WWW 2025 论文列表

Zero-shot named entity recognition (NER) aims to develop entity recognition systems from unannotated text corpora. This task presents substantial challenges due to minimal human intervention. Recent work has adapted large language models (LLMs) for zero-shot NER by crafting specialized prompt templates. And it advances the models’ self-learning ability by incorporating self-annotated demonstrations. Two important challenges persist: (i) Correlations between contexts surrounding entities are overlooked, leading to wrong type predictions or entity omissions. (ii) The indiscriminate use of task demonstrations, retrieved through shallow similarity-based strategies, severely misleads the inferences made by LLMs. In this paper, we introduce CMAS, or cooperative multi-agent system, a framework for zero-shot NER that uses the collective intelligence and tailored abilities of multiple agents to address the challenges outlined above. Cooperative multi-agent system (CMAS) has four main agents: (i) a self-annotator, (ii) a type-related feature (TRF) extractor, (iii) a demonstration discriminator, and (iv) an overall predictor. To explicitly capture correlations between contexts surrounding entities, CMAS reformulates NER into two subtasks: recognizing named entities and identifying entity type-related features within the target sentence. Moreover, pseudo-labels for TRFs are generated using mutual-information criteria without requiring human effort, facilitating the prediction of the TRF extractor. To assess the quality of demonstrations, a demonstration discriminator is established to incorporate the self-reflection mechanism, automatically evaluating helpfulness scores for the target sentence and enabling controllable utilization of demonstrations. Experimental results show that CMAS significantly improves zero-shot NER performance across six benchmarks, including both domain-specific and general-domain scenarios. Furthermore, CMAS demonstrates its effectiveness in few-shot settings and with various LLM backbones.

Large language models (LLMs) are expected to offer structured Markdown responses for the sake of readability in web chatbots (e.g., ChatGPT). Although there are a myriad of metrics to evaluate LLMs, they fail to evaluate the readability from the view of output content structure. To this end, we focus on an overlooked yet important metric --- Markdown Awareness, which directly impacts the readability and structure of the content generated by these language models. In this paper, we introduce MDEval, a comprehensive benchmark to assess Markdown Awareness for LLMs, by constructing a dataset with 20K instances covering 10 subjects in English and Chinese. Unlike traditional model-based evaluations, MDEval provides excellent interpretability by combining model-based generation tasks and statistical methods. Our results demonstrate that MDEval achieves a Spearman correlation of 0.791 and an accuracy of 84.1% with human, outperforming existing methods by a large margin. Extensive experimental results also show that through fine-tuning over our proposed dataset, less performant open-source models are able to achieve comparable performance to GPT-4o in terms of Markdown Awareness. To ensure reproducibility and transparency, MDEval is open sourced at https://anonymous.4open.science/r/MDEval-Benchmark-1730/.

Federated heterogeneous graph learning, an extension of federated learning, enables effective representation of complex multidimensional relationships while preserving data privacy. In horizontal federated heterogeneous graph learning, data from different parties often differ in topology and semantic distributions, causing sensitivity to distribution imbalance and amplifying the complexity of the topological structure. This interaction makes it difficult for models to learn shared representations, leading to increased instability during training. To address these challenges, this paper proposes a novel multi-scale adaptive horizontal federated heterogeneous graph learning method MAFedHGL. A random masking mechanism forces the model to infer missing connections. The model also captures multi-hop and multi-path connections using high-order topology mining, enhancing robustness against structural heterogeneity. Dynamic semantic consistency modeling uses a masking matrix to recover and integrate diverse node attributes, ensuring both global and local semantic consistency. Using clustering coefficients as aggregation weights enables clients with richer structural information to contribute more effectively to the global model, improving adaptability and performance across varying data distributions in horizontal federated heterogeneous graph learning. Extensive experiments on multiple public heterogeneous graph datasets validate that the proposed method outperforms state-of-the-art methods in both performance and robustness across various data distribution scenarios.

Recent studies have indicated that Large Language Models (LLMs) harbor an inherent understanding of truthfulness, yet often fail to express fully and generate false statements. This gap between "knowing" and "telling" poses a challenge for ensuring the truthfulness of generated content. Inspired by recent work on the practice of encoding human-interpretable concepts linearly within large language models, we treat truthfulness as a specially linearly encoded concept within LLMs, and introduce Adaptive Activation Steering (ACT), a tuning-free method that adaptively shifts LLM's activations in the "truthful" direction during inference. ACT addresses diverse categories of hallucinations by utilizing diverse truthfulness-related steering vectors and adjusting the steering intensity adaptively. Applied as an add-on across various models, ACT significantly improves truthfulness in LLaMA ($\uparrow$ 142\%), LLaMA2 ($\uparrow$ 24\%), Alpaca ($\uparrow$ 36\%), Vicuna ($\uparrow$ 28\%), LLaMA2-Chat ($\uparrow$ 19\%), and LLaMA3($\uparrow$ 34\%). Furthermore, we verify ACT's scalability across larger models (13B, 33B, 65B), underscoring the adaptability of ACT to large-scale language models. Code: https://anonymous.4open.science/r/ACT24.

In today's technology-driven world, web services have opened up new opportunities for blind and visually impaired people to interact independently. Securing interactions with these services is crucial; however, currently deployed methods of web authentication mainly concentrate on sighted users, overlooking the specific needs of the blind and visually impaired community. In this paper, we address this critical gap by investigating the security and accessibility aspects of these web authentication methods when adopted by blind and visually impaired users. We model web authentication for such users as screen reader assisted authentication and introduce an evaluation framework called Authentication Workflows Accessibility Review and Evaluation (AWARE). Using AWARE, we then systematically assessed popular PC-based and smartphone-based screen readers against different types of deployed web authentication methods, including variants of 2FA and passwordless schemes, to simulate real-world scenarios for blind and visually impaired individuals. We analyzed these screen reader assisted authentication interactions with authentication methods in three settings: using a terminal (PC) with screen readers, a combination of the terminal (PC) and smartphone with screen readers, and smartphones with integrated screen readers. The results of our study underscore significant weaknesses in all of our observed screen reader assisted authentication scenarios for real-life authentication methods. These weaknesses, encompassing specific accessibility issues caused by imprecise screen reader instructions, highlight vulnerability concerning observed scenarios for both real-world and research literature based attacks, including phishing, concurrency, fatigue, cross-service, and shoulder surfing. Broadly, our AWARE framework can be used by authentication system designers as a precursor to user studies which are typically time-consuming and tedious to perform, independently allowing to unfold security and accessibility problems early which designers can address prior to full-fledged user testing of more isolated issues.

Federated reinforcement learning (FRL) allows agents to jointly learn a global decision-making policy under the guidance of a central server. While FRL has advantages, its decentralized design makes it prone to poisoning attacks. To mitigate this, Byzantine-robust aggregation techniques tailored for FRL have been introduced. Yet, in our work, we reveal that these current Byzantine-robust techniques are not immune to our newly introduced Normalized attack. Distinct from previous attacks that targeted enlarging the distance of policy updates before and after an attack, our Normalized attack emphasizes on maximizing the angle of deviation between these updates. To counter these threats, we develop an ensemble FRL approach that is provably secure against both known and our newly proposed attacks. Our ensemble method involves training multiple global policies, where each is learnt by a group of agents using any foundational aggregation rule. These well-trained global policies then individually predict the action for a specific test state. The ultimate action is chosen based on a majority vote for discrete action systems or the geometric median for continuous ones. Our experimental results across different settings show that the Normalized attack can greatly disrupt non-ensemble Byzantine-robust methods, and our ensemble approach offers substantial resistance against poisoning attacks.

Do real-world hypergraphs obey any patterns? Are power laws fundamental in hypergraphs as they are in real-world graphs? What generator can reproduce these patterns? A hypergraph is a generalization of a conventional graph, and it consists of nodes and hyperedges, with each hyperedge joining any number of nodes. Hypergraphs are adept at representing group interactions where two or more entities interact simultaneously, such as collaborative research and group discussions. In a wide range of real-world hypergraphs, we discover power-law or log-logistic distributions in eight structural properties. To simulate these observed patterns, we introduce HyRec, a tractable and realistic generative model leveraging the Kronecker product. We mathematically demonstrate that HyRec accurately reproduces both the patterns we observed and typical evolutionary trends found in real-world hypergraphs. To fit the parameters of HyRec to large-scale hypergraphs, we design SingFit, a fast and space-efficient algorithm successfully applied to eleven real-world hypergraphs with up to one million nodes and hyperedges. This paper makes the following contributions: (a) Discoveries: we identify multiple patterns that real-world hypergraphs obey, (b) Model: we propose HyRec, a tractable and realistic model capable of reproducing real-world hypergraphs efficiently (spec., with fewer than 1,000 parameters) with the support of SingFit, and (c) Proofs: we prove that HyRec adheres to these patterns.

Over the past few years, the European Commission has made significant steps to reduce disinformation in cyberspace. One of those steps has been the introduction of the 2022 "Strengthened Code of Practice on Disinformation". Signed by leading online platforms, this Strengthened Code of Practice on Disinformation is an attempt to combat disinformation on the Web. The Code of Practice includes a variety of measures including the demonetization of disinformation, urging, for example, advertisers "to avoid the placement of advertising next to Disinformation content". In this work, we set out to explore what was the impact of the Code of Practice and especially to explore to what extent ad networks continue to advertise on dis-/mis-information sites. We perform a historical analysis and find that, although at a hasty glance things may seem to be improving, there is really no significant reduction in the amount of advertising relationships among popular misinformation websites and major ad networks. In fact, we show that ad networks have withdrawn mostly from unpopular misinformation websites with very few visitors, but still form relationships with highly unreliable websites that account for the majority of misinformation traffic. To make matters worse, we show that ad networks continue to place advertisements of legitimate companies next to misinformation content. In fact we show that major ad networks place ads in almost 400 misinformation websites of our dataset.

Recent research on query generation has focused on using Large Language Models (LLMs), which despite bringing state-of-the-art performance, also introduce issues with hallucinations in the generated queries. In this work, we introduce relevance hallucination and factuality hallucination as a new typology for hallucination problems brought by query generation based on LLMs. We propose an effective way to separate content from form in LLM-generated queries, which preserves the factual knowledge extracted and integrated from the inputs and compiles the syntactic structure, including function words, using the powerful linguistic capabilities of the LLM. Specifically, we introduce a model-agnostic and training-free method that turns the **Large** Language Model into a **P**o**i**nter-**G**enerator (**LargePiG**), where the pointer attention distribution leverages the LLM's inherent attention weights, and the copy probability is derived from the difference between the vocabulary distribution of the model’s high layers and the last layer. To validate the effectiveness of LargePiG, we constructed two datasets for assessing the hallucination problems in query generation, covering both document and video scenarios. Empirical studies on various LLMs demonstrated the superiority of LargePiG on both datasets. Additional experiments also verified that LargePiG could reduce hallucination in large vision language models and improve the accuracy of document-based question answering and factuality evaluation tasks.

The expression of mental health symptoms through non-traditional means, such as memes, has gained remarkable attention over the past few years, with users often highlighting their mental health struggles through figurative intricacies within memes. While humans rely on commonsense knowledge to interpret these complex expressions, current Multimodal Language Models (MLMs) struggle to capture these figurative aspects inherent in memes. To address this gap, we introduce a novel dataset, AxiOM, derived from the GAD anxiety questionnaire, which categorizes memes into six fine-grained anxiety symptoms. Next, we propose a commonsense and domain-enriched framework, M3H, to enhance MLMs’ ability to interpret figurative language and commonsense knowledge. The overarching goal remains to first understand and then classify the mental health symptoms expressed in memes. We benchmark M3H against 6 competitive baselines (with 20 variations), demonstrating substantial improvements in both quantitative and qualitative metrics, including a detailed human evaluation. We observe a clear improvement of 4.20% and 4.66% on weighted-F1 metric. To assess the generalizability, we perform extensive experiments on a publicly available dataset, RESTORE, for depressive symptom identification, presenting an extensive ablation study that highlights the contribution of each module in both datasets. Our findings reveal key limitations in existing models and the advantage of employing commonsense understanding to enhance figurative understanding.

Personalized user preference driven recommendations have seamlessly intertwined with our daily lives. However, item providers may expect specific items to gradually increase their appeal to users over the course of users’ long-term interactions with the system, but few studies pay attention to this problem. In this paper, we propose a novel intention-based targeted multi-round proactive recommendation method, dubbed ITMPRec. Specifically, we first choose a set of target items from the target category, by conducting a pre-match strategy. Afterward, we utilize a multi-round nudging recommendation method, in which we design a module to quantify the intention-level dynamic evolution of users so that we could choose more appropriate intermediate items during guidance. Besides, we model each user’s sensitivity to the changes in representation induced by the intermediate items they accept. Finally, we propose a design for a Large Language Model (LLM) agent as a pluggable component to simulate user feedback. This design offers an alternative to the traditional click model based on distribution, relying on the agent’s external knowledge and reasoning capabilities. Through extensive experiments on four public datasets, we demonstrate the superiority of ITMPRec compared to seven baseline models. The code repository is available at https://anonymous.4open.science/r/ITMPRec-D821.

Addressing the critical challenge of data incompleteness in inner product matrix estimation, we introduce a novel eigenvalue correction method designed to precisely reconstruct true inner product matrices from incomplete data. Utilizing random matrix theory, our method adjusts the eigenvalue distribution of the estimated inner product matrix to align with the ground-truth. This approach significantly reduces estimation errors for both inner product matrices and the derived Euclidean distance matrices, thereby enhancing the effectiveness of similarity searches on incomplete data. Our method surpasses traditional data imputation and similarity calibration techniques in both maximum inner product search and nearest neighbor search tasks, demonstrating marked advancements in managing incomplete data. It exhibits robust performance across various missing rates and diverse scenarios.

Review personalization aims at presenting the most relevant reviews of a product according to the preferences of the individual user. Existing studies of review personalization use the reviews authored by the user as a proxy for their preferences, and henceforth as a means for learning and evaluating personalization quality. In this work, we suggest using review votes rather than authorship for personalization. We suggest MAGLLM, an approach that leverages heterogeneous graphs for modeling the relationships among reviews, products, and users, with large language model (LLM) to enrich user representation on the graph. Our evaluation over a unique public dataset that includes user voting information indicates that the vote signal yields substantially higher personalization performance across a variety of recommendation methods and e-commerce domains. It also indicates that our graph-LLM approach outperforms comparative baselines and algorithmic alternatives. We conclude with concrete recommendations for e-commerce platforms seeking to enhance their review personalization experience.

As the number and sophistication of cyber attacks have increased, threat hunting has become a critical aspect of active security, enabling proactive detection and mitigation of threats before they cause significant harm. Open-source cyber threat intelligence (OSCTI) is a valuable resource for threat hunters, however, it often comes in unstructured formats that require further manual analysis. Previous studies aimed at automating OSCTI analysis are limited since (1) they failed to provide actionable outputs, (2) they did not take advantage of images present in OSCTI sources, and (3) they focused on on-premises environments, overlooking the growing importance of cloud environments. To address these gaps, we propose LLMCloudHunter, a novel framework that leverages large language models (LLMs) to automatically generate generic-signature detection rule candidates from textual and visual OSCTI data. We evaluated the quality of the rules generated by the proposed framework using 20 annotated real-world cloud threat reports. The results show that our framework achieved a precision of 83\% and recall of 99\% for the task of accurately extracting API calls made by the threat actor and a precision of 99\% with a recall of 97\% for IoCs. Additionally, 99.18\% of the generated detection rule candidates were successfully compiled and converted into Splunk queries.

Discontinuous Named Entity Recognition (DNER) presents a challenging problem where entities may be scattered across multiple non-adjacent tokens, making traditional sequence labelling approaches inadequate. Existing methods predominantly rely on custom tagging schemes to handle these discontinuous entities, resulting in models tightly coupled to specific tagging strategies and lacking generalisability across diverse datasets. To address these challenges, we propose TriG-NER, a novel Triplet-Grid Framework that introduces a generalisable approach to learning robust token-level representations for discontinuous entity extraction. Our framework applies triplet loss at the token level, where similarity is defined by word pairs existing within the same entity, effectively pulling together similar and pushing apart dissimilar ones. This approach enhances entity boundary detection and reduces the dependency on specific tagging schemes by focusing on word-pair relationships within a flexible grid structure. We evaluate TriG-NER on three benchmark DNER datasets and demonstrate significant improvements over existing grid-based architectures. These results underscore our framework’s effectiveness in capturing complex entity structures and its adaptability to various tagging schemes, setting a new benchmark for discontinuous entity extraction.

Off-policy evaluation (OPE) is a crucial problem in reinforcement learning (RL), where the goal is to estimate the long-term cumulative reward of a target policy using historical data generated by a potentially different behaviour policy. In many real-world applications, such as precision medicine and recommendation systems, unobserved confounders may influence the action, reward, and state transition dynamics, which leads to biased estimates if not properly addressed. While existing methods for handling unobserved confounders in OPE focus on single-action settings, they are less effective in multi-action scenarios commonly found in practical applications, where an agent can take multiple actions simultaneously. In this paper, we propose a novel auxiliary variable-aided method for OPE in multi-action settings with unobserved confounders. Our approach overcomes the limitations of traditional auxiliary variable methods for multi-action scenarios by requiring only a single auxiliary variable, relaxing the need for as many auxiliary variables as the actions. Through theoretical analysis, we prove that our method provides an unbiased estimation of the target policy value. Empirical evaluations demonstrate that our estimator achieves better performance compared to existing baseline methods, highlighting its effectiveness and reliability in addressing unobserved confounders in multi-action OPE settings.

Increasing studies focus on machine unlearning as it upholds users' right to be forgotten, under which individuals can request the removal of their specified samples from trained models. However, the auditing of machine unlearning processes remains significantly underexplored. Although some existing methods offer unlearning auditing by leveraging backdoors, these backdoor-based approaches are inefficient and impractical, as they necessitate involvement in the initial model training process to embed the backdoors. In this paper, we propose a TAilored Posterior diffErence (TAPE) method to provide unlearning auditing independently of original model training. We observe that the process of machine unlearning inherently introduces changes in the model, which contains information related to the erased data. TAPE leverages unlearning model differences to assess how much information has been removed through the unlearning operation. Firstly, TAPE mimics the unlearned posterior differences by quickly building unlearned shadow models based on first-order influence estimation. Secondly, we train a Reconstructor model to extract and evaluate the private information of the unlearned posterior differences to audit unlearning. Existing privacy reconstructing methods based on posterior differences are only feasible for model updates of a single sample. To enable the reconstruction effective for multi-sample unlearning requests, we propose two strategies, unlearned data perturbation and unlearned influence-based division, to augment the posterior difference. Extensive experimental results indicate the significant superiority of TAPE over the state-of-the-art unlearning verification methods, at least 4.5$\times$ efficiency speedup and supporting the auditing for broader unlearning scenarios.

Diffusion models (DMs) have emerged as promising approaches for sequential recommendation due to their strong ability to model data distributions and generate high-quality items. Existing work typically adds noise to the next item and progressively denoises it guided by the user's interaction sequence, generating items that closely align with user interests. However, we identify two key issues in this paradigm. First, the sequences are often heterogeneous in length and content, exhibiting noise due to stochastic user behaviors. Using such sequences as guidance may hinder DMs from accurately understanding user interests. Second, DMs are prone to data bias and tend to generate only the popular items that dominate the training dataset, thus failing to meet the personalized needs of different users. To address these issues, we propose Distinguished Quantized Guidance for Diffusion-based Sequence Recommendation (DiQDiff), which aims to extract robust guidance to understand user interests and generate distinguished items for personalized user interests within DMs. To extract robust guidance, DiQDiff introduces Semantic Vector Quantization (SVQ) to quantize sequences into semantic vectors (e.g., collaborative signals and category interests) using a codebook, which can enrich the guidance to better understand user interests. To generate distinguished items, DiQDiff personalizes the generation through Contrastive Discrepancy Maximization (CDM), which maximizes the distance between denoising trajectories using contrastive loss to prevent biased generation for different users. Extensive experiments are conducted to compare DiQDiff with multiple baseline models across four widely-used datasets. The superior recommendation performance of DiQDiff demonstrates its effectiveness in the sequential recommendation.

Road trajectory generation creates synthetic yet realistic trajectories to tackle data collection costs and privacy concerns. Existing methods generate a trajectory either segment-by-segment using sequence models or holistically in one step using diffusion models. Sequence-based models have good regularity and consistency (i.e., resemble the input trajectories) but lack diversity, while diffusion-based models enhance diversity but sacrifice regularity and consistency. To combine the merits of existing methods, we propose $\textit{Seed}$, by bridging $\underline{\textit{se}}$quenc$\underline{\textit{e}}$ and $\underline{\textit{d}}$iffusion models for trajectory generation. In particular, Seed adopts a \textit{conditional diffusion structure}, where a Transformer models the movement of each trajectory along the road segments, and conditioned on the Transformer's output, a diffusion model recovers the next road segment from random noise. The rationale is that the Transformer captures sequential movement patterns for regularity and consistency, while the diffusion model introduces diversity by recovering from noise. To train Seed, we adopt Node2vec to learn embeddings for the road segments to prepare model inputs, supervise learning using the task of trajectory reconstruction, and design a curriculum learning strategy to accelerate convergence. We compare Seed with 8 state-of-the-art trajectory generation methods on 3 datasets, and the results show that Seed improves the best-performing baseline by over 50\%.