WWW 2025 Papers

Graph neural networks have been demonstrated as a powerful paradigm for effectively learning graph-structured data for downstream task analysis. Current leading graph models require a large number of labeled samples for training, which unavoidably leads to overfitting in few-shot scenarios. Recent research has sought to alleviate this issue by simultaneously leveraging graph learning and meta-learning paradigms. However, these graph meta-learning models assume the availability of numerous meta-training tasks to learn transferable meta-knowledge. Such an assumption may not be feasible in the real world due to the difficulty of constructing tasks and the substantial costs involved. Therefore, we propose a SiMple yet effectIve approach for graph few-shot Learning with fEwer tasks, named SMILE. We introduce a dual-level mixup strategy, encompassing both within-task and across-task mixup, to simultaneously enrich the available nodes and tasks in meta-learning. Moreover, we explicitly leverage the prior information provided by the node degrees in the graph to encode expressive node representations. Theoretically, we demonstrate that SMILE can enhance the model generalization ability. Empirically, SMILE consistently outperforms other competitive models by a large margin across all evaluated datasets with in-domain and cross-domain settings. Our anonymous code can be found here.

Retrieval-Augmented Generation (RAG) systems have become acrucial tool to augment large language models (LLMs) with external knowledge for better task performance. However, existing traditional RAG methods inject knowledge directly in the context, resulting in several limitations. First, these methods highly rely on the in-context learning capability of LLMs, which often leads to excessively long contexts. This is inefficient due to the quadratic complexity of self-attention, leading to significant increases in inference time. Second, the extended context and the nature of self-attention can cause the LLMs to lose important information in the context, thereby degrading the original capabilities of LLMs. Furthermore, the effectiveness of knowledge injection is perturbed by the permutation of knowledge within the extended context, reducing the robustness of existing RAG methods. To tackle the above problems, we propose DecoupledRAG, a method that decouples external knowledge from the context within the RAG framework. Specifically, we introduce a cross-attention based method that injects retrieved knowledge directly to the inference process of LLM on the fly, without modifying its parameters or the input context. The external knowledge could be utilized robustly in a permutation-independent manner. To the best of our knowledge, this is the first work that explore how to utilize cross-attention to inject knowledge with low training cost in decoder-only LLM era. By leveraging cross-attention operation, DecoupledRAG enables seamless knowledge aggregation without creating extended context. Experimental results demonstrate that our method achieves high efficiency while maintaining strong performance, which indicates that RAG frameworks have the potential to benefit further from more knowledge.

With the rise of generative large language models (LLMs) like LLaMA and ChatGPT, these models have significantly transformed daily life and work by providing advanced insights. However, as jailbreak attacks continue to circumvent built-in safety mechanisms, exploiting carefully crafted scenarios or tokens, the safety risks of LLMs have come into focus. While numerous defense strategies—such as prompt detection, modification, and model fine-tuning—have been proposed to counter these attacks, a critical question arises: do these defenses compromise the utility and usability of LLMs for legitimate users? Existing research predominantly focuses on the effectiveness of defense strategies without thoroughly examining their impact on performance, leaving a gap in understanding the trade-offs between LLM safety and performance. Our research addresses this gap by conducting a comprehensive study on the utility degradation, safety elevation, and exaggerated-safety escalation of LLMs with jailbreak defense strategies. We propose _**USEBench**_, a novel benchmark designed to evaluate these aspects, along with _**USEIndex**_, a comprehensive metric for assessing overall model performance. Through experiments on seven state-of-the-art LLMs, we found that mainstream jailbreak defenses fail to ensure both safety and performance simultaneously. Although model-finetuning performs the best overall, their effectiveness varies across LLMs. Furthermore, vertical comparisons reveal that developers commonly prioritize performance over safety when iterating or fine-tuning their LLMs.

Canvas fingerprinting has become one of the most effective techniques for tracking users online, allowing websites to identify and track visitors without their consent. In this paper, we investigate four primary defense techniques designed to counter canvas fingerprinting, systematically analyzing their adoption across 18 browser extensions in Chrome and Firefox, as well as built-in protections from five major browsers: Chrome, Firefox, Brave, Tor, and Safari. Our analysis reveals significant disparities in the implementation and effectiveness of these defenses, with randomization-based techniques being the most widely adopted, particularly across nine extensions and in the privacy-focused browser, Brave. Despite their sophistication, we demonstrate successful attacks on all these randomization mechanisms, revealing that their supposed non-deterministic behavior can, in fact, be predicted and exploited. In summary, we demonstrate that, unfortunately, no fully deployable defense against canvas fingerprinting attacks exists currently. We conclude by proposing recommendations to strengthen existing defenses and enhance their resistance to future attacks.

Retrieval-Augmented Generation (RAG) has been shown to improve knowledge capabilities and alleviate the hallucination problem of LLMs. The Web is a major source of external knowledge used in RAG systems, and many commercial systems such as ChatGPT and Perplexity have used Web search engines as their major retrieval systems. Typically, such RAG systems retrieve search results, download HTML sources of the results, and then extract plain texts from the HTML sources. Plain text documents or chunks are fed into the LLMs to augment the generation. However, much of the structural and semantic information inherent in HTML, such as headings and table structures, is lost during this plain-text-based RAG process. To alleviate this problem, we propose HtmlRAG, which uses HTML instead of plain text as the format of retrieved knowledge in RAG. We believe HTML is better than plain text in modeling knowledge in external documents, and most LLMs possess robust capacities to understand HTML. However, utilizing HTML presents new challenges. HTML contains additional content such as tags, JavaScript, and CSS specifications, which bring extra input tokens and noise to the RAG system. To address this issue, we propose HTML cleaning, compression, and pruning strategies, to shorten the HTML while minimizing the loss of information. Specifically, we design a two-step block-tree-based pruning method that prunes useless HTML blocks and keeps only the relevant part of the HTML. Experiments on six QA datasets confirm the superiority of using HTML in RAG systems.

As multimedia content continues to grow on the Web, the integration of visual and textual data has become a crucial challenge for Web applications, particularly in recommendation systems. Large Vision Language Models (LVLMs) have demonstrated considerable potential in addressing this challenge across various tasks that require such multimodal integration. However, their application in multimodal sequential recommendation (MSR) has not been extensively studied, despite their potential to significantly enhance the performance of web-based multimodal recommendations. To bridge this gap, we introduce MSRBench, the first comprehensive benchmark designed to systematically evaluate different LVLM integration strategies in web-based recommendation scenarios. We benchmark three state-of-the-art LVLMs, i.e., GPT-4 Vision, GPT4o, and Claude-3-Opus, on the next item prediction task using the constructed Amazon Review Plus dataset, which includes additional item descriptions generated by LVLMs. Our evaluation examines five integration strategies: using LVLMs as recommender, item enhancer, reranker, and various combinations of these roles. The benchmark results reveal that 1) using LVLMs as rerankers is the most effective strategy, significantly outperforming others that rely on LVLMs to directly generate recommendations or only enhance items; 2) GPT-4o consistently achieves the best performance across most scenarios, particularly when employed as a reranker; 3) the computational inefficiency of LVLMs presents a major barrier to their widespread adoption in real-time multimodal recommendation systems. Our codes and datasets will be made publicly available upon acceptance.

Hypergraph neural networks have been promising tools for handling learning tasks involving higher-order data, with notable applications in web graphs, such as modeling multi-way hyperlink structures and complex user interactions. Yet, their generalization abilities in theory are less clear to us. In this paper, we seek to develop margin-based generalization bounds for four representative classes of hypergraph neural networks, including convolutional-based methods (UniGCN), set-based aggregation (AllDeepSets), invariant and equivariant transformations (M-IGN), and tensor-based approaches (T-MPHN). Through the PAC-Bayes framework, our results reveal the manner in which hypergraph structure and spectral norms of the learned weights can affect the generalization bounds, where the key technical challenge lies in developing new perturbation analysis for hypergraph neural networks, which offers a rigorous understanding of how variations in the model's weights and hypergraph structure impact its generalization behavior. Our empirical study examines the relationship between the practical performance and theoretical bounds of the models over synthetic and real-world datasets. One of our primary observations is the strong correlation between the theoretical bounds and empirical loss, with statistically significant consistency in most cases.

Livestreaming by VTubers -- animated 2D/3D avatars controlled by real individuals -- have recently garnered substantial global followings and achieved significant monetary success. Despite prior research highlighting the importance of realism in audience engagement, VTubers deliberately conceal their identities, cultivating dedicated fan communities through virtual personas. While previous studies underscore that building a core fan community is essential to a streamer's success, we lack an understanding of the characteristics and behaviors of viewers of this new type of streamer. Gaining a deeper insight into these viewers is critical for VTubers to enhance audience engagement, foster a more robust fan base, and attract a larger viewership. To address this gap, we conduct a comprehensive analysis of VTuber viewers on Bilibili, a leading livestreaming platform where nearly all VTubers in China stream. By compiling a first-of-its-kind dataset covering 2.7M livestreaming sessions, we investigate the characteristics, engagement patterns, and influence of VTuber viewers. Our research yields several valuable insights, which we then leverage to develop a tool to "recommend" future subscribers to VTubers. By reversing the typical approach of recommending streams to viewers, this tool assists VTubers in pinpointing potential future fans to pay more attention to, and thereby effectively growing their fan community.

Ethereum, as a representative of Web3, adopts a novel framework called Proposer Builder Separation (PBS) to prevent the centralization of block profits in the hands of institutional Ethereum stakers. Introducing builders to generate blocks based on public transactions, PBS aims to ensure that block profits are distributed among all stakers. Through the auction among builders, only one will win the block in each slot. Ideally, the equilibrium strategy of builders under public information would lead them to bid all block profits. However, builders are now capable of extracting profits from private order flows. In this paper, we explore the effect of PBS with private order flows. Specifically, we propose the asymmetry auction model of MEV-Boost auction. Moreover, we conduct empirical study on Ethereum blocks from January 2023 to May 2024. Our analysis indicates that private order flows contribute to 54.59% of the block value, indicating that different builders will build blocks with different valuations. Interestingly, we find that builders with more private order flows (i.e., higher block valuations) are more likely to win the block, while retain larger proportion of profits. In return, such builders will further attract more private order flows, resulting in a monopolistic market gradually. Our findings reveal that PBS in current stage is unable to balance the profit distribution, which just transits the centralization of block profits from institutional stakers to the monopolistic builder.

Tables are ubiquitous across various domains for concisely representing structured information. Empowering large language models (LLMs) to reason over tabular data represents an actively explored direction. However, since typical LLMs only support one-dimensional (1D) inputs, existing methods often flatten the two-dimensional (2D) table structure into a sequence of tokens, which can severely disrupt the spatial relationships and result in an inevitable loss of vital contextual information. In this paper, we first empirically demonstrate the detrimental impact of such flattening operations on the performance of LLMs in capturing the spatial information of tables through two elaborate proxy tasks. Subsequently, we introduce a simple yet effective positional encoding method, termed ``2D-TPE'' (Two-Dimensional Table Positional Encoding), to address this challenge. 2D-TPE enables each attention head to dynamically select a permutation order of tokens within the context for attending to them, where each permutation represents a distinct traversal mode for the table, such as column-wise or row-wise traversal. 2D-TPE effectively mitigates the risk of losing essential spatial information while preserving computational efficiency, thus better preserving the table structure. Extensive experiments across five benchmarks demonstrate that 2D-TPE outperforms strong baselines, underscoring the importance of preserving the table structure for accurate table comprehension. Comprehensive analysis further reveals the substantially better scalability of 2D-TPE to large tables than baselines.

There are a large number of text processing tasks in web applications, such as sentiment classification, summary extraction, and question answering. Recently, fine-tuning pre-trained language models (PLMs) to adapt to downstream text-processing tasks has attracted much attention. However, due to the differences in data, model, and tasks between the pre-training and fine-tuning processes, the fine-tuning process may suffer from catastrophic forgetting of pre-training knowledge, which may implicitly limit the model’s performance and generalization ability. To address these challenges, we propose a novel dual-model framework, termed as \emph{co}nsistency \emph{a}l\emph{i}gnment (CoAi). The insight of CoAi lies in building an auxiliary model that simulates the distribution of pre-training knowledge in real-time according to the current task, and co-training the task-specific model and the auxiliary model to balance the pre-training knowledge and task-specific knowledge during fine-tuning. Specifically, the auxiliary model is constructed on-the-fly to maintain the pre-training knowledge. Subsequently, CoAi simulates the pre-training process by performing distributional exploration in the parameter space, which is built upon our novel insight into the transformation between data and model parameter space. However, the objectives leveraged to construct the auxiliary model lead to the misalignment between the pre-training and task-specific knowledge. To alleviate the inconsistency, we employ an auxiliary variable to align the prediction distribution of the task-specific and the auxiliary models, inspired by constrastive clustering. We validate the effectiveness of CoAi on ten classic classification tasks and three generation tasks, showing consistent and significant improvements compared with state-of-the-art methods.

The advanced capabilities of Large Language Models (LLMs) have inspired the development of various interactive web services or applications, such as ChatGPT, which offer query inference services for users. Unlike traditional DNN model, the inference of LLM entails different iterations of forward computation for different queries, which result in efficiency challenges for existing run-to-completion batch-wise inference. Hence, some methods refine batch-wise inference to iteration-level by duplicating all nonlinear layers of LLM. However, this approach not only increases resource usage but also introduces idle computations to the batch due to the prefilling of newly added queries. Therefore, we propose BATON, an efficient batch-wise LLM inference scheme by dynamically adjusting processing batch, which can achieve near-zero idle computations without incurring additional resource consumption. To do so, BATON 1) shapes the vectors involved in the inference of the newly inserted query and processing batch to align dimensions and generates a new attention mask based on vector shaping to ensure inference correctness, which enables query inserting without consuming additional resource; 2) embeds prefilled \textit{Keys} and \textit{Values} of the new query into the \textit{KV\_Cache} of the processing batch by leveraging the prefilling and decoding separation mechanism, eliminating idle computations to the batch introduced by the prefilling process of the new query. Experimental results show that compared to the state-of-the-art solution Orca, \name outperforms improves query processing by up to 1.75$\times$.

Ubiquitous sensors and mobile devices have spurred the growth of Web-of-Things (WoT) services in smart cities, making accurate spatio-temporal traffic predictions increasingly crucial. Leveraging advances in deep learning, recent Spatio-Temporal Graph Neural Networks (STGNNs) have achieved remarkable results. However, these methods address scenario-specific spatio-temporal heterogeneity by designing model architectures, often overlooking the importance of selecting optimal spatio-temporal knowledge (i.e., model inputs). In this paper, we propose an automated framework for spatio-temporal knowledge optimization to address this challenge. Our framework seamlessly integrates with downstream models, enhancing their performance across various prediction tasks. Specifically, we design a knowledge search space composed of parameters that represent scenario-specific spatio-temporal correlations within data. Additionally, we employ a bandit-based multi-fidelity algorithm for knowledge optimization to solve the constraint of limited resource. Furthermore, we adopt a meta-learner to extract transferable meta-knowledge about optimal knowledge, facilitating efficient exploration of the search space. Extensive experiments on five widely used real-world datasets demonstrate the effectiveness of our proposed framework. To the best of our knowledge, we are the first to automatically optimize spatio-temporal knowledge for spatio-temporal traffic prediction.

We consider a robust aggregation problem in the presence of both truthful and adversarial experts. The truthful experts will report their private signals truthfully, while the adversarial experts can report arbitrarily. We assume experts are marginally symmetric in the sense that they share the same common prior and marginal posteriors. The rule maker needs to design an aggregator to predict the true world state from these experts' reports, without knowledge of the underlying information structures or adversarial strategies. We aim to find the optimal aggregator that outputs a forecast minimizing regret under the worst information structure and adversarial strategies. The regret is defined by the difference in expected loss between the aggregator and a benchmark who aggregates optimally given the information structure and reports of truthful experts. We focus on binary states and reports. Under L1 loss, we show that the truncated mean aggregator is optimal. When there are at most k adversaries, this aggregator discards the k lowest and highest reported values and averages the remaining ones. For L2 loss, the optimal aggregators are piecewise linear functions. All the optimalities hold when the ratio of adversaries is bounded above by a value determined by the experts' priors and posteriors. The regret only depends on the ratio of adversaries, not on their total number. For hard aggregators that output a decision, we prove that a random version of the truncated mean is optimal for both L1 and L2. This aggregator randomly follows a remaining value after discarding the $k$ lowest and highest reported values. We evaluate our aggregators numerically in an ensemble learning task. We also obtain negative results for general adversarial aggregation problems under broader information structures and report spaces.

Modern online advertising systems usually involve a large amount of advertisers in each auction, causing scalability issues. To mitigate the problem, two-stage auctions are designed and deployed in practice, enabling efficient allocations of ad slots among numerous candidate advertisers within a short response time. Such a design uses a fast but coarse model to select a small subset of advertisers in the first stage, and a slow yet refined model to finally decide the winners. However, existing two-stage auction mechanisms primarily focus on optimizing welfare, ignoring other crucial objectives of the platform, such as revenue. In this paper, we propose ad-wise selection metrics (namely Max-Wel and Max-Rev) that are based on an ad's contribution to the platform's objective (welfare or revenue). Then we provide theoretical guarantees for the proposed metrics. Our method is applicable to both welfare and revenue optimizations and can be easily implemented using neural networks. We conduct extensive experiments on both synthetic and industrial data to demonstrate the advantages of our proposed selection metrics over existing baselines.

Prior studies on Visual Sentiment Understanding (VSU) primarily rely on the explicit scene information (e.g., facial expression) to judge visual sentiments, which largely ignore implicit scene information (e.g., human action, objection relation and visual background), while such information is critical for precisely discovering visual sentiments. Motivated by this, this paper proposes a new Omni-scene driven visual Sentiment Identifying, Locating and Attributing in videos (Omni-SILA) task, aiming to interactively and precisely identify, locate and attribute visual sentiments through both explicit and implicit scene information. Furthermore, this paper believes that this Omni-SILA task faces two key challenges: modeling scene and highlighting implicit scene beyond explicit. To this end, this paper proposes an Implicit-enhanced Causal MoE (ICM) approach for addressing the Omni-SILA task. Specifically, a Scene-Balanced MoE (SBM) and an Implicit-Enhanced Causal (IEC) blocks are tailored to model scene information and highlight the implicit scene information beyond explicit, respectively. Extensive experimental results on our constructed explicit and implicit Omni-SILA datasets demonstrate the great advantage of the proposed ICM approach over advanced Video-LLMs.

Accurate measurement of mobile network performance is crucial for optimizing user experience and ensuring regulatory compliance. Traditional methods like crowdsourcing approaches, though effective, depend heavily on user participation and extensive infrastructure. In this paper, we introduce adNPM, a novel technique for measuring download speed by embedded measurement code in ads displayed across web browsers and mobile apps, without requiring user participation. Through controlled lab tests and real-world deployments in 15 countries, we show that adNPM produces results comparable to well-established tools such as Speedtest by Ookla and Opensignal while consuming significantly less data. Our solution leverages ad campaigns to collect extensive data from diverse demographics and geographic regions, providing deep insights into the performance of major Internet Service Providers (ISPs). Furthermore, adNPM can segment download speed analyses by demographic factors and operating systems, making it a versatile and scalable tool for network performance assessment.

Efficient processing of large-scale graphs with billions to trillions of edges is essential for training graph-based large language models (LLMs) in web-scale systems. The increasing complexity and size of these models create significant communication challenges due to the extensive message exchanges required across distributed nodes. Current graph engines struggle to effectively scale across hundreds of computing nodes because they often overlook variations in communication costs within the interconnection hierarchy. To address this challenge, we introduce TuComm, a communication hierarchy-aware engine specifically designed to optimize distributed training of graph-based LLMs. By leveraging hierarchical network topology, TuComm dynamically aggregates and transfers messages, fully accounting for the underlying communication domains, thereby enhancing the efficiency of distributed model training across large-scale systems. We implemented TuComm on top of the message passing interface (MPI), incorporating innovations such as dynamic buffer expansion and active buffer switching to enhance scalability. Evaluations conducted on synthetic and real-world datasets, utilizing up to 79,024 nodes and over 1.2 million processor cores, demonstrate that TuComm surpasses leading graph-parallel systems and state-of-the-art counterparts in both throughput and scalability. Moreover, we have deployed TuComm on a production supercomputer, where it consistently outperforms top solutions on the Graph500 list. These results highlight TuComm’s potential to significantly enhance the efficiency of distributed large-scale graph-based LLM training by optimizing communication among distributed systems, making it an invaluable communication engine for web-scale model training.

The standard framework of online bidding algorithm design assumes that the seller commits himself to faithfully implementing the rules of the adopted auction. However, the seller may attempt to cheat in execution to increase his revenue if the auction belongs to the class of non-credible auctions. For example, in a second-price auction, the seller could create a fake bid between the highest bid and the second highest bid. This paper focuses on one such case of online bidding in repeated second-price auctions. At each time $t$, the winner with bid $b_t$ is charged not the highest competing bid $d_t$ but a manipulated price $p_t = \alpha_0 d_t + (1-\alpha_0) b_t$, where the parameter $\alpha_0 \in [0, 1]$ in essence measures the seller's credibility. Unlike classic repeated-auction settings where the bidder has access to samples $(d_s)\_{s=1}^{t-1}$, she can only receive mixed signals of $(b_s)\_{s=1}^{t-1}$, $(d_s)\_{s=1}^{t-1}$ and $\alpha_0$ in this problem. The task for the bidder is to learn not only the bid distributions of her competitors but also the seller's credibility. We establish regret lower bounds in various information models and provide corresponding online bidding algorithms that can achieve near-optimal performance. Specifically, we consider three cases of prior information based on whether the credibility $\alpha_0$ and the distribution of the highest competing bids are known. Our goal is to characterize the landscape of online bidding in non-credible second-price auctions and understand the impact of the seller's credibility on online bidding algorithm design under different information structures.