from which we propose a white-box statistical framework that directly translates various user behavior assumptions in watching (short) videos into statistical watch time models. These behavior assumptions are portrayed by our domain knowledge on users' behavior modes in video watching. We further employ bucketization to cope with user's non-stationary watching probability over the video progress bar, which additionally helps to respect the constraint of video length and facilitate the practical compatibility between the continuous regression event of watch time and other binary classification events. We test our models extensively on two public datasets, a large-scale offline industrial dataset, and an online A/B test on a short video platform with hundreds of millions of daily-active users. On all experiments, our models perform competitively against strong relevant baselines, demonstrating the efficacy of our user-centric perspective and proposed framework.
Textual description of a physical location, commonly known as an address, plays an important role in location-based services(LBS) such as on-demand delivery and navigation. However, the prevalence of abnormal addresses, those containing inaccuracies that fail to pinpoint a location, have led to significant costs. Address rewriting has emerged as a solution to rectify these abnormal addresses. Despite the critical need, existing address rewriting methods are limited, typically tailored to correct specific error types, or frequently require retraining to process new address data effectively. In this study, we introduce AddrLLM, an innovative framework for address rewriting that is built upon a retrieval augmented large language model. AddrLLM overcomes aforementioned limitations through a meticulously designed Supervised Fine-Tuning module, an Address-centric Retrieval Augmented Generation module and a Bias-free Objective Alignment module. To the best of our knowledge, this study pioneers the application of LLM-based address rewriting approach to solve the issue of abnormal addresses. Through comprehensive offline testing with real-world data on a national scale and subsequent online deployment, AddrLLM has demonstrated superior performance in integration with existing logistics system. It has significantly decreased the rate of parcel re-routing by approximately 43%, underscoring its exceptional efficacy in real-world applications.
Click-through Rate (CTR) prediction is crucial for online personalization platforms. Recent advancements have shown that modeling rich user behaviors can significantly improve the performance of CTR prediction. Current long-term user behavior modeling algorithms predominantly follow two cascading stages. The first stage retrieves subsequence related to the target item from the long-term behavior sequence, while the second stage models the relationship between the subsequence and the target item. Despite significant progress, these methods have two critical flaws. First, the retrieval query typically includes only target item information, limiting the ability to capture the user's diverse interests. Second, relational information, such as sequential and interactive information within the subsequence, is frequently overlooked. Therefore, it requires to be further mined to more accurately model user interests. To this end, we propose Multi-granularity Interest Retrieval and Refinement Network (MIRRN). Specifically, we first construct queries based on behaviors observed at different time scales to obtain subsequences, each capturing users' interest at various granularities. We then introduce an noval multi-head Fourier transformer to efficiently learn sequential and interactive information within the subsequences, leading to more accurate modeling of user interests. Finally, we employ multi-head target attention to adaptively assess the impact of these multi-granularity interests on the target item. Extensive experiments have demonstrated that MIRRN significantly outperforms state-of-the-art baselines. Furthermore, an A/B test shows that MIRRN increases the average number of listening songs by 1.32% and the average time of listening songs by 0.55% on the Huawei Music App. The implementation code is publicly available at https://github.com/USTC-StarTeam/MIRRN.
Disclosing Actual Controllers (ACs) of a company has been the basis for financial risk governance. A shareholder in a winning stable coalition, where members make consistent decisions and win in votes, is considered an AC. However, existing methods fail to discover stable coalitions due to the ignorance of various relations other than the shareholding relation among shareholders, such as kinship, subsidiary and so on. Moreover, the above relations form a large-scale equity network, which brings challenges for efficiently identifying winning stable coalitions. We construct an Equity Knowledge Graph (EKG) to represent the semantic and structural information of the equity network. In this paper, we propose an AC disclosure method based on Equity Knowledge Graph Learning (EKGL). Specifically, to discover stable coalitions, EKGL designs a multi-relational aggregation module to aggregate the information of different relations horizontally. Based on the aggregated information, EKGL leverages a metapath-based aggregation module to encode the shareholding structure by capturing different shareholding paths on EKG vertically. To identify winning stable coalitions, we propose a control neural network to simulate the voting process of shareholders. Experiments and a case study on the EKG constructed from real datasets demonstrate that EKGL outperforms baselines by achieving 0.33 improvement in F1 score and reducing time cost.
Short-video sharing platforms engaging billions of users have attracted intense interest recently. A key insight is that user feedback on these platforms is heavily influenced by preceding exposed videos in the same request, called context cumulative effects. For example, multiple repeated videos in a request often cause user fatigue and influence user feedback. However, related factors, such as the other exposed items in the same request, are available during model training but not accessible during online serving. Vanilla distillation methods mitigate the training-inference inconsistency, struggling to capture the dynamic dependence between context cumulative effects and user feedback. To address this problem, we propose the Mutual Information-aware Knowledge Distillation (MIKD) framework, which fuses such effects and user-item matching degrees by evaluating their impacts on user feedback based on mutual information estimation. Rigorous analysis and extensive experiments demonstrate that MIKD precisely extracts personal interests and consistently improves performance. We conduct online A/B testing on a leading short-video sharing mobile app, and the results demonstrate the effectiveness of the proposed method. MIKD has been successfully deployed online to serve the main traffic and optimize user experiences.
Logistics services have become a core business in online-to-offline e-commerce like Amazon, Alibaba, and JD. In logistics services, a city is partitioned into distinct geographical areas, and each area is assigned a worker, responsible for all delivery tasks within it. Due to varying geographic conditions (e.g., high-rise buildings, buildings without elevators), the difficulty of completing tasks can differ significantly between areas, which results in unbalanced workloads and salaries for workers. The necessity for scalable data-driven methods to assess area difficulty in logistics is well-recognized. However, the significant expenses associated with ground truth data collection limit the capabilities of current machine learning methods. In this paper, we consider a frequently overlooked resource, i.e., the workers' firsthand knowledge of areas, to address this problem in a human-AI collaboration fashion. In particular, we design RAICA (Ranking-Aggregated Isotonic Calibration Assessment) framework, which includes two key modules: (i) a Judgment Rank Aggregation module, which aggregates individual workers' judgment rankings collected from surveys into an overall ranking to mitigate personal biases and inconsistency between different workers; (ii) an Isotonic Calibration module, which calibrates the assessment from existing machine learning models with the aggregated ranking through Isotonic regression to enhance the accuracy of area difficulty assessment with theoretical guarantees. Extensive evaluation based on real-world data including over 2 million orders collected from 97 areas during 6 months by one of the largest logistics companies in the world shows that RAICA outperforms existing methods, increasing F1 score by 0.25. More importantly, RAICA has been deployed by this logistics company, which significantly improved crowdsourcing couriers' salary fairness with a 0.2 decrease in the Gini coefficient across over 1,200 delivery stations nationwide and increased the on-time delivery rates for full-time couriers by 1.67%.
Last-mile Delivery Parcel Volume (LDPV) quantifies the number of parcels destined for a specific region, particularly a manually divided Area-Of-Interest (AOI). Accurate prediction of AOI-level LDPV is crucial for the efficient management of logistics resources. However, the straightforward adaptation of existing prediction models often falls short, primarily due to (I) a lack of consideration for the intuition behind AOI divisions, and (II) a reliance solely on fully observed historical data, which may not inform future trends. To overcome the above pitfalls, leveraging rich AOI data and advanced parcel travel time estimation services in JD Logistics, this paper introduces a novel framework called Dual-view Prediction Networks (DualPNs). It combines a Vector-Quantified AutoEncoder (VQ-AE) and a Template-Augmented Zero-Inflated Poisson (TA-ZIP), enabling both point and probabilistic distribution predictions of AOI-level LDPV. Specifically, VQ-AE utilizes a vector quantization technique to distill a large number of AOIs into representative templates, thereby addressing the first pitfall. Subsequently, TA-ZIP dynamically integrates fully observed and lookahead features, aligning them with template-specific decoders to parameterize the probabilistic distributions, thus resolving the second pitfall. We conduct extensive experiments in two cities, comprising over 47,000 and 126,000 AOIs respectively, to demonstrate the superiority of our DualPNs over other baselines. Moreover, a real-world case study highlights the effectiveness of DualPNs for enhancing downstream courier allocation by yielding an average improvement of 1.51% in the on-time delivery rate.
An up-to-date city-scale lane-level map is an indispensable infrastructure and a key enabling technology for ensuring the safety and user experience of autonomous driving systems. In industrial scenarios, reliance on manual annotation for map updates creates a critical bottleneck. Lane-level updates require precise change information and must ensure consistency with adjacent data while adhering to strict standards. Traditional methods utilize a three-stage approach -- construction, change detection, and updating -- which often necessitates manual verification due to accuracy limitations. This results in labor-intensive processes and hampers timely updates. To address these challenges, we propose LDMapNet-U, which implements a new end-to-end paradigm for city-scale lane-level map updating. By reconceptualizing the update task as an end-to-end map generation process grounded in historical map data, we introduce a paradigm shift in map updating that simultaneously generates vectorized maps and change information. To achieve this, a Prior-Map Encoding (PME) module is introduced to effectively encode historical maps, serving as a critical reference for detecting changes. Additionally, we incorporate a novel Instance Change Prediction (ICP) module that learns to predict associations with historical maps. Consequently, LDMapNet-U simultaneously achieves vectorized map element generation and change detection. To demonstrate the superiority and effectiveness of LDMapNet-U, extensive experiments are conducted using large-scale real-world datasets. In addition, LDMapNet-U has been successfully deployed in production at Baidu Maps since April 2024, supporting lane-level map updating for over 360 cities and significantly shortening the update cycle from quarterly to weekly, thereby enhancing the timeliness and accuracy of lane-level map. The nationwide, high-frequency city-scale lane-level map has been instrumental in the development of the lane-level navigation product serving hundreds of millions of users, while also integrating into the autonomous driving systems of several leading vehicle companies.
The ever-increasing sensor service, though opening a precious path and providing a deluge of earth system data for deep-learning-oriented earth science, sadly introduce a daunting obstacle to their industrial level deployment. Concretely, earth science systems rely heavily on the extensive deployment of sensors, however, the data collection from sensors is constrained by complex geographical and social factors, making it challenging to achieve comprehensive coverage and uniform deployment. To alleviate the obstacle, traditional approaches to sensor deployment utilize specific algorithms to design and deploy sensors. These methods dynamically adjust the activation times of sensors to optimize the detection process across each sub-region. Regrettably, formulating an activation strategy generally based on historical observations and geographic characteristics, which make the methods and resultant models were neither simple nor practical. Worse still, the complex technical design may ultimately lead to a model with weak generalizability. In this paper, we introduce for the first time the concept of spatio-temporal data dynamic sparse training and are committed to adaptively, dynamically filtering important sensor distributions. To our knowledge, this is the first proposal (termed DynST) of an industry-level deployment optimization concept at the data level. However, due to the existence of the temporal dimension, pruning of spatio-temporal data may lead to conflicts at different timestamps. To achieve this goal, we employ dynamic merge technology, along with ingenious dimensional mapping to mitigate potential impacts caused by the temporal aspect. During the training process, DynST utilize iterative pruning and sparse training, repeatedly identifying and dynamically removing sensor perception areas that contribute the least to future predictions. DynST demonstrates tremendous capability on industrial-grade data from JD Technology TaxiBJ+ and practical deployment scenarios such as meteorology, combustion dynamics, and turbulence. It seamlessly integrates with relevant models and efficiently prunes image and graph-type data, leading to significantly higher inference speeds without introducing noticeable performance degradation.
It has become increasingly clear that recommender systems that overly focus on short-term engagement prevents users from exploring diverse interests, ultimately hurting long-term user experience. To tackle this challenge, numerous diversification algorithms have been proposed as the final stage of recommender systems. These algorithms typically rely on measures of item similarity, aiming to maximize the dissimilarity across items in the final set of recommendations. However, in this work, we demonstrate the benefits of going beyond item-level similarities by utilizing higher-level user understanding-specifically, user intents that persist across multiple interactions or recommendation sessions-in diversification. Our approach is motivated by the observation that user behaviors on online platforms are largely driven by their underlying intents. Therefore, final recommendations should ensure that a diverse set of user intents is accurately represented. While user intent has primarily been studied in the context of search, it is less clear how to incorporate real-time dynamic intent predictions in recommender systems. To address this gap, we develop a probabilistic intent-based whole-page diversification framework for the final stage of a recommender system. Starting with a prior belief of user intents, the proposed framework sequentially selects items for each position based on these beliefs and subsequently updates posterior beliefs about the intents. This approach ensures that different user intents are represented on a page, towards optimizing long-term user experience. We experiment with the intent diversification framework on YouTube, the world's largest video recommendation platform, serving billions of users daily. Live experiments on a diverse set of intents show that the proposed framework increases Daily Active Users (DAU) and overall user enjoyment, validating its effectiveness in facilitating long-term planning. Specifically, it enables users to consistently discover and engage with diverse content that aligns with their underlying intents over time, leading to an improved long-term user experience.
Applying large language models (LLMs) to academic API usage shows promise in reducing researchers' efforts to seek academic information. However, current LLM methods for using APIs struggle with the complex API coupling commonly encountered in academic queries. To address this, we introduce SoAy, a solution-based LLM methodology for academic information seeking. SoAy enables LLMs to generate code for invoking APIs, guided by a pre-constructed API calling sequence referred to as a solution. This solution simplifies the model's understanding of complex API relationships, while the generated code enhances reasoning efficiency. LLMs are aligned with this solution-oriented, code-based reasoning method by automatically enumerating valid API coupling sequences and transforming them into queries and executable code. To evaluate SoAy, we introduce SoAyBench, an evaluation benchmark accompanied by SoAyEval, built upon a cloned environment of APIs from AMiner. Experimental results demonstrate a 34.58-75.99% performance improvement compared to state-of-the-art LLM API-based baselines. All datasets, codes, tuned models, and deployed online services are publicly accessible at https://github.com/RUCKBReasoning/SoAy.
Recommender systems play a crucial role in online platforms, providing personalized recommendations for purchases, content consumption, and interpersonal connections. These systems involve two sides: producers (sellers, content creators, service providers, etc.) and consumers (buyers, viewers, customers, etc.). To optimize online recommender systems, A/B tests serve as the golden standard for comparing different ranking models and evaluating their impacts on both sides. While consumer-side experiments are relatively straightforward to design and commonly employed to assess ranking changes' effects on the behavior of consumers (buyers, viewers, etc.), designing producer-side experiments for an online recommender/ranking system is notably more complex. This complexity arises from the necessity of ranking producer items in the treatment and control groups by different models and then merging them into a unified ranking for presentation to each consumer. Existing design solutions in the literature lack rigorous guiding principles, leading to ad hoc approaches. In this paper, we address the limitations of current methods and propose the principles of consistency and monotonicity for designing producer-side experiments in online recommender systems. Building upon these principles, we also present a systematic solution based on counterfactual interleaving designs to accurately measure the impacts of ranking changes on the producers (sellers, content creators, etc.).
In this paper, we present the practical problems and the lessons learned at short-video services from Kuaishou. In industry, a widely-used multi-task framework is the Mixture-of-Experts (MoE) paradigm, which always introduces some shared and specific experts for each task and then uses gate networks to measure related experts' contributions. Although the MoE achieves remarkable improvements, we still observe three anomalies that seriously affect model performances in our iteration: (1) Expert Collapse: We found that experts' output distributions are significantly different, and some experts have over 90% zero activations with ReLU, making it hard for gate networks to assign fair weights to balance experts. (2) Expert Degradation: Ideally, the shared-expert aims to provide predictive information for all tasks simultaneously. Nevertheless, we find that some shared-experts are occupied by only one task, which indicates that shared-experts lost their ability but degenerated into some specific-experts. (3) Expert Underfitting: In our services, we have dozens of behavior tasks that need to be predicted, but we find that some data-sparse prediction tasks tend to ignore their specific-experts and assign large weights to shared-experts. The reason might be that the shared-experts can perceive more gradient updates and knowledge from dense tasks, while specific-experts easily fall into underfitting due to their sparse behaviors. Motivated by those observations, we propose HoME to achieve a simple, efficient and balanced MoE system for multi-task learning. Specifically, we conduct three insightful modifications: (1) Expert normalization&Swish mechanism to align expert output distributions and avoid expert collapse. (2) Hierarchy mask mechanism to enhance sharing efficiency between tasks to reduce occupancy issues and away from expert degradation. (3) Feature-gate&Self-gate mechanisms to ensure each expert could obtain appropriate gradient to maximize its effectiveness. To our knowledge, this paper is the first work to focus on improving multi-task MoE system stability, and we conduct extensive offline&online (average improves 0.52% GAUC offline & 0.954% play-time per user online) experiments and ablation analyses to demonstrate our HoME effectiveness. From March 2024, our HoME has been widely deployed on various services at Kuaishou, supporting 400 Million active users daily.
In a single-slot recommendation system, users are only exposed to one item at a time, and the system cannot collect user feedback on multiple items simultaneously. Therefore, only pointwise modeling solutions can be adopted, focusing solely on modeling the likelihood of clicks or conversions for items by users to learn user-item preferences, without the ability to capture the ranking information among different items directly. However, since user-side information is often much more abundant than item-side information, the model can quickly learn the differences in user intrinsic tendencies, which are independent of the items they are exposed to. This can cause these intrinsic tendencies to become a shortcut bias for the model, leading to insufficient mining of the most concerned user-item preferences. To solve this challenge, we introduce the Breaker model. Breaker integrates an auxiliary task of user representation clustering with a multi-tower structure for cluster-specific preference modeling. By clustering user representations, we ensure that users within each cluster exhibit similar characteristics, which increases the complexity of the pointwise recommendation task on the user side. This forces the multi-tower structure with cluster-driven parameter learning to better model user-item preferences, ultimately eliminating shortcut biases related to user intrinsic tendencies. In terms of training, we propose a delayed parameter update mechanism to enhance training stability and convergence, enabling end-to-end joint training of the auxiliary clustering and classification tasks. Both offline and online experiments demonstrate that our method surpasses the baselines. It has already been deployed and is actively serving tens of millions of users daily on Meituan, one of the most popular e-commerce platforms for services.
Within the realm of drug development, the transition from successful animal trials to human clinical efficacy remains a daunting challenge. While initial outcomes may appear promising in animal studies, ensuring similar effectiveness in humans, especially across specific target tissues, presents a significant obstacle. To address this pressing concern, we introduce a novel generative model tailored to optimize molecules that have demonstrated efficacy in rats for enhanced performance in specific human tissues. Central to our solution is the transformer architecture, enhanced with intricate mechanisms such as molecule self-attention within the encoder and a novel dedicated tissue-specific generator. Intuitively, by learning to generate molecules simultaneously from multiple tissues, the generative model enhances its ability to perform the necessary adaptations from rats to humans. Through rigorous empirical evaluation across various tissues, our model consistently exhibits remarkable efficacy compared to existing methods. We anticipate that this model has the potential to minimize the requirement for lengthy and inconclusive trials, thereby streamlining the drug development process.
In book search, relevant book information should be returned in response to a query. Books contain complex, multi-faceted information such as metadata, outlines, and main text, where the outline provides hierarchical information between chapters and sections. Generative retrieval (GR) is a new retrieval paradigm that consolidates corpus information into a single model to generate identifiers of documents that are relevant to a given query. How can GR be applied to book search? Directly applying GR to book search is a challenge due to the unique characteristics of book search: The model needs to retain the complex, multi-faceted information of the book, which increases the demand for labeled data. Splitting book information and treating it as a collection of separate segments for learning might result in a loss of hierarchical information. We propose an effective Generative retrieval framework for Book Search (GBS) that features two main components: data augmentation and outline-oriented book encoding. For data augmentation, GBS constructs multiple query-book pairs for training; it constructs multiple book identifiers based on the outline, various forms of book contents, and simulates real book retrieval scenarios with varied pseudo-queries. This includes coverage-promoting book identifier augmentation, allowing the model to learn to index effectively, and diversity-enhanced query augmentation, allowing the model to learn to retrieve effectively. Outline-oriented book encoding improves length extrapolation through bi-level positional encoding and retentive attention mechanisms to maintain context over long sequences. Experiments on a proprietary Baidu dataset demonstrate that GBS outperforms strong baselines, achieving a 9.8% improvement in terms of MRR@20, over the state-of-the-art RIPOR method. Experiments on public datasets confirm the robustness and generalizability of GBS, highlighting its potential to enhance book retrieval.