Large Language Models (LLMs) enable Web-scale multilingual content analysis but face critical challenges in scaling to long-tail languages and ensuring robustness. Current research is split between two isolated trajectories: a Macro-Paradigm (system-level engineering) and a Micro-Paradigm (internal model intervention). We argue that a true Web-scale solution requires their systematic fusion, balancing large-scale data processing with fine-grained model control. We introduce the Control-Tower Framework (CTF), a novel methodology designed to systematically enhance powerful, pre-trained base models. Inspired by control-theoretic ideas, CTF transforms a base model into a controllable analysis engine via three synergistic stages: (1) Micro-enhanced pre-training that injects linguistic priors (e.g., syntax) to build a robust semantic foundation; (2) a control-inspired fine-tuning stage where a heuristic dynamic feedback loop, driven by micro-level error signals (e.g., knowledge editing loss), actively adjusts the macro-scale learning curriculum; and (3) Macro-optimized inference using Minimum Bayes Risk (MBR) decoding to enhance robustness on noisy user-generated content (UGC). Extensive experiments show that CTF surpasses the leading open-weights model, Tower+ 9B FT, by a substantial margin of +2.18 XCOMET-XXL on low-resource languages (WMT24++). Crucially, CTF unlocks large-scale cross-lingual Web mining by converting unstructured Web text into machine-analyzable assets. We evidence this with substantial gains across both document-level (on MARC) and aspect-based (on SemEval-2016) sentiment analysis tasks. Our work offers a practical pathway toward building more reliable, scalable, and controllable global information ecosystems.
论文检索
输入标题、作者或关键词，从 15,207 篇学术成果中精准定位
As a key to accessing research impact, citation dynamics underpins research evaluation, scholarly recommendation, and the study of knowledge diffusion. Citation prediction is particularly critical for newborn papers, where early assessment must be performed without citation signals and under highly long-tailed distributions. We identify two key research gaps: (i) insufficient modeling of implicit factors of scientific impact, leading to reliance on coarse proxies; and (ii) a lack of bias-aware learning that can deliver stable predictions on lowly cited papers. We address these gaps by proposing a Bias-Aware Citation Prediction Framework, which combines multi-agent feature extraction with robust graph representation learning. First, a multi-agent × graph co-learning module derives fine-grained, interpretable signals, such as reproducibility, collaboration network, and text quality, from metadata and external resources, and fuses them with heterogeneous-network embeddings to provide rich supervision even in the absence of early citation signals. Second, we incorporate a set of robust mechanisms: a two-stage forward process that routes explicit factors through an intermediate exposure estimate, GroupDRO to optimize worst-case group risk across environments, and a regularization head that performs what-if analyses on controllable factors under monotonicity and smoothness constraints. Comprehensive experiments on two real-world datasets demonstrate the effectiveness of our proposed model. Specifically, our model achieves around a 13% reduction in error metrics (MALE and RMSLE) and a notable 5.5% improvement in the ranking metric (NDCG) over the baseline methods.
Retrieval-Augmented Generation (RAG)-enhanced LLM systems, while powerful, introduce substantial inference costs due to the inclusion of an extra multi-stage pipeline that dynamically retrieves and synthesizes information from external knowledge sources. This high operational cost exposes a critical vulnerability to Inference Cost Attacks (ICAs). However, existing ICAs often rely on the impractical assumption of direct prompt manipulation. We argue that a more feasible and potent threat to RAG-enhanced LLM systems arises from poisoning external knowledge bases (e.g., web knowledge from the Internet). In this work, we introduce the Retrieval-Augmented Inference Cost Attack (RA-ICA), a novel attacking paradigm that targets the computational cost of RAG-enhanced LLM systems by injecting malicious documents into external knowledge corpus. To operationalize this attack, we propose Computational Resource Exhaustion via External Poisoning (CREEP), a novel framework that leverages LLM agents to automatically craft malicious documents that are both semantically relevant for retrieval and potent for inducing an abnormal increase in token consumption during the inference phase. To enhance the attack's effectiveness, we introduce Memory-Augmented Group Relative Policy Optimization (MA-GRPO), a novel reinforcement learning algorithm that fine-tunes the agents by learning from a dynamic memory of historical best adversarial documents. Extensive experiments across three real-world datasets demonstrate that RA-ICA increases token consumption by up to 13.12 times with an over 90% success rate, without degrading the integrity of the generated answer.
Real-world multimodal knowledge graphs (MKGs) are inherently heterogeneous, modeling entities that are associated with diverse modalities. Traditional knowledge graph embedding (KGE) methods excel at learning continuous representations of entities and relations, yet they are typically designed for unimodal settings. Recent approaches extend KGE to multimodal settings but remain constrained, often processing modalities in isolation, resulting in weak cross-modal alignment, and relying on simplistic assumptions such as uniform modality availability across entities. Vision--Language Models (VLMs) offer a powerful way to align diverse modalities within a shared embedding space. We propose Vision--Language Knowledge Graph Embeddings (VL-KGE), a framework that integrates cross-modal alignment from VLMs with structured relational modeling to learn unified multimodal representations of knowledge graphs. Experiments on WN9-IMG and two novel fine art MKGs, WikiArt-MKG-v1 and WikiArt-MKG-v2, demonstrate that VL-KGE consistently improves over traditional unimodal and multimodal KGE methods in link prediction tasks. Our results highlight the value of VLMs for multimodal KGE, enabling more robust and structured reasoning over large-scale heterogeneous knowledge graphs.
Federated learning offers a promising paradigm for privacy-preserving traffic prediction, yet its performance is often challenged by the non-identically and independently distributed (non-IID) nature of decentralized traffic data. Existing federated methods frequently struggle with this data heterogeneity, typically entangling globally shared patterns with client-specific local dynamics within a single representation. In this work, we postulate that this heterogeneity stems from the entanglement of two distinct generative sources: client-specific localized dynamics and cross-client global spatial-temporal patterns. Motivated by this perspective, we introduce FedDis, a novel framework that, to the best of our knowledge, is the first to leverage causal disentanglement for federated spatial-temporal prediction. Architecturally, FedDis comprises a dual-branch design wherein a Personalized Bank learns to capture client-specific factors, while a Global Pattern Bank distills common knowledge. This separation enables robust cross-client knowledge transfer while preserving high adaptability to unique local environments. Crucially, a mutual information minimization objective is employed to enforce informational orthogonality between the two branches, thereby ensuring effective disentanglement. Comprehensive experiments conducted on four real-world benchmark datasets demonstrate that FedDis consistently achieves state-of-the-art performance, promising efficiency, and superior expandability. To ease the reproducibility, we release our implementation code online. https://github.com/Jlu-zcy/www2026_FedDis.
While Large Language Models (LLMs) excel at many reasoning tasks, their native inability to produce calibrated, multi-class probability distributions limits their use in high-stakes Web applications like content moderation and fraud detection. Existing methods to elicit probabilities from LLMs either sacrifice their crucial Chain-of-Thought (CoT) reasoning capabilities or suffer from poor calibration. To address this, we introduce a new paradigm, Verbalized Probability Distribution, and a novel training framework, RLVP (Reinforcement Learning with Verbalized Probabilities). RLVP fine-tunes an LLM to generate both an interpretable CoT and a complete, verbalized probability distribution. We overcome the ''insufficient reward granularity'' problem in standard Reinforcement Learning (RL) for classification by using soft probabilities from expert tabular models as a dense reward curriculum. Through large-scale joint training on 169 tabular tasks, we demonstrate that a single RLVP-trained model can surpass a strong, task-specific XGBoost baseline on up to 55% of tasks. More importantly, the trained model achieves state-of-the-art few-shot performance on unseen, heterogeneous Web benchmarks that mix structured data with free text, achieving performance comparable to or superior than expert models trained on the same limited data. This showcases a strong capability for generalization and knowledge transfer to complex Web data. Our work presents a viable path toward building general-purpose, probabilistically-sound, and interpretable foundation models for the Web.
Knowledge Tracing (KT) is a core task in intelligent tutoring systems, designed to model the dynamic evolution of students' knowledge states by predicting their performance on specific problems. However, existing KT models, primarily developed for traditional subject domains, exhibit poor generalization to programming tasks due to several distinct challenges. First, student responses in programming tasks are characterized by high uncertainty. Second, programming tasks typically provide continuous scores based on the proportion of passed test cases, in contrast to the binary correctness assumed by traditional KT models, with these test-case-based scores often being noisy. Third, most KT models rely exclusively on the understanding of knowledge concepts for performance prediction, whereas success in programming tasks is heavily contingent upon logical reasoning and problem-specific comprehension. To address these issues, we propose an LLM-driven Interaction Enrichment Framework (MIE) to mitigate high uncertainty and problematic labeling, and introduce the Multi-Level Programming Knowledge Tracing (MLPKT) model to capture students' knowledge states across multiple dimensions. MLPKT conducts multi-layer analysis of student submissions to identify the root causes of errors and assign semantically meaningful fine-grained labels. Additionally, we propose a three-level, three-phase KT architecture that captures knowledge dynamics across three dimensions—problems, concepts, and logical skills—through the phases of learning, forgetting/reinforcement, and application. Extensive experiments on three datasets demonstrate that MIE+MLPKT consistently outperforms 18 baseline methods. Our code is available at: https://anonymous.4open.science/r/MIE-MLPKT-D654.
Time series anomaly detection aims to identify samples that deviate from a normal sample distribution in a time series, enabling various web-centric applications. Most existing approaches are static, targeting pre-defined types of anomalies. These methods thus fail to work well on streaming time series with changing data distributions and anomaly formats. To contend with such streaming time series and to accommodate memory constraints, we propose the first data-efficient streaming time series anomaly detection framework, called DESS. To accumulate historical knowledge, DESS includes a novel evolving proxy generation module to synthesize a small but informative proxy summarizing the historical data, facilitating data efficiency. Next, DESS employs an innovative heterogeneous temporal feature extraction module to explicitly capture correlations of multi-level time series semantics. Finally, DESS enables fast streaming anomaly detection by employing a parameter-efficient training scheme that only activates a subset of lightweight parameters while ensuring performance. Extensive experiments on real data offer insight into the effectiveness and efficiency of DESS, showing that it is able to outperform the best baselines by up to 17.53% while reducing the training time by up to 64.88%.
Towards Multi-Label Text Interpretation with Chain-of-Thought Prompting and Contextualized Knowledge
Existing multi-label topic models face several challenges when interpreting texts annotated with multiple labels: (1) they often associate irrelevant text segments with incorrect labels, which negatively impacts both segment and label interpretation; (2) they fail to effectively capture the semantic relationships between tokens and labels within the segment; (3) they do not integrate contextualized knowledge that could improve interpretability. To overcome these issues, we introduce the Contextualized Prompting Topic Model (CPTM). CPTM utilizes Chain-of-Thought (CoT) prompting to better align text segments with their semantically relevant labels. Furthermore, it integrates label-specific token visualization and topic mining procedure to facilitate the interpretation of tokens and labels. Experimental evaluations conducted on three multi-label text datasets show that CPTM significantly outperforms existing models in both segment and label interpretation. Human assessments also verify CPTM's effectiveness in accurately identifying label-relevant tokens within segments and providing insightful token-level interpretation.
The proliferation of misinformation on video-sharing platforms demands robust detection of video fake news. Existing methods struggle to integrate external world knowledge with internal multimodal cues, limiting their generalization and robustness. In this work, we propose TrueLens, a new framework for video fake news detection that gathers and consolidates dual-level evidence \zznotethrough three primary components, \ie, External Precedent Retriever, Adversarial Contrastor, and Internal Evidential Logic Fusion. At the external level, the External Precedent Retriever first decomposes the query video into textual, visual, and audio queries while leveraging multimodal large language models (MLLMs) to enhance the overall semantic representation. It then applies an entropy-guided multimodal retrieval mechanism to identify the two most similar reference videos from a gallery of real and fake samples, \ie, one real and one fake video. The Adversarial Contrastor integrates these references with the input video through contrastive attention, enhancing contextual reasoning. At the internal level, our Evidential Logic Fusion module aggregates multimodal signals from the Adversarial Contrastor to produce consistent, robust predictions. Extensive experiments on three benchmarks show that the proposed TrueLens consistently surpasses competitive baselines under both temporal and event settings by a clear margin, yielding up to a +21.60% F1 improvement under the event setting and achieving 93.73%, 90.64%, and 98.83% accuracy on the FakeSV, FakeTT, and FVC datasets under the temporal setting. The code for our project is available at https://github.com/JunyiChen-ai/TrueLens.
Multimodal information extraction (MIE) constitutes a set of essential tasks aimed at extracting structural information from Web texts with integrating images, to facilitate the structural construction of Web-based semantic knowledge. To address the expanding category set including newly emerging entity types or relations on websites, prior research proposed the zero-shot MIE (ZS-MIE) task which aims to extract unseen structural knowledge with textual and visual modalities. However, the ZS-MIE models are limited to recognizing the samples that fall within the unseen category set, and they struggle to deal with real-world scenarios that encompass both seen and unseen categories. The shortcomings of existing methods can be ascribed to two main aspects. On one hand, these methods construct representations of samples and categories within Euclidean space, failing to capture the hierarchical semantic relationships between the two modalities within a sample and their corresponding category prototypes. On the other hand, there is a notable gap in the distribution of semantic similarity between seen and unseen category sets, which impacts the generative capability of the ZS-MIE models. To overcome the above disadvantages, we delve into the generalized zero-shot MIE (GZS-MIE) task and propose the hyperbolic multimodal generative representation learning framework (HMGRL). The variational information bottleneck and autoencoder networks are reconstructed with hyperbolic space for modeling the multi-level hierarchical semantic correlations among samples and prototypes. Furthermore, the proposed model is trained with the unseen samples generated by the decoder, and we introduce the semantic similarity distribution alignment loss to enhance the model's generalization performance. Experimental evaluations on two benchmark datasets underscore the superiority of HMGRL compared to existing baseline methods.
Knowledge tracing (KT) aims to personalize online education on large-scale web-based platforms by modeling students' evolving knowledge states from their interaction sequences. However, most KT models rely on a single encoder architecture (e.g., self-attention or RNN), with fixed inductive biases that fails to capture the diversity of learning behaviors. Specifically, student learning unfolds across multiple timescales, and interaction sequences contain diverse frequency components ranging from short-term variations to long-term trends. Our data-driven analysis reveals that existing encoders exhibit characteristic frequency biases (e.g., self-attention tends to emphasize low-frequency patterns), highlighting the limitations of any single architecture. To address this problem, we propose FA-KT, a frequency-aware mixture of heterogeneous experts framework. FA-KT combines self-attention, Mamba, CNN, and LSTM experts, each with complementary frequency biases. A frequency-aware router analyzes each sequence's frequency characteristics and adaptively combines experts to create dynamic, personalized encoders for individual students. Across five benchmark datasets, FA-KT consistently outperforms 20 strong KT baselines in predicting future performance. Code is available at https://pykt.org/.
Zero-shot stance detection (ZSSD) aims to classify stances towards previously unseen targets without direct supervision on those topics during training. While recent approaches have explored various strategies, they often suffer from limited linguistic diversity or unstable semantic representations, restricting generalization to new domains. To address these challenges, we propose DPSD , a dynamic framework that integrates LLM-assisted data augmentation, multi-granularity feature fusion with contrastive learning, and adaptive prototype updating. Our method enriches the training data with both diverse targets and stylistically varied texts. A gate-controlled fusion mechanism combines deep contextualized features from BERT with shallow lexical patterns via TF-IDF, while contrastive learning refines the feature space by pulling similar instances closer and pushing dissimilar ones apart, thereby improving representation discriminability. Furthermore, we introduce a sliding-window prototype pool that dynamically maintains class-specific prototypes while preserving historical knowledge, ensuring stable and interpretable inference over time. We also incorporate LLM-calibrated semantic similarity as an auxiliary scorer for controlled reasoning. Experimental results on three benchmark datasets -SEM16, P-Stance, and VAST - show that DPSD achieves strong performance in various zero-shot settings, especially in cross-dataset and unseen target scenarios.
Large Language Model-based Recommender Systems (LRSs) have recently emerged as a new paradigm in sequential recommendation by directly adopting LLMs as backbones. While LRSs demonstrate strong knowledge utilization and instruction-following abilities, they have not been systematically studied under the long-standing long-tail problem. In this paper, we conduct an empirical study and reveal that LRSs face two distinct types of long-tail: i) prior long-tail, inherited implicitly from pretraining corpora, and ii) data long-tail, originating from skewed recommendation datasets. Our analysis shows that both contribute to the performance disparity between head and tail items, with the intersection of the two heads exhibiting an even stronger head effect. Nevertheless, the overall performance distribution in LRSs, especially on the tail, remains dominated by the data long-tail. To address this challenge, we propose Efficient Item-wise Sharpness-Aware Minimization (EISAM), a novel optimization framework that improves tail-item performance by adaptively regularizing the loss landscape at the item level. EISAM introduces an efficient penalty design that captures fine-grained item-specific sharpness while maintaining computational scalability for LLMs. In addition, we derive a generalization bound for EISAM. Our theoretical analysis shows that the bound decreases at a faster rate under our item-wise regularization, offering theoretical support for its effectiveness. Extensive experiments on three real-world datasets demonstrate that EISAM significantly boosts tail-item recommendation performance while preserving overall quality, establishing the first systematic solution to the long-tail problem in LRSs.
Large Language Models (LLMs) have demonstrated strong potential for generative recommendation by leveraging rich semantic knowledge. However, existing LLM-based recommender systems struggle to effectively incorporate collaborative filtering (CF) signals, due to a fundamental mismatch between item-level preference modeling in CF and token-level next-token prediction (NTP) optimization in LLMs. Prior approaches typically treat CF as contextual hints or representation bias, and resort to multi-stage training to reduce behavioral–semantic space discrepancies, leaving CF unable to explicitly regulate LLM generation. In this work, we propose Token-level Collaborative Alignment for Recommendation (TCA4Rec), a model-agnostic and plug-and-play framework that establishes an explicit optimization-level interface between CF supervision and LLM generation. TCA4Rec consists of (i) Collaborative Tokenizer, which projects raw item-level CF logits into token-level distributions aligned with the LLM token space, and (ii) Soft Label Alignment, which integrates these CF-informed distributions with one-hot supervision to optimize a soft NTP objective. This design preserves the generative nature of LLM training while enabling collaborative alignment with essential user preference of CF models. We highlight TCA4Rec is compatible with arbitrary traditional CF models and generalizes across a wide range of decoder-based LLM recommender architectures. Moreover, it provides an explicit mechanism to balance behavioral alignment and semantic fluency, yielding generative recommendations that are both accurate and controllable. Extensive experiments demonstrate that TCA4Rec consistently improves recommendation performance across a broad spectrum of CF models and LLM-based recommender systems. Our code is available at https://github.com/critical88/TCA4Rec
Cross-domain recommendation (CDR) offers an effective strategy for improving recommendation quality in a target domain by leveraging auxiliary signals from source domains. Nonetheless, emerging evidence shows that CDR can inadvertently heighten group-level unfairness. In this work, we conduct a comprehensive theoretical and empirical analysis to uncover why these fairness issues arise. Specifically, we identify two key challenges: (i) Cross-Domain Disparity Transfer, wherein existing group-level disparities in the source domain are systematically propagated to the target domain; and (ii) Unfairness from Cross-Domain Information Gain, where the benefits derived from cross-domain knowledge are unevenly allocated among distinct groups. To address these two challenges, we propose a Cross-Domain Fairness Augmentation (CDFA) framework composed of two key components. Firstly, it mitigates cross-domain disparity transfer by adaptively integrating unlabeled data to equilibrate the informativeness of training signals across groups. Secondly, it redistributes cross-domain information gains via an information-theoretic approach to ensure equitable benefit allocation across groups. Extensive experiments on multiple datasets and baselines demonstrate that our framework significantly reduces unfairness in CDR without sacrificing overall recommendation performance, while even enhancing it. Our source code is publicly available at https://github.com/weixinchen98/CDFA.
Cross-domain recommendation (CDR) addresses the data sparsity and cold-start problems in the target domain by leveraging knowledge from data-rich source domains. However, existing CDR methods often rely on domain-specific features or identifiers that lack transferability across different domains, limiting their ability to capture inter-domain semantic patterns. To overcome this, we propose \model, a semantics-driven framework for cross-domain sequential recommendation that leverages large language models (LLMs) to construct a unified semantic space. \model creates multi-view item features by integrating LLM-generated domain-agnostic semantics with domain-specific content, aligned by contrastive regularization. \model systematically creates LLM-generated domain-specific and domain-agnostic semantics, and employs adaptive fusion to generate unified preference representations. Furthermore, it aligns cross-domain behavior sequences with an adaptive fusion mechanism to synthesize interaction sequences from source, target, and mixed domains. Extensive experiments on real-world datasets show that \model consistently outperforms state-of-the-art baselines, demonstrating its effectiveness in capturing coherent intra-domain patterns while facilitating knowledge transfer across domains. Our code is available online. https://github.com/huangshanqiang/SemaCDR.
Large Language Models (LLMs) have been widely applied across multiple domains for their broad knowledge and strong reasoning capabilities. However, applying them to recommendation systems is challenging since it is hard for LLMs to extract user preferences from large, sparse user-item logs, and real-time per-user ranking over the full catalog is too time-consuming to be practical. Moreover, many existing recommender systems focus solely on ranking items while overlooking explanations, which could help improve predictive accuracy and make recommendations more convincing to users. Inspired by recent works that achieve strong recommendation performance by forecasting near-term item popularity, we propose TRAIL (TRend and explAnation Integrated Learner). TRAIL is a fine-tuned LLM that jointly predicts short-term item popularity and generates faithful natural-language explanations. It employs contrastive learning with positive and negative pairs to align its scores and explanations with structured trend signals, yielding accurate and explainable popularity predictions. Extensive experiments show that TRAIL outperforms strong baselines and produces coherent, well-grounded explanations.
Large language models (LLMs), owing to their extensive open-domain knowledge and semantic reasoning capabilities, have been increasingly integrated into recommender systems (RS). However, a substantial gap remains between the pre-training objectives of LLMs and the specific requirements of recommendation tasks. To address this gap, supervised fine-tuning (SFT) is commonly performed on specially curated recommendation datasets to further enhance their predictive ability. Despite its success, SFT exhibits a critical limitation: it induces Context Bias, whereby the model over-relies on auxiliary tokens—such as task descriptions and prefix-generated tokens—while underutilizing core user interaction tokens that encode user-specific preferences. This bias not only undermines recommendation accuracy but also raises unfairness concerns. To address this issue, we propose Group Distributionally Robust Optimization-based Tuning (GDRT), a novel fine-tuning paradigm that enforces consistent model performance across token groups with varying degrees of relevance to auxiliary tokens. By adaptively upweighting underperforming groups, typically those weakly correlated with auxiliary tokens, GDRT shifts the model's attention from superficial auxiliary cues to informative user interaction tokens, thereby mitigating context bias. Extensive experiments conducted on three public datasets demonstrate that GDRT effectively mitigates context bias, yielding substantial improvements in recommendation accuracy (with an average NDCG@10 gain of 24.29%) and significantly enhancing recommendation fairness. The code is available at https://github.com/WANGBohaO-jpg/GDRT.
Expert-sharing patterns have emerged as promising paradigms for multi-task learning (MTL) in recommender systems, enabling efficient resource allocation and dynamic modeling of diverse tasks. In this paper, we observe that high-gating (leader) and low-gating (auxiliary) experts play distinct roles in MTL: leader experts dominate task-specific predictions, while auxiliary experts, despite their lower gating scores, often contain complementary knowledge that can enhance model performance. However, critical challenges persist: how to effectively identify and utilize the knowledge of the leader and auxiliary experts in a synergistic manner? To address this, we propose a novel Dynamic Experts Synergy (DES) mechanism that integrates Entropy-driven Experts Classification (EEC) and Multi-view Knowledge Recycle (MVKR). EEC dynamically partitions experts into leader and auxiliary groups by analyzing task-specific prediction and gating entropy, enabling adaptive allocation aligned with real-time task difficulty. MVKR effectively revisits knowledge from auxiliary experts through utility, diversity, and task-relatedness perspectives, ensuring comprehensive knowledge utilization. Extensive experiments on five datasets demonstrate the superiority of our DES against state-of-the-art methods.