Zhaopeng Feng, Yupu Liang, Shaosheng Cao, Jiayuan Su, Jiahan Ren, Zhijie Zhou, Wenxuan Huang, Jian Wu, Zuozhu Liu. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Zhiwei Zhang, Fei Zhao, Rui Wang, Zezhong Wang, Bin Liang, Jiakang Wang, Yao Hu, Shaosheng Cao, Kam-Fai Wong. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Ranking relevance is a fundamental task in search engines, aiming to identify the items most relevant to a given user query. Traditional relevance models typically produce scalar scores or directly predict relevance labels, limiting both interpretability and the modeling of complex relevance signals. Inspired by recent advances in Chain-of-Thought (CoT) reasoning for complex tasks, we investigate whether explicit reasoning can enhance both interpretability and performance in relevance modeling. However, existing reasoning-based Generative Relevance Models (GRMs) primarily rely on supervised fine-tuning on large amounts of human-annotated or synthetic CoT data, which often leads to limited generalization. Moreover, domain-agnostic, free-form reasoning tends to be overly generic and insufficiently grounded, limiting its potential to handle the diverse and ambiguous cases prevalent in open-domain search. In this work, we formulate relevance modeling in Xiaohongshu search as a reasoning task and introduce a Reinforcement Learning (RL)-based training framework to enhance the grounded reasoning capabilities of GRMs. Specifically, we incorporate practical business-specific relevance criteria into the multi-step reasoning prompt design and propose Stepwise Advantage Masking (SAM), a lightweight process-supervision strategy which facilitates effective learning of these criteria through improved credit assignment. To enable industrial deployment, we further distill the large-scale RL-tuned model to a lightweight version suitable for real-world search systems. Extensive offline evaluations and online A/B tests demonstrate that our approach consistently delivers significant improvements across key relevance and business metrics, validating its effectiveness, robustness, and practicality for large-scale industrial search systems.

Heterogeneous graph neural networks (HGNNs) are effective for modeling multi-relational structured data. Existing HGNNs usually assume the training samples are relatively sufficient, thus focusing on improving the predictive performance by complicating the model architecture with more learnable parameters. In this paper, we instead explore how to design HGNNs when training labels are scarce, under which we observe that existing HGNNs suffer from serious overfitting issues. Inspired by the graph random neural network (GRAND)-a consistency regularization framework for graph learning, we propose a simple yet efficient R-GRAND framework to overcome the issues above. R-GRAND is a general relation-aware consistency regularized training method with both labeled and unlabeled nodes to facilitate the model's generalization capability. It designs a lightweight relational graph convolution neural network (SRGC) as the backbone model to deal with the heterogeneous information. To enable regularized training, we further advance the data augmentation methods of GRAND with a Multi-block DropEdge strategy. The proposed training framework not only excels with its default SRGC backbone but also effectively enhances the performance of other HGNN architectures, such as RGCN and Simple-HGN. Extensive experiments on seven heterogeneous graph datasets demonstrate that R-GRAND can achieve remarkable performance improvements over state-of-theart HGNNs with better generalization ability and high efficiency.

As interest in using Large Language Models for interactive and emotionally rich experiences grows, virtual pet companionship emerges as a novel yet underexplored application. Existing approaches focus on basic pet role-playing interactions without systematically benchmarking LLMs for comprehensive companionship. In this paper, we introduce PET-BENCH, a dedicated benchmark that evaluates LLMs across both self-interaction and human-interaction dimensions. Unlike prior work, PET-BENCH emphasizes self-evolution and developmental behaviors alongside interactive engagement, offering a more realistic reflection of pet companionship. It features diverse tasks such as intelligent scheduling, memory-based dialogues, and psychological conversations, with over 7,500 interaction instances designed to simulate pet behaviors. Evaluation of 28 LLMs reveals significant performance variations linked to model size and inherent capabilities, underscoring the need for specialized optimization in this domain. PET-BENCH serves as a foundational resource for benchmarking pet-related LLM abilities and advancing emotionally immersive human-pet interactions.

Due to the high coverage and low cost nature of trajectory data, an increasing number of works have utilized trajectory data to infer maps. Nevertheless, limited by the sparse trajectories in some areas and intermingled trajectories on parallel roads, the existing inferring methods still face a high missed detection rate of the roads. In view of that, we propose a Complementarity and Disparity-aware Map Inference Framework, called CDMap, consisting of grid dual feature extraction, contextual road difference-embedded grid representation, dual feature complementary network-based road topology prediction and parallel roads disparity-enhanced model optimization. To improve the prediction accuracy of the roads in areas with sparse trajectories, we extract point-wise features and segment-wise features separately for the grids, then design a dual feature complementary network to adaptively model the importance of both types of features in different road scenarios. Further, to proliferate the detection accuracy of parallel roads, we incorporate the contextual roads' differences between parallel roads into grid representations, then put forward a parallel roads disparity-enhanced model optimization strategy. Extensive comparative experiments conducted on three real-world datasets demonstrate the superiority of CDMap over the state-of-the-art methods, especially by achieving the most significant reduction in missed detection rate (30.23%) on the trajectory data collected from DidiChuxing platform.

Multimodal Large Language Models (MLLMs) have achieved remarkable success in vision understanding, reasoning, and interaction. However, the inference computation and memory increase progressively with the generation of output tokens during decoding, directly affecting the efficacy of MLLMs. Existing methods attempt to reduce the vision context redundancy to achieve efficient MLLMs. Unfortunately, the efficiency benefits of the vision context reduction in the prefill stage gradually diminish during the decoding stage. To address this problem, we proposed a dynamic vision-language context sparsification framework Dynamic-LLaVA, which dynamically reduces the redundancy of vision context in the prefill stage and decreases the memory and computation overhead of the generated language context during decoding. Dynamic-LLaVA designs a tailored sparsification inference scheme for different inference modes, i.e., prefill, decoding with and without KV cache, to achieve efficient inference of MLLMs. In practice, Dynamic-LLaVA can reduce computation consumption by $\sim$75\% in the prefill stage. Meanwhile, throughout the entire generation process of MLLMs, Dynamic-LLaVA reduces the $\sim$50\% computation consumption under decoding without KV cache, while saving $\sim$50\% GPU memory overhead when decoding with KV cache, due to the vision-language context sparsification. Extensive experiments also demonstrate that Dynamic-LLaVA achieves efficient inference for MLLMs with negligible understanding and generation ability degradation or even performance gains compared to the full-context inference baselines. Code is available at https://github.com/Osilly/dynamic_llava.

With the rapid advancement of Social Networking Services (SNS), the need for intelligent and efficient interaction within diverse platforms has become more crucial. Large Language Models (LLMs) play an important role in SNS as they possess the potential to revolutionize user experience, content generation, and communication dynamics. However, recent studies focus on isolated SNS tasks rather than a comprehensive evaluation. In this paper, we introduce SNS-Bench, specially constructed for assessing the abilities of large language models from different Social Networking Services, with a wide range of SNS-related information. SNS-Bench encompasses 8 different tasks such as note classification, query content relevance, and highlight words generation in comments. Finally, 6,658 questions of social media text, including subjective questions, single-choice, and multiple-choice questions, are concluded in SNS-Bench. Further, we evaluate the performance of over 25+ current diverse LLMs on our SNS-Bench. Models with different sizes exhibit performance variations, yet adhere to the scaling law. Moreover, we hope provide more insights to revolutionize the techniques of social network services with LLMs.

Recent advances in Multimodal Large Language Models (MLLMs) have significantly improved 2D visual understanding, prompting interest in their application to complex 3D reasoning tasks. However, it remains unclear whether these models can effectively capture the detailed spatial information required for robust real-world performance, especially cross-view consistency, a key requirement for accurate 3D reasoning. Considering this issue, we introduce Viewpoint Learning, a task designed to evaluate and improve the spatial reasoning capabilities of MLLMs. We present the Viewpoint-100K dataset, consisting of 100K object-centric image pairs with diverse viewpoints and corresponding question-answer pairs. Our approach employs a two-stage fine-tuning strategy: first, foundational knowledge is injected to the baseline MLLM via Supervised Fine-Tuning (SFT) on Viewpoint-100K, resulting in significant improvements across multiple tasks; second, generalization is enhanced through Reinforcement Learning using the Group Relative Policy Optimization (GRPO) algorithm on a broader set of questions. Additionally, we introduce a hybrid cold-start initialization method designed to simultaneously learn viewpoint representations and maintain coherent reasoning thinking. Experimental results show that our approach significantly activates the spatial reasoning ability of MLLM, improving performance on both in-domain and out-of-domain reasoning tasks. Our findings highlight the value of developing foundational spatial skills in MLLMs, supporting future progress in robotics, autonomous systems, and 3D scene understanding.

User-generated content (UGC) communities, especially those featuring multimodal content, improve user experiences by integrating visual and textual information into results (or items).The challenge of improving user experiences in complex systems with search and recommendation (S&R) services has drawn significant attention from both academia and industry these years.However, the lack of high-quality datasets has limited the research progress on multimodal S&R.To address the growing need for developing better S&R services, we present a novel multimodal information retrieval dataset in this paper, namely Qilin.The dataset is collected from Xiaohongshu, a popular social platform with over 300 million monthly active users and an average search penetration rate of over 70%.In contrast to existing datasets, Qilin offers a comprehensive collection of user sessions with heterogeneous results like image-text notes, video notes, commercial notes, and direct answers, facilitating the development of advanced multimodal neural retrieval models across diverse task settings.To better model user satisfaction and support the analysis of heterogeneous user behaviors, we also collect extensive APP-level contextual signals and genuine user feedback.Notably, Qilin contains user-favored answers and their referred results for search requests triggering the

Social chatbots have become essential companions in daily scenarios ranging from emotional support to personal interaction. However, conventional chatbots with passive response mechanisms usually rely on users to initiate or sustain dialogues by bringing up new topics, resulting in diminished engagement and shortened dialogue duration. In this paper, we present PaRT, a novel framework enabling context-aware proactive dialogues for social chatbots through personalized real-time retrieval and generation. Specifically, PaRT first integrates user profiles and dialogue context into a large language model (LLM), which is initially prompted to refine user queries and recognize underlying intents for the upcoming conversation. Guided by refined intents, the LLM generates personalized dialogue topics as targeted queries to retrieve relevant passages from RedNote. Finally, we prompt LLMs with summarized passages to generate knowledge-grounded and engagement-optimized responses. Our approach has been running stably in a real-world production environment for more than 30 days, achieving a 21.77% improvement in the average duration of dialogues.

Temporal (ST) prediction is crucial for making informed decisions in urban location-based applications like ride-sharing. However, existing ST models often require region partition as a prerequisite, resulting in two main pitfalls. Firstly, location-based services necessitate ad-hoc regions for various purposes, requiring multiple ST models with varying scales and zones, which can be costly to support. Secondly, different ST models may produce conflicting outputs, resulting in confusing predictions. In this paper, we propose One4All-ST, a framework that can conduct ST prediction for arbitrary modifiable areal units using only one model. To reduce the cost of getting multi-scale predictions, we design an ST network with hierarchical spatial modeling and scale normalization modules to efficiently and equally learn multi-scale representations. To address prediction inconsistencies across scales, we propose a dynamic programming scheme to solve the formulated optimal combination problem, minimizing predicted error through theoretical analysis. Besides, we suggest using an extended quad-tree to index the optimal combinations for quick response to arbitrary modifiable areal units in practical online scenarios. Extensive experiments on two real-world datasets verify the efficiency and effectiveness of One4All-ST in ST prediction for arbitrary modifiable areal units. The source codes and data of this work are available at https://github.com/uctb/One4All-ST.

Along with the increase in strict regulation of drunk driving behavior in China, the demands for designated driving services have risen in popularity. In the absence of specialized cycling map for the designated drivers who use foldable electric bicycles, they tend to take a detour or are lost on the way to the car owners’ appointed parking places. With gradual popularization of chauffeur services, cycling trajectories generated by designated drivers almost spread all over the city. It provides a chance for inferring the cycling map dedicated to the designated drivers. However, to infer an accurate map using trajectories faces severe challenges stemming from random cycling behaviors of designated drivers, including (i) trajectories contain a lot of noises and incomplete segments, (ii) turning trajectories at minor intersections are very sparse and (iii) trajectories on the roads of distinct shapes are obviously different. To address the above challenges, we propose a three-phase map inference framework, called SAMI, consisting of trajectory refinement, intersection pinpointing, and road curve interlinking. Specifically, cycling behavioral differences from neighbor regions are incorporated into intersection identification process to ensure obtaining high detection precision even when trajectory data is sparse. Further, shape-aware based centerline fitting strategy is put forward to guarantee that inferred road curves are consistent with real road shape as possible. Finally, extensive comparative experiments on two real data sets demonstrate that SAMI significantly outperforms state-of-the-art methods by 13.31% in F1-score of map inference and by 44.88% in recall rate of minor intersection detection.

MAUP (modifiable areal unit problem) is a fundamental problem for spatial data management and analysis. As an instantiation of MAUP in online transportation platforms, region generation (i.e., specifying the areal unit for service operations) is the first and vital step for supporting spatiotemporal transportation services such as ride-sharing and freight transport. Most existing region generation methods are manually specified (e.g., fixed-size grids), suffering from poor spatial semantic meaning and inflexibility to meet service operation requirements. In this paper, we propose RegionGen, a data-driven region generation framework that can specify regions with key characteristics (e.g., good spatial semantic meaning and predictability) by modeling region generation as a multi-objective optimization problem. First, to obtain good spatial semantic meaning, RegionGen segments the whole city into atomic spatial elements based on road networks and obstacles (e.g., rivers). Then, it clusters the atomic spatial elements into regions by maximizing various operation characteristics, which is formulated as a multi-objective optimization problem. For this optimization problem, we propose a multi-objective co-optimization algorithm. Extensive experiments verify that RegionGen can generate more suitable regions than traditional methods for spatiotemporal service management.

In the ride-hailing business, compensation is mostly used to motivate consumers to place more orders and grow the market scale. However, most of the previous studies focus on car-hailing services. Few works investigate localized smart transportation innovations, such as intra-city freight logistics and designated driving. In addition, satisfying consumer fairness and improving consumer surplus, with the objective of maximizing revenue, are also important. In this paper, we propose a consumer compensation system, where a transfer learning enhanced uplift modeling is designed to measure the elasticity, and a model predictive control based optimization is formulated to control the budget accurately. Our implementation is effective and can keep the online environment lightweight. The proposed system has been deployed in the production environment of the real-world ride-hailing platform for 300 days, which outperforms the expert strategy by using 0.5% less subsidy and achieving 14.4% more revenue.

Transportation demand forecasting is a critical precondition of optimal online transportation dispatch, which will greatly reduce drivers’ wasted mileage and customers’ waiting time, contributing to economic and environmental sustainability. Though various methods have been developed, the core spatio-temporal complexity remains challenging from three perspectives: (1) Compound spatial relationships. According to our empirical analysis, these relationships widely exist. Previous studies focus on capturing different spatial relationships using multi-homogeneous graphs. However, the information flow across various spatial relationships is not modeled explicitly. (2) Heterogeneity in spatial relationships. A region’s neighbors under the same spatial relationship may have different weights for this region. Meanwhile, different relationships may also weigh differently. (3) Synchronicity between compound spatial relationships and temporal relationships. Previous research considers synchronous influences from spatial and temporal relationships in a homogeneous fashion while compound spatial relationships are not captured for this synchronicity. To address the aforementioned perspectives, we propose the S patio- T emporal H eterogeneous graph A ttention N etwork (STHAN), where the key intuition is capturing the compound spatial relationships via meta-paths explicitly. We first construct a spatio-temporal heterogeneous graph including multiple spatial relationships and temporal relationships and use meta-paths to depict compound spatial relationships. To capture the heterogeneity, we use hierarchical attention, which contains node level attention and meta-path level attention. The synchronicity between temporal relationships and spatial relationships, including compound ones, is modeled in meta-path-level attention. Our framework outperforms state-of-the-art models by reducing 6.58%, 4.57%, and 4.20% of WMAPE in experiments on three real-world datasets, respectively.

With the booming of the Internet finance and e-payment business, telecom and online fraud has become a serious problem which grows rapidly. In China, 351 billion RMB (approximately 0.3% of China's GDP) was lost in 2018 due to telecommunication and online fraud, influencing tens of millions of individual customers. Anti-fraud algorithms have been widely adopted by major Internet finance companies to detect and block transactions induced by scam. However, due to limited contextual information, most systems would probably mistakenly block the normal transactions, leading to poor user experience. On the other hand, if the transactions induced by scam are detected yet not fully explained to the users, the users will continue to pay, suffering from direct financial losses.

To reduce the customer service pressure of small and medium-sized enterprises, we propose an intelligent cloud customer service system, called IntelliTag. Unlike traditional customer service, a cloud service based system has difficulty in collecting user personal information. Therefore, we add a tag recommendation function to quickly capture the user's question intent by clicking on the tags. Specifically, IntelliTag is elaborately designed with the consideration of the following three aspects. First, how to mine high-quality tags is a challenging problem. Second, in the tag recommendation tasks, we have multifarious data types and relations that are used to build a sequential recommendation model. Finally, system implementation and deployment also need to be carefully designed to satisfy online service requirements. In this paper, we show the details of data construction, model designs, system implementation and deployment, and the empirical results compared with several state-of-the-art methods. Nowadays, our IntelliTag has already supported hundreds of thousands of enterprises and millions of users in our industrial production environment.

With the rapid growth of B2B (Business-to-Business), how to efficiently respond to various customer questions is becoming an important issue. In this scenario, customer questions always involve many aspects of the products, so there are usually multiple customer service agents to response respectively. To improve efficiency, we propose a human-machine cooperation solution called ServiceGroup, where relevant agents and customers are invited into the same group, and the system can provide a series of intelligent functions, including question notification, question recommendation and knowledge extraction. With the assistance of our developed ServiceGroup, the response rate within 15 minutes is improved twice. Until now, our ServiceGroup has already supported thousands of enterprises by means of millions of groups in instant messaging softwares.

With the explosive growth of e-commerce and the booming of e-payment, detecting online transaction fraud in real time has become increasingly important to Fintech business. To tackle this problem, we introduce the TitAnt, a transaction fraud detection system deployed in Ant Financial, one of the largest Fintech companies in the world. The system is able to predict online real-time transaction fraud in mere milliseconds. We present the problem definition, feature extraction, detection methods, implementation and deployment of the system, as well as empirical effectiveness. Extensive experiments have been conducted on large real-world transaction data to show the effectiveness and the efficiency of the proposed system.

We propose cw2vec, a novel method for learning Chinese word embeddings. It is based on our observation that exploiting stroke-level information is crucial for improving the learning of Chinese word embeddings. Specifically, we design a minimalist approach to exploit such features, by using stroke n-grams, which capture semantic and morphological level information of Chinese words. Through qualitative analysis, we demonstrate that our model is able to extract semantic information that cannot be captured by existing methods. Empirical results on the word similarity, word analogy, text classification and named entity recognition tasks show that the proposed approach consistently outperforms state-of-the-art approaches such as word-based word2vec and GloVe, character-based CWE, component-based JWE and pixel-based GWE.

We present a novel approach to learning word embeddings by exploring subword information (character n-gram, root/affix and inflections) and capturing the structural information of their context with convolutional feature learning. Specifically, we introduce a convolutional neural network architecture that allows us to measure structural information of context words and incorporate subword features conveying semantic, syntactic and morphological information related to the words. To assess the effectiveness of our model, we conduct extensive experiments on the standard word similarity and word analogy tasks. We showed improvements over existing state-of-the-art methods for learning word embeddings, including skipgram, GloVe, char n-gram and DSSM.

In this work, we propose to actively detect implicit fraudulent transactions. A novel machine learning method is introduced to distinguish anomalous electronic transactions based on the historical records. The transferor will be alerted during the on-going payment when the fraud probability is recognized as large enough. Compared with elaborative rule-based approaches, our model is much more effective in fraud detection.

In this paper, we propose a novel model for learning graph representations, which generates a low-dimensional vector representation for each vertex by capturing the graph structural information. Different from other previous research efforts, we adopt a random surfing model to capture graph structural information directly, instead of using the sampling-based method for generating linear sequences proposed by Perozzi et al. (2014). The advantages of our approach will be illustrated from both theorical and empirical perspectives. We also give a new perspective for the matrix factorization method proposed by Levy and Goldberg (2014), in which the pointwise mutual information (PMI) matrix is considered as an analytical solution to the objective function of the skip-gram model with negative sampling proposed by Mikolov et al. (2013). Unlike their approach which involves the use of the SVD for finding the low-dimensitonal projections from the PMI matrix, however, the stacked denoising autoencoder is introduced in our model to extract complex features and model non-linearities. To demonstrate the effectiveness of our model, we conduct experiments on clustering and visualization tasks, employing the learned vertex representations as features. Empirical results on datasets of varying sizes show that our model outperforms other stat-of-the-art models in such tasks.

In this paper, we present {GraRep}, a novel model for learning vertex representations of weighted graphs. This model learns low dimensional vectors to represent vertices appearing in a graph and, unlike existing work, integrates global structural information of the graph into the learning process. We also formally analyze the connections between our work and several previous research efforts, including the DeepWalk model of Perozzi et al. as well as the skip-gram model with negative sampling of Mikolov et al.