Yinghao Tang, Xueding Liu, Boyuan Zhang, Tingfeng Lan, Yupeng Xie, Jiale Lao, Yiyao Wang, Haoxuan Li, Tingting Gao, Bo Pan, Luoxuan Weng, Xiuqi Huang, Minfeng Zhu, Yingchaojie Feng, Yuyu Luo, Wei Chen. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Large language models (LLMs) have achieved remarkable performance across a wide range of natural language tasks. Understanding how LLMs internally represent knowledge remains a significant challenge. Despite Sparse Autoencoders (SAEs) have emerged as a promising technique for extracting interpretable features from LLMs, SAE features do not inherently align with human-understandable concepts, making their interpretation cumbersome and labor-intensive. To bridge the gap between SAE features and human concepts, we present ConceptViz, a visual analytics system designed for exploring concepts in LLMs. ConceptViz implements a novel Identification ⇒ Interpretation ⇒Validation pipeline, enabling users to query SAEs using concepts of interest, interactively explore concept-to-feature alignments, and validate the correspondences through model behavior verification. We demonstrate the effectiveness of ConceptViz through two usage scenarios and a user study. Our results show that ConceptViz enhances interpretability research by streamlining the discovery and validation of meaningful concept representations in LLMs, ultimately aiding researchers in building more accurate mental models of LLM features. Our code and user guide are publicly available at https://github.com/Happy-Hippo209/conceptViz.

The security of individual privacy is paramount for trajectory publication, while preserving trajectory utility is also essential to serve analysis tasks such as urban planning and transportation development. To assess and maintain trajectory utility, existing studies consider geographic context. However, innate semantic characteristics of trajectories (e.g., origin, destination, stay point, path) have been overlooked, which prevents data owners from specifying task-specific utility measurements and, consequently, from achieving a delicate balance between privacy and utility. This paper proposes an interactive trajectory publishing approach driven by flexible utility considerations, which processes trajectory points according to their semantics to fulfill diverse utility requirements. Concretely, we decouple trajectories into origin-destination (OD) and path components: ODs are generalized into regions to satisfy $k$k-anonymity, and paths are sanitized within each OD group using road-network-aware differential privacy under predefined privacy constraints. We also develop a visual interface to support exploration and comprehension of privacy-preserving solutions, through which we incorporate human knowledge into the privacy scheme. Experiments on real-world urban datasets demonstrate the effectiveness of our approach.

Business intelligence (BI) transforms large volumes of data within modern organizations into actionable insights for informed decision-making. Recently, large language model (LLM)-based agents have streamlined the BI workflow by automatically performing task planning, reasoning, and actions in executable environments based on natural language (NL) queries. However, existing approaches primarily focus on individual BI tasks such as NL2SQL and NL2VIS. The fragmentation of tasks across different data roles and tools lead to inefficiencies and potential errors due to the iterative and collaborative nature of BI. In this paper, we introduce DataLab, a unified BI platform that integrates a one-stop LLM-based agent framework with an augmented computational notebook interface. DataLab supports various BI tasks for different data roles in data preparation, analysis, and visualization by seamlessly combining LLM assistance with user customization within a single environment. To achieve this unification, we design a domain knowledge incorporation module tailored for enterprise-specific BI tasks, an inter-agent communication mechanism to facilitate information sharing across the BI workflow, and a cell-based context management strategy to enhance context utilization efficiency in BI notebooks. Extensive experiments demonstrate that DataLab achieves state-of-the-art performance on various BI tasks across popular research benchmarks. Moreover, DataLab maintains high effectiveness and efficiency on real-world datasets from Tencent, achieving up to a 58.58% increase in accuracy and a 61.65 % reduction in token cost on enterprise-specific BI tasks.

Machine learning (ML) has become a cornerstone of key applications at ByteDance. As model complexity and data volumes surge, data management for large-scale ML workloads faces substantial challenges, particularly with recent advances in large recommendation models (LRMs) and large multimodal models (LMMs). Traditional approaches exhibit limitations in storage efficiency, metadata scalability, update mechanisms, and integration with ML frameworks. To address these challenges, we propose Magnus, a holistic data management system built upon Apache Iceberg. Magnus integrates innovative optimizations across resource-efficient storage formats optimized for large wide tables and multimodal data, built-in support for vector and inverted indexes to accelerate data retrieval, scalable metadata planning with Git-like branching and tagging capabilities, and high-performance update/upsert based on lightweight merge-on-read (MOR) strategies. Additionally, Magnus provides native support and specialized enhancement for LRM and LMM training workloads. Experimental results demonstrate significant performance gains in real-world ML scenarios. Magnus has been deployed at ByteDance for over five years, enabling robust and efficient data infrastructure for large-scale ML workloads.

Graph visualization techniques provide a foundational framework for offering comprehensive overviews and insights into cloud computing systems, facilitating efficient management and ensuring their availability and reliability. Despite the enhanced computational and storage capabilities of larger-scale cloud computing architectures, they introduce significant challenges to traditional graph-based visualization due to issues of hierarchical heterogeneity, scalability, and data incompleteness. This paper proposes a novel abstraction approach to visualize large multi-layer networks. Our method leverages graphons, a probabilistic representation of network layers, to encompass three core steps: an inner-layer summary to identify stable and volatile substructures, an inter-layer mixup for aligning heterogeneous network layers, and a context-aware multi-layer joint sampling technique aimed at reducing network scale while retaining essential topological characteristics. By abstracting complex network data into manageable weighted graphs, with each graph depicting a distinct network layer, our approach renders these intricate systems accessible on standard computing hardware. We validate our methodology through case studies, quantitative experiments and expert evaluations, demonstrating its effectiveness in managing large multi-layer networks, as well as its applicability to broader network types such as transportation and social networks.

Public cloud providers embrace multi-tenancy as a strategy to enhance the utilization and efficiency of resources. However, co-located virtual machines (VMs) suffer from qualityof-service (QoS) degradation caused by shared resource interference. Existing solutions for predicting QoS degradation often rely on the assumption of online access to application-level information. However, in a production environment, this assumption proves invalid as the VMs are black boxes to the providers. This intrinsic characteristic of the IaaS cloud necessitates the prediction model to generalize to unfamiliar applications and imposes specific criteria on the monitorable metrics.To meet the black-box scenario under Infrastructure as a Service (IaaS) cloud computing, we present a novel framework, FEDGE, that can predict interference-aware QoS (IA-QoS) of co-located VMs using only low-level monitorable metrics before migration. Specifically, FEDGE utilizes a stochastic gates layer to select the most informative features from the high-dimensional resource and hardware metrics, which helps to reduce the monitoring overhead. Furthermore, we design a multi-domain MMD-based adversarial denoising autoencoder to regularize the learned hidden representations and prevent over-fitting on the source domains. Next, we employ a multi-layer perceptron (MLP) to accurately predict complex QoS degradation using the learned representations with domain generalization. Experimental results demonstrate that FEDGE outperforms other state-of-the-art methods in terms of both generalizability and effectiveness.

Urban public nuisance events, like garbage exposure, illegal parking, facilities damage, and etc., impair the quality of life for city residents. Predicting and controlling these nuisances is crucial but complicated due to their ties to subjective and psychological factors. In this study, we reveal a significant correlation between such nuisances and meteorological indicators, influenced by the impact of climate on people's psychological states. We employ meteorology predictions that are integrated in Hawkes processes to enhance the accuracy of predicting the category and timing of these nuisances. To this end, we propose Spatial-Temporal Two-Tower Transformer (ST-T3), which simultaneously considers spatial data and further improves the prediction accuracy. Evaluated by about three-year data from both downtown and suburban Shanghai, our method outperforms both traditional and advanced prediction systems. We share a portion of the de-identified dataset for open research.

Visible light communication (VLC), which uses an irregular visible band to offer ultra-high bandwidth, ultra-high data rates, lower latency, low energy consumption and reduced implementation costs, is one of the most promising complementary wireless technologies for Sixth Generation Mobile Communications(6G). With the deployment of massive access points for bidirectional VLC network, the problem of user service quality is also brought correspondingly, such as how to increase user Quality of Experience (QoE) and reduce user access delay. In this paper, a space-wavelength-time division multiple access (SWT-MA) scheme is proposed based on Tri-color Light Emitting Diode(LEDs). The principle of SWT-MA scheme is descripted. The metrics of the VLC based on the SWT-MA scheme is analyzed in throughput, access delay in detail respectively. And finally, the experimental measurements of BER for each color signals are illustrated. The access scheme proposed in this paper solves the above problems by improving the data transmission rate through space division multiplexing and wavelength division multiplexing, and increasing the system capacity through time division multiplexing. The feasibility of this scheme is verified by theoretical analysis, simulation and experiment. We showed that the proposed access scheme is suitable for large-scale bidirectional VLC access networks and can be used for 6G.

At ByteDance, where we execute over a million Spark jobs and handle 500PB of shuffled data daily, ensuring resource efficiency is paramount for cost savings. However, achieving optimization of resource efficiency in large-scale production environments poses significant challenges. Drawing from our practical experiences, we have identified three key issues critical to addressing resource efficiency in real-world production settings: 1 slow I/Os leading to excessive CPU and memory idleness, 2 coarse-grained resource control causing wastage, and 3 sub-optimal job configurations resulting in low utilization. To tackle these issues, we propose a resource efficiency governance framework for Spark workloads. Specifically, 1 we devise the multi-mechanism shuffle services, including Enhanced External Shuffle Service (ESS) and Cloud Shuffle Service (CSS), where CSS employs a push-based approach to enhance I/O efficiency through sequential reading. 2 We modify the Spark configuration parameter protocol, allowing for fine-grained resource control by introducing several new parameters such as milliCores and memoryBurst, as well as supporting operators with additional spill modes. 3 We design a two-stage configuration autotuning method, comprising rule-based and algorithm-based tuning, providing more reliable Spark configuration optimizations. By deploying these techniques on millions of Spark jobs in production over the last two years, we have achieved over 22% CPU utilization increase, 5% memory utilization increase, and 10% shuffle block time ratio decrease, effectively saving millions of CPU cores and petabytes of memory daily.

Trend forecasting is vital to optimize the workload performance. It becomes even more urgent with an increasing number of applications and database configurations. However, DBAs mainly target at historical workloads and may give suboptimal configuration advice when the workload trends have changed. Although there are some studies on trend forecasting, they have several limitations. First, they mainly predict the changes of query numbers, which do not combine other critical factors (e.g., disk utilization) and cannot fully reflect the future workload trends. Besides, there are numerous queries in the workloads and exact clustering algorithms like K-means cannot effectively merge similar queries which contain noises like time shifts. Second, basic machine learning models like RNN may have relatively low prediction accuracy on complex workloads (e.g., no cycles but random bursts). Third, real-world workloads may have diverse patterns, while previous models cannot efficiently and reliably predict for all the different workload patterns.To address these challenges, we propose a trend forecasting system (DBAugur) that utilizes adversarial neural networks to predict the trends of different workloads. First, DBAugur collects the important features (e.g., queries, resource metrics) to characterize workloads, and reduces the number of involved queries by separately merging similar queries based on the SQL semantics and trend patterns. Second, DBAugur utilizes Generative Adversarial Networks (GANs) to capture the latent patterns, correlations between different metrics, and occasional bursts within the complicated and time-varying workloads. Moreover, we further propose a time-sensitive ensemble algorithm that takes advantage of various machine learning models (e.g., generative models, convolutional models, feed-forward models) to accommodate the various workload patterns. The experimental results show that DBAugur outperformed state-of-the-art methods on various real-world workloads.

Social networks are widely recognized as highly effective information sources for social recommendation services. However, previous social recommendation methods assumed that a user’s preference factor and social trust factor shared a common latent feature space. Additionally, few studies have explored the incorporation of social information into the item domain for recommendations. To address these gaps, we propose ExpoEv, a deep collaborative filtering recommendation model that integrates social exposure based on feature evolution for social recommendation services. Specifically, we propose a social exposure module for both user and item domains that considers the number of items that a user’s social friends interact with. Furthermore, we introduce a feature evolution component that enables the incorporation of social exposure information with social trust and attribute factors in the context of social recommendation services. Experiments demonstrate the effectiveness of our model in the quality of recommendation service.

In large-scale systems, due to system complexity and demand volatility, diverse and dynamic workloads make accurate predictions difficult. In this work, we address an online interval prediction problem (OnPred-Int) and adopt ensemble learning to solve it. We depict that the ensemble learning for OnPred-Int is a dynamic deterministic Markov Decision Process (Dd-MDP) and convert it into a stateful online learning task. Then we propose IPOC, a lightweight and flexible model able to produce effective confidence intervals, adapting the dynamics of real-time workload streams. At each time, IPOC selects a target model and executes chasing for it by a designed chasing oracle, during which process IPOC produces accurate confidence intervals. The effectiveness of IPOCis theoretically validated through sublinear regret analysis and satisfaction of confidence interval requirements. Besides, we conduct extensive experiments on 4 real-world datasets comparing with 19 baselines. To the best of our knowledge, we are the first to apply the frontier theory of online learning to time series prediction tasks.

A major challenge confronting today’s distributed metadata management schemes is how to meet the dynamic requirements of various applications through effectively mapping and migrating metadata nodes to different metadata servers (MDS’s). Most of the existing works dynamically reallocate nodes to different servers adopting history-based coarse-grained methods, failing to make a timely and efficient update on the distribution of nodes. In this paper, we present the first deep reinforcement learning-leveraged distributed metadata management scheme, AdaM, to address the aforementioned dilemma. AdaM is an adaptive fine-grained metadata management scheme that trains an actor-critic network to migrate “hot” metadata nodes to different MDS’s based on its observations of the current “states” (i.e., access pattern, the structure of namespace tree and current distribution of nodes on MDS’s). Adaptive to varying access patterns, AdaM can automatically migrate hot metadata nodes among servers to keep load balancing while maintaining metadata locality. Besides, we propose a self-adaptive metadata cache policy, which dynamically combines the two strategies of managing caches on the server side and the client side to gain better query performance. Last but not least, we design a distributed metadata processing 2PC Protocol called MST-based 2PC to ensure data consistency. Experiments on a real-world dataset demonstrate the superiority of our proposed method over other schemes.

Nowadays, Database-as-a-Service (DBaaS) has become more and more popular among users as it can largely reduce the complexity of managing databases and applications. Considering the increasing complexity of different applications, system management and automation such as self-provisioning and performance tuning can be challenging. To better achieve autonomous optimization of the system, the ability to predict future workload patterns is of great essence. In this paper, we propose a novel lightweight probabilistic workload forecasting framework (LIGHTPRO) that is easy to train and robust, to help the system predict future workload patterns, leveraging multi-head attention mechanism and convolution operations. Experiments on real-world query traces demonstrate the superiority of LIGHTPRO in reducing training time and capturing both long-term and short-term temporal patterns of the workload compared with other baselines.

The development of global positioning system stimulates the popularity of location-based social network (LBSN) services. With a large volume of data containing locations, texts, check-in information, and social relationships, spatial keyword queries in LBSNs have become increasingly complex. In this paper, we identify and solve the Social-based Time-aware Spatial Keyword Query (STSKQ) that returns the top-k objects by considering geo-spatial score, keywords similarity, visiting time score, and social relationship effect. To tackle STSKQ, we propose a two-layer hybrid index structure called Network Embedding Time-aware R-tree (NETR-Tree). In the user layer, we exploit the network embedding strategy to measure the relationship effect in users' relationship network. In the location layer, we build a Time-aware R-tree (TR-tree) considered spatial objects' spatiotemporal check-in information, and present a corresponding query processing algorithm. Finally, extensive experiments on two different real-life LBSNs demonstrate the effectiveness and efficiency of our methods, compared with existing state-of-the-art methods.