论文检索

Rui Liu, Rui Xie 0002, Zijun Yao 0001, Yanjie Fu, Dongjie Wang

Feature selection removes redundant features to enhance both performance and computational efficiency in downstream tasks. Existing methods often struggle to capture complex feature interactions and adapt to diverse scenarios. Recent advances in this domain have incorporated generative intelligence to address these drawbacks by uncovering intricate relationships between features. However, two key limitations remain: 1) embedding feature subsets in a continuous space is challenging due to permutation sensitivity, as changes in feature order can introduce biases and weaken the embedding learning process; 2) gradient-based search in the embedding space assumes convexity, which is rarely guaranteed, leading to reduced search effectiveness and suboptimal subsets. To address these limitations, we propose a new framework that can: 1) preserve feature subset knowledge in a continuous embedding space while ensuring permutation invariance; 2) effectively explore the embedding space without relying on strong convex assumptions. For the first objective, we develop an encoder-decoder paradigm to preserve feature selection knowledge into a continuous embedding space. This paradigm captures feature interactions through pairwise relationships within the subset, removing the influence of feature order on the embedding. Moreover, an inducing point mechanism is introduced to accelerate pairwise relationship computations. For the second objective, we employ a policy-based reinforcement learning (RL) approach to guide the exploration of the embedding space. The RL agent effectively navigates the space by balancing multiple objectives. By prioritizing high-potential regions adaptively and eliminating the reliance on convexity assumptions, this search strategy effectively reduces the risk of converging to local optima. Finally, we conduct extensive experiments to demonstrate the effectiveness, efficiency, robustness and explicitness of our model. Our code and dataset are publicly accessible on GitHub. https://github.com/RayLiu1103/CAPS.

Runchang Liang, Qi Hao 0001, Yue Gao 0015, Kunpeng Liu 0001, Lu Jiang 0007, Pengyang Wang, Minghao Yin

Multivariate time series forecasting in practical deployment faces a critical challenge termed Variable Subset Forecasting (VSF), where certain variables accessible during training are entirely missing during inference. This creates a stark discrepancy between the training (source domain with full variables) and inference (target domain with partial variables) environments, disrupting cross-variable dependencies and fragmenting global temporal patterns. Existing imputation methods, limited to transferring local knowledge (e.g., temporal neighbors or pairwise correlations), fail to capture essential global dynamics, leading to severe performance degradation under distribution shifts. To address these challenges, we redefine VSF as a cross-domain knowledge transfer problem and propose VIDA, a framework that systematically transfers Variable Invariant knowledge from complete to partial observations through Domain Adaptation. Key to our approach is (1) Global time-frequency joint representation learning, which encodes temporal dynamics via dilated convolutions and captures low-frequency spectral consistency using Fourier neural operators, and (2) Sinkhorn-regularized distribution alignment to bridge non-overlapping feature supports across domains via optimal transport. Unlike imputation-first methods, VIDA enforces task-driven consistency by jointly optimizing predictions on reconstructed and original data, ensuring the transferred knowledge directly enhances forecasting robustness. Extensive experiments across four real-world datasets show that VIDA outperforms state-of-the-art imputation methods by 25% on average with partially observed variables. This work establishes a new paradigm for variable-missing scenarios by unifying imputation and forecasting through principled knowledge transfer.

Ruifeng Li 0002, Mingqian Li, Wei Liu, Hongyang Chen 0001

Effective molecular representation learning is crucial for advancing molecular property prediction and drug design. Mainstream molecular representation learning methods rely on Graph Neural Networks (GNNs) to model complex molecular structures. However, these GNN-based approaches struggle with three significant issues: insufficient annotations, molecular diversity, and architectural limitations such as over-squashing, which leads to the loss of critical structural details. To address these issues, we introduce a new class of GNNs that integrates the Kolmogorov-Arnold Networks (KANs), known for their robust data-fitting capabilities and high accuracy in small-scale AI + Science tasks. By incorporating KANs into GNNs, our model enhances the representation of molecular structures. We further advance this approach with a variant called SwallowKAN (SKAN), which employs adaptive Radial Basis Functions (RBFs) as the core of the non-linear neurons. This innovation improves both computational efficiency and adaptability to diverse molecular structures. Building on the strengths of SKAN, we propose a new class of GNNs, GNN-SKAN, and its augmented variant, GNN-SKAN+, which incorporates a SKAN-based classifier to further boost the performance. Experiments across 6 classification datasets, 6 regression datasets, and 4 few-shot learning datasets demonstrate that our approach achieves new state-of-the-art performance in terms of accuracy and computational cost.

Ke Li 0019, Lisi Chen 0001, Shuo Shang, Christian S. Jensen, Panos Kalnis

With the proliferation of GPS-enabled devices such as smartphones, the querying of moving objects has attracted substantial attention, with studies covering joins, range and kNN queries, similarity queries, etc. Challenges arise due to variable sampling frequencies, potential inaccuracies in location samples, and the unavailability of locations between samples. Existing similarity joins often rely on discrete location samples, which fail to capture movement uncertainty and may miss meaningful interactions. To address this limitation, we propose Intersection Similarity Join (IS-Join), a novel approach that identifies object pairs based on the overlap of their motion ranges rather than location-based proximity. We define motion ranges as the spatial regions an object may traverse within a given time period, and introduce an intersection similarity measure that quantifies their overlap. To efficiently process IS-Join queries, we develop a Hybrid Ball-tree indexing structure with a repartitioning strategy, enabling scalable candidate filtering. Additionally, we introduce pre-checking and pruning techniques to further reduce computational overhead. Extensive experiments on two real-world trajectory datasets demonstrate that IS-Join significantly outperforms well-designed baselines, achieving up to a 3x reduction in runtime. Our work opens new opportunities for applications such as urban mobility analysis, traffic monitoring, wildlife tracking, and contact tracing.

Junjun Li, Zeyuan Ma, Ting Huang 0001, Yue-Jiao Gong

Lifelong Multi-Agent Path Finding (LMAPF) focuses on planning conflict-free paths for agents, like autonomous vehicles, that are continuously assigned new tasks. The synergy of search-based and learning-based methods holds promise for striking a balance in-between effectiveness and efficiency but still faces several challenges such as inferior initial paths, weak search-learning synergy and low sample utilization rate. To address these issues, this paper proposes a new synergized LMAPF approach, named Synergistic Multi-Agent Path Optimization (SMAPO), which consists of two tightly-coupled phases: Primordial Planning and Decision Refinement. In the Primordial Planning phase, we introduce a novel load-balanced A* algorithm that integrates planned and perceived congestion costs, which enhances initial solution quality by evenly distributing spatiotemporal traffic loads, thereby mitigating potential conflicts. In the Decision Refinement phase, we propose a novel Encoder-Decoder based neural network to learn a collaborative optimization policy through multi-agent reinforcement learning. In addition, we leverage dual transformations to augment trajectory samples during online learning, enhancing both the sample utilization rate and overall learning stability. Extensive experiments reveal that our SMAPO is superior to the state-of-the-art baselines in effectiveness, efficiency, and generalization capability. Source code is available at https://github.com/ByteUser-blues/SMAPO.

Jian Lang, Rongpei Hong, Zhangtao Cheng, Ting Zhong, Yong Wang 0046, Fan Zhou 0002

Traditional multimodal learning approaches often assume that all modalities are available during both the training and inference phases. However, this assumption is often impractical in real-world scenarios due to challenges such as sensor failures, data corruption, or privacy concerns. While recent efforts focus on enhancing the robustness of pre-trained Multimodal Transformers (MTs) under missing modality conditions, mainstream work in this field often overlook reconstructing the missing modalities and rely on static, sample-agnostic prompt-tuning techniques, undermining their efficacy in severe modality missing scenarios. To address these limitations, we propose REDEEM, a novel REtrieval-guiDEd conditional gEnerative fraMework that largely alleviates the modality missing problems on pre-trained MTs. REDEEM consists of a new adaptive retrieval mechanism to identify relevant instances for both modality-complete and -incomplete samples. It then conditions on the remaining modalities and utilizes the retrieved data as experts to effectively recover the missing ones in modality-incomplete instances through a within-modal reconstruction manner. Finally, REDEEM generates sample-aware inter-modal prompts from the retrieved instances to guide MTs in tackling severe modality missing challenges. Comprehensive experiments on three diverse multimodal classification benchmarks demonstrate that REDEEM significantly outperforms competitive baselines.

Ricardo Knauer, Mario Koddenbrock, Raphael Wallsberger, Nicholas M. Brisson, Georg N. Duda, Deborah Falla, David W. Evans, Erik Rodner

Large language models (LLMs) provide powerful means to leverage prior knowledge for predictive modeling when data is limited. In this work, we demonstrate how LLMs can use their compressed world knowledge to generate intrinsically interpretable machine learning models, i.e., decision trees, without any training data. We find that these zero-shot decision trees can even surpass data-driven trees on some small-sized tabular datasets and that embeddings derived from these trees perform better than data-driven tree-based embeddings on average. Our decision tree induction and embedding approaches can therefore serve as new knowledge-driven baselines for data-driven machine learning methods in the low-data regime. Furthermore, they offer ways to harness the rich world knowledge within LLMs for tabular machine learning tasks. Our code and results are available at https://github.com/ml-lab-htw/llm-trees.

Hyunsik Jeon, Satoshi Koide, Yu Wang 0170, Zhankui He, Julian J. McAuley

Conversational recommender systems engage users in dialogues to refine their needs and provide more personalized suggestions. Although textual information suffices for many domains, visually driven categories such as fashion or home decor potentially require detailed visual information related to color, style, or design. To address this challenge, we propose LaViC (Large Vision-Language Conversational Recommendation Framework), a novel approach that integrates compact image representations into dialogue-based recommendation systems. LaViC leverages a large vision-language model in a two-stage process: (1) visual knowledge self-distillation, which condenses product images from thousands of tokens into a small set of visual tokens in a self-distillation manner, significantly reducing computational overhead, and (2) recommendation fine-tuning, which enables the model to incorporate both dialogue context and distilled visual tokens, providing a unified mechanism for capturing textual and visual features. To support rigorous evaluation of visually-aware conversational recommendation, we construct a new dataset by aligning Reddit conversations with Amazon product listings across multiple visually oriented categories (e.g., fashion, beauty, and home). This dataset covers realistic user queries and product appearances in domains where visual details are crucial. Extensive experiments demonstrate that LaViC significantly outperforms text-only conversational recommendation methods and open-source vision-language baselines. Moreover, LaViC achieves competitive or superior accuracy compared to prominent proprietary baselines (e.g., GPT-3.5-turbo, GPT-4o-mini, and GPT-4o), demonstrating the necessity of explicitly using visual data for capturing product attributes and showing the effectiveness of our vision-language integration. Our code and dataset are available at https://github.com/jeon185/LaViC.

Zijian He, Reyna Abhyankar, Vikranth Srivatsa, Yiying Zhang 0005

Today's gen-AI workflows that involve multiple ML model calls, tool/API calls, data retrieval, or generic code execution are often tuned manually in an ad-hoc way that is both time-consuming and error-prone. In this paper, we propose a systematic approach for automatically tuning gen-AI workflows. Our key insight is that gen-AI workflows can benefit from structure, operator, and prompt changes, but unique properties of gen-AI workflows require new optimization techniques. We propose AdaSeek, an adaptive hierarchical search algorithm for autotuning gen-AI workflows. AdaSeek organizes workflow tuning methods into different layers based on the user-specified total search budget and distributes the budget across different layers based on the complexity of each layer. During its hierarchical search, AdaSeek redistributes the search budget from less useful to more promising tuning configurations based on workflow-level evaluation results. We implement AdaSeek in a workflow autotuning framework called Cognify and evaluate Cognify using six types of workflows such as RAG-based QA and text-to-SQL transformation. Overall, Cognify improves these workflows' generation quality by up to 2.8×, reduces execution monetary cost by up to 10×, and reduces end-to-end latency by 2.7×.

Mohsen Dehghankar, Abolfazl Asudeh

Large language models (LLMs) have quickly emerged as practical and versatile tools that provide new solutions for a wide range of domains. In this paper, we consider the application of LLMs on symmetric tasks where a query is asked on an (unordered) bag of elements. Examples of such tasks include answering aggregate queries on a dataset. In general, when the bag contains a large number of elements, LLMs tend to overlook some elements, leading to challenges in generating accurate responses to the query. LLMs receive their inputs as ordered sequences. On the other hand, the input of a symmetric task is not ordered; hence reordering should not affect the LLM's response. Observing that LLMs are less likely to miss elements at certain positions of the input, we introduce the problem of LLM input reranking: to find a ranking of the input that maximizes the LLM's accuracy for the given query without making explicit assumptions about the query. Finding the optimal ranking requires identifying (i) the relevance of each input element for answering the query and (ii) the importance of each rank position for the LLM's attention. We develop algorithms for estimating these values efficiently utilizing a helper LLM. We conduct comprehensive experiments on different synthetic and real datasets to validate our proposal and to evaluate the effectiveness of our proposed algorithms. Our experiments confirm that our reranking approach improves the accuracy of the LLMs on symmetric tasks by up to 99% proximity to the optimum upper bound.

Xiangxiang Dai, Xiaowei Sun, Jinhang Zuo, Xutong Liu 0002, John C. S. Lui

Co-branding has become a vital strategy for businesses aiming to expand market reach within recommendation systems. However, identifying effective cross-industry partnerships remains challenging due to resource imbalances, uncertain brand willingness, and ever-changing market conditions. In this paper, we provide the first systematic study of this problem and propose a unified online-offline framework to enable co-branding recommendations. Our approach begins by constructing a bipartite graph linking ''initiating'' and ''target'' brands to quantify co-branding probabilities and assess market benefits. During the online learning phase, we dynamically update the graph in response to market feedback, while striking a balance between exploring new collaborations for long-term gains and exploiting established partnerships for immediate benefits. To address the high initial co-branding costs, our framework mitigates redundant exploration, thereby enhancing short-term performance while ensuring sustainable strategic growth. In the offline optimization phase, our framework consolidates the interests of multiple sub-brands under the same parent brand to maximize overall returns, avoid excessive investment in single sub-brands, and reduce unnecessary costs associated with over-prioritizing a single sub-brand. We present a theoretical analysis of our approach, establishing a highly nontrivial sublinear regret bound for online learning in the complex co-branding problem, and enhancing the approximation guarantee for the NP-hard offline budget allocation optimization. Experiments on both synthetic and real-world co-branding datasets demonstrate the practical effectiveness of our framework, with at least 12% improvement.

Christopher D. Manning

Language Models have been around for decades but have suddenly taken the world by storm. In a surprising third act for anyone doing Natural Language Processing (NLP) in the 70s, 80s, 90s, or 2000s, in much of the popular media, artificial intelligence is now synonymous with language models. In this talk, I want to take a look backward at where language models came from and why they were so slow to emerge, a look inward to give a few thoughts on meaning, intelligence, and what language models understand and know, and a look forward at some topics of recent research: the possibilities for progress with new versions of Universal Transformers and using Large Language Models (LLMs) to build intelligent language-using agents for the digital world. I will argue that material beyond language is not necessary to having meaning and understanding, but it is very useful in most cases, and that composability, adaptability, and learning are vital to intelligence. Rather than simply continuing to build ever-larger LLMs from huge passive collections of text data, we need to explore developing better neural architectures and agents that can learn through interactions. I will discuss how web agents can learn through interactions and how such an interaction-first learning approach can work very effectively in web environments with relatively small language models.

Qinchen Yang 0001, Zhiqing Hong, Dongjiang Cao, Haotian Wang 0008, Zejun Xie, Tian He 0001, Yunhuai Liu, Yu Yang 0010, Desheng Zhang 0002

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.

Deguo Xia, Weiming Zhang 0006, Xiyan Liu, Wei Zhang 0088, Chenting Gong, Xiao Tan 0001, Jizhou Huang, Mengmeng Yang 0001, Diange Yang

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.

Rajdeep Mukherjee, Sonali Singh, Sachin Farfade

Products on e-commerce platforms are usually organized based on seller-provided product attributes. Customers looking for a product typically have certain needs or use cases in mind, such as headphones for gym classes, or a printer for school projects. However, they often struggle to map these use cases to product attributes, thereby failing to find the product they need. To help customers shop online confidently, we propose a Use case-Based Shopping (UBS) system that facilitates customer experiences based on use cases (Fig. 1). UBS consists of three steps: 1) use case phrase extraction from product reviews, 2) clustering the extracted use case phrases to identify the dominant ones, and 3) mapping products in a given category to one or more of these dominant use cases. In this work, we utilize instruction-tuned LMs to primarily focus on the first two steps. However, the way we design them also helps us to seamlessly solve the third step to complete the design of our proposed system. First, we define the novel task of joint Use Uase, Uentiment Uxtraction (UCSE) from product reviews which can be used for both steps 1 and 3. We harness the task adaptation capability of instruction-tuned FLAN-T5 models and gradually improve their zero-shot UCSE performance through instruction tuning, multi-task training, and few-shot iterative re-training for new categories, achieving around ~90% reduction in annotation bandwidth. We then employ Anthropic's Claude 2 LLM to propose an unsupervised approach for hierarchical use case phrase clustering that demonstrates better clustering and cluster naming capabilities when compared to K-Means and LDA. In an online experiment targeting the top 7 product categories, UBS recommendations on search, browse, and product pages resulted in a revenue lift of 0.77%, 0.94%, and 0.44% respectively, and an average click rate lift of 0.15%.

Seungyeon Lee 0002, Ruoqi Liu, Feixiong Cheng, Ping Zhang 0016

Drug repurposing identifies new therapeutic uses for existing drugs, reducing the time and costs compared to traditional de novo drug discovery. Most existing drug repurposing studies using real-world patient data often treat the entire population as homogeneous, ignoring the heterogeneity of treatment responses across patient subgroups. This approach may overlook promising drugs that benefit specific subgroups but lack notable treatment effects across the entire population, potentially limiting the number of repurposable candidates identified. To address this, we introduce STEDR, a novel drug repurposing framework that integrates subgroup analysis with treatment effect estimation. Our approach first identifies repurposing candidates by emulating multiple clinical trials on real-world patient data and then characterizes patient subgroups by learning subgroup-specific treatment effects. We deploy STEDR to Alzheimer's Disease (AD), a condition with few approved drugs and known heterogeneity in treatment responses. We emulate trials for over one thousand medications on a large-scale real-world database covering over 8 million patients, identifying 14 drug candidates with beneficial effects to AD in characterized subgroups. Experiments demonstrate STEDR superior capability in identifying repurposing candidates compared to existing approaches. Additionally, our method can characterize clinically relevant patient subgroups associated with important AD-related risk factors, paving the way for precision drug repurposing.

Zhipeng Jin, Wen Tao, Yafei Li, Yi Yang 0031, Cong Han 0002, Shuanglong Li, Lin Liu

Recent advances in generative artificial intelligence have revolutionized information retrieval and content generation, opening up new opportunities for the e-commerce industry. Alignment learning between small models and parallel corpora cannot meet current needs. The success of ChatGPT demonstrates that large models need to first establish a fundamental understanding, and then utilize high-quality corpora for generation. Having a large model foundation is indispensable. In this paper, we establish a fundamental 10B multimodal model foundation for multimodal generation tasks and propose a scene-based alignment learning approach called conditional sample supervised fine-tuning for downstream generation tasks. Meanwhile, diffusion models are known to be vulnerable to outliers in training data. To address this, we utilize an alternative diffusion loss function that preserves the high quality of generated data like the original squared L2 loss while being robust to outliers.In practical test sets, the multimodal foundation fully demonstrates its alignment and comprehension abilities for graphic and textual content. Additionally, conditional fine-tuning and the design of the loss function significantly enhance the quality of generated content. The quality rate of images has increased by 34.3 percentage points, and prompt control has improved by 19.8 percentage points. The application of our framework in Baidu Search Ads has led to significant revenue growth. For instance, ads with generated image creatives have achieved a 29% higher click-through rate (CTR), resulting in a daily consumption of 3 million yuan.

Medhasree Ghosh, Chirag Dinesh Jain, Raju Halder, Joydeep Chandra

Phishing scams on the Ethereum network have become a serious threat, especially with the influx of new users into the cryptocurrency market. Current detection methods are mainly focused on long-term consistent transaction patterns with smooth temporal dynamics. However, these methods often struggle to differentiate between phishing and non-phishing users, whose behaviours may appear deceptively similar. Additionally, they face challenges such as network sparsity and data leakage, leading to significant performance limitations. To address these issues, we introduce TEMPER, a novel sequential learning framework designed to jointly capture the subtle distinctions between long- and short-term user behaviours and their correlations to provide more comprehensive insights. TEMPER effectively generates distinguishable user embeddings, enabling the accurate identification of phishing users. Unlike previous approaches, TEMPER mitigates data leakage through a novel sequential transaction sampling algorithm and addresses network sparsity with short-term temporal learning. Through extensive experimentation on three real-world Ethereum datasets, TEMPER demonstrates its efficacy by achieving a 3-4% improvement in the F1-Score compared to existing baseline models, representing a significant advancement in Ethereum phishing user detection.

Jianian Wang, Rui Song 0006

To represent the causal relationships between variables, a directed acyclic graph (DAG) is widely utilized in many areas, such as social sciences, epidemics, and genetics. Many causal structure learning approaches are developed to learn the hidden causal structure using deep learning approaches. However, these approaches have a hidden assumption that the causal relationship remains unchanged over time, which may not hold in real life. In this paper, we develop a new framework to model the dynamic causal graph where the causal relations are allowed to be time-varying. We incorporate the basis approximation method into the score-based causal discovery approach to capture the dynamic pattern of causal graphs. Utilizing the autoregressive model structure, we could capture both contemporaneous and time-lagged causal relationships while allowing them to vary with time. We propose an algorithm that could provide both past-time estimates and future-time predictions on the causal graphs, and conduct simulations to demonstrate the usefulness of the proposed method. We also apply the proposed method for the covid-data analysis, and provide causal estimates on how the effect of policy restriction changes.

Ravisri Valluri, Akash Kumar Mohankumar, Kushal Dave 0001, Amit Singh 0003, Jian Jiao 0007, Manik Varma, Gaurav Sinha 0001

Generative Retrieval introduces a new approach to Information Retrieval by reframing it as a constrained generation task, leveraging recent advancements in Autoregressive (AR) language models. However, AR-based Generative Retrieval methods suffer from high inference latency and cost compared to traditional dense retrieval techniques, limiting their practical applicability. This paper investigates fully Non-autoregressive (NAR) language models as a more efficient alternative for generative retrieval. While standard NAR models alleviate latency and cost concerns, they exhibit a significant drop in retrieval performance (compared to AR models) due to their inability to capture dependencies between target tokens. To address this, we question the conventional choice of limiting the target token space to solely words or sub-words. We propose PIXNAR, a novel approach that expands the target vocabulary of NAR models to include multi-word entities and common phrases (up to 5 million tokens), thereby reducing token dependencies. PIXNAR employs inference optimization strategies to maintain low inference latency despite the significantly larger vocabulary. Our results demonstrate that PIXNAR achieves a relative improvement of 31.0% in MRR@10 on MS MARCO and 23.2% in Hits@5 on Natural Questions compared to standard NAR models with similar latency and cost. Furthermore, online A/B experiments on a large commercial search engine show significant increase in clicks and revenue.