Toxicity detection mitigates the dissemination of toxic content (e.g., hateful comments, posts, and messages within online social actions) to safeguard a healthy online social environment. However, malicious users persistently develop evasive perturbations to disguise toxic content and evade detectors. Traditional detectors or methods are static over time and are inadequate in addressing these evolving evasion tactics. Thus, continual learning emerges as a logical approach to dynamically update detection ability against evolving perturbations. Nevertheless, disparities across perturbations hinder the detector's continual learning on perturbed text. More importantly, perturbation-induced noises distort semantics to degrade comprehension and also impair critical feature learning to render detection sensitive to perturbations. These amplify the challenge of continual learning against evolving perturbations. In this work, we present ContiGuard, the first framework tailored for continual learning of the detector on time-evolving perturbed text (termed continual toxicity detection) to enable the detector to continually update capability and maintain sustained resilience against evolving perturbations. Specifically, to boost the comprehension, we present an LLM powered semantic enriching strategy, where we dynamically incorporate possible meaning and toxicity-related clues excavated by LLM into the perturbed text to improve the comprehension. To mitigate non-critical features and amplify critical ones, we propose a discriminability driven feature learning strategy, where we strengthen discriminative features while suppressing the less-discriminative ones to shape a robust classification boundary for detection. Additionally, we introduce a historical capability replay strategy to preserve previously learned features via feature alignment to alleviate capability forgetting. To the best of our knowledge, this work is the first study on continual toxicity detection against time-evolving evasive perturbed text. Extensive experiments prove the superior performance of ContiGuard over both existing detectors and continual methods. Code and dataset are available at https://github.com/khk-abc/ContiGuard. Warning: This paper contains discussions of harmful content that may be disturbing to some readers.
论文检索
输入标题、作者或关键词，从 3,225 篇学术成果中精准定位
Detecting social bots is critical to ensuring the security of online discourse and maintaining trust in social networks. Early feature-based and text-based methods often fail against bots that mimic human behavior, and graph-based approaches have emerged to better exploit structural signals. However, most existing Graph Neural Networks (GNNs) still focus on pairwise connections, overlooking higher-order relational patterns, and their multi-relation fusion strategies are typically simplistic, ignoring dependencies between relations and user-specific preferences. To overcome these limitations, we propose MPS-Bot, a model that integrates higher-order structure modeling with user-specific cross-relation dependency learning. MPS-Bot introduces a simplex convolutional layer that leverages simplexes derived from network structures to capture group coordination patterns beyond pairwise connections. In addition, a cross-relation dependency attention mechanism adaptively fuses relation-specific representations according to each user's relational preferences, leading to more discriminative and robust multi-relation representations. Extensive experiments on two widely used Twitter bot detection benchmarks, MGTAB and TwiBot-22, show that MPS-Bot generally outperforms state-of-the-art baselines. These findings highlight the effectiveness of higher-dimensional message passing over simplexes as a powerful approach to unmasking bots in social networks.
Spamming activities (e.g., fake reviews, click farming, and deceptive content promotion) are increasingly conducted through collusive groups that exploit collective dynamics to manipulate platform metrics and mislead users, posing serious threats to the fairness, credibility, and functionality of online systems. To counteract these harmful behaviors, the task of spam detection has emerged as a critical area of research. However, existing detection methods generally remain limited in three key aspects: (i) They treat detection as a standard classification task, where representation learning and optimization are loosely coupled and suboptimal for capturing complex behaviors; (ii) They rely primarily on individual-level representation modeling, making it difficult to detect collective cheating strategies; (iii) They lack dedicated objective functions explicitly designed to characterize group-level spamming activities. To overcome these limitations, we introduce a collusion-aware Set-level learning framework (SetDet) that redefines the spam Detection task as a unified setwise optimization problem. Our approach offers three core advantages: (i) It enables end-to-end optimization by jointly learning representations and performing detection in a single, integrated process; (ii) It incorporates a model-level design for collusion representation, effectively capturing the temporal and relational patterns of coordinated spam; (iii) It pioneers a dedicated set-level optimization criterion that aligns closely with the structural characteristics of group-based cheating behaviors and accounts for class imbalance in real-world scenarios. Extensive experiments confirm the generalizability and superior performance of our framework across diverse spam scenarios and collusion strategies.
The widespread proliferation of online content has intensified concerns about clickbait, deceptive or exaggerated headlines designed to attract attention. While Large Language Models (LLMs) offer a promising avenue for addressing this issue, their effectiveness is often hindered by Sycophancy, a tendency to produce reasoning that matches users' beliefs over truthful ones, which deviates from instruction-following principles. Rather than treating sycophancy as a flaw to be eliminated, this work proposes a novel approach that initially harnesses this behavior to generate contrastive reasoning from opposing perspectives. Specifically, we design a Self-renewal Opposing-stance Reasoning Generation (SORG) framework that prompts LLMs to produce high-quality ''agree'' and ''disagree'' reasoning pairs for a given news title without requiring ground-truth labels. To utilize the generated reasoning, we develop a local Opposing Reasoning-based Clickbait Detection (ORCD) model that integrates three BERT encoders to represent the title and its associated reasoning. The model leverages contrastive learning, guided by soft labels derived from LLM-generated credibility scores, to enhance detection robustness. Experimental evaluations on three benchmark datasets demonstrate that our method consistently outperforms LLM prompting, fine-tuned smaller language models, and state-of-the-art clickbait detection baselines. Our code is available in https://github.com/126541/ORCD.
Privacy leakage on Web-based platform has become a critical challenge as users continually share personal information through online social networks, health tracking platforms, and other Web services. Information recipients and third-party can progressively aggregate shared content across the Web, enabling increasingly accurate profiling of individuals. However, existing studies typically treat each disclosure as independent, overlooking the cumulative privacy risks that arise in continual information sharing. In addition, the subjective cognition of both users and adversaries in Web environments, where users and adversaries can dynamically adapt based on observable actions, remains underexplored. To address these challenges, we propose a dynamic Stackelberg game model for continual information sharing scenarios, where user's sequential privacy decisions are optimized to balance privacy protection and data utility. The model explicitly captures the cognitive behaviors of both the user and the adversary, allowing their subjective perceptions to shape the Stackelberg equilibrium. Building on this formulation, we develop a reinforcement learning-based algorithm to derive approximately optimal strategies for mitigating privacy leakage in the context of continual information sharing. Experiments on real-world datasets demonstrate that our method significantly reduces cumulative privacy risks while preserving the utility of shared content. The proposed model further provides actionable insights for the design of privacy-enhancing technologies and web platform policies.
The proliferation of face forgery content on the Web poses a severe threat to online trust, social media security, and the credibility of digital information. Existing detection approaches often fail to generalize across diverse forgery types and unseen scenarios commonly encountered in web-scale applications. Recent studies have utilized visual large language models (VLMs) to answer not only ''Is this face a forgery?'' but also ''Why is the face a forgery?'' These studies introduced forgery-related attributes, such as forgery location and type, to construct deepfake VQA datasets and train VLMs, achieving high accuracy while providing human-understandable explanatory text descriptions. However, these methods still have limitations. For example, they do not fully leverage face quality-related attributes, which are often abnormal in forged faces, and they lack effective training strategies for forgery-aware VLMs. In this paper, we extend the VQA dataset to create DD-VQA+, which features a richer set of attributes and a more diverse range of samples. Furthermore, we introduce a novel forgery detection framework, MGFFD-VLM, which integrates an Attribute-Driven Hybrid LoRA Strategy to enhance the capabilities of Visual Large Language Models (VLMs). Additionally, our framework incorporates Multi-Granularity Prompt Learning and a Forgery-Aware Training Strategy. By transforming classification and forgery segmentation results into prompts, our method not only improves forgery classification but also enhances interpretability. To further boost detection performance, we design multiple forgery-related auxiliary losses. Experimental results demonstrate that our approach surpasses existing methods in both text-based forgery judgment and analysis, achieving superior accuracy.
The proliferation of online hate speech poses a significant threat to the harmony of the web. While explicit hate is easily recognized through overt slurs, implicit hate speech is often conveyed through sarcasm, irony, stereotypes, or coded language---making it harder to detect. Existing hate speech detection models, which predominantly rely on surface-level linguistic cues, fail to generalize effectively across diverse stylistic variations. Moreover, hate speech spread on different platforms often targets distinct groups and adopts unique styles, potentially inducing spurious correlations between them and labels, further challenging current detection approaches. Motivated by these observations, we hypothesize that the generation of hate speech can be modeled as a causal graph involving key factors: contextual environment, creator motivation, target, and style. Guided by this graph, we propose CADET, a causal representation learning framework that disentangles hate speech into interpretable latent factors and then controls confounders, thereby isolating genuine hate intent from superficial linguistic cues. Furthermore, CADET allows counterfactual reasoning by intervening on style within the latent space, naturally guiding the model to robustly identify hate speech in varying forms. CADET demonstrates superior performance in comprehensive experiments, highlighting the potential of causal priors in advancing generalizable hate speech detection.
Graph-level anomaly detection (GLAD) is a critical task to identify graphs with abnormal properties in various domains, ranging from fraudulent social networks to malicious botnets on online platforms. The dominant paradigm for existing GLAD detectors has been partially unsupervised, relying on training data composed exclusively of normal samples. However, this partially unsupervised paradigm inevitably requires a costly expert filtering process to ensure the training data is free of anomalies. This creates a significant gap between current approaches and the real-world necessity of a fully unsupervised paradigm, which involves training a model directly on real-world data ''as-is'', with its inherent mix of normal and anomalous samples. To bridge this gap, we incorporate uncertainty learning into GLAD to promote fully unsupervised learning. We propose two frameworks: Score Uncertainty Learning (SUL) and Graph-data Uncertainty Learning (GUL). Specifically, SUL enhances existing GLAD detectors by modeling uncertainty through Gaussian distributions over the detectors' predictions, adaptively attenuating the influence of potential anomalies. GUL is an end-to-end framework that iteratively optimizes anomaly detection and uncertainty modeling via an Expectation-Maximization algorithm. In addition, we develop a dedicated loss that utilizes potential anomalies to enhance the effectiveness and robustness of GUL. Empirical results on sixteen benchmark datasets, covering real-world graphs from social networks and online platforms, demonstrate the superiority of our methods and highlight the promise of incorporating uncertainty into fully unsupervised GLAD.
Reinforcement learning-based recommendation systems (RLRS) are increasingly favored for their ability to leverage online interactive feedback, enabling adaptive and personalized decision-making. In this setting, user feedback serves as both a behavioral signal and an optimization target, making it essential for policy learning. However, collecting such feedback, e.g., clicks, ratings, and engagement traces, raises serious privacy concerns, posing critical challenges for value estimation, online adaptation, and privacy protection. In this paper, we propose FeedGuard, a critic-guided reinforcement learning framework with privacy-preserving feedback. FeedGuard enhances trajectory modeling via critic guidance, enables joint online fine-tuning with effective exploration–exploitation tradeoffs, and enforces end-to-end privacy protection across the feedback lifecycle via split federated learning and differential privacy. We further provide a formal analysis of its differential privacy guarantees. Extensive experiments on four public recommendation datasets and the VirtualTB platform show that FeedGuard performs well in both offline and online settings, while maintaining rigorous privacy guarantees with minimal degradation.
Online market platforms play an increasingly powerful role in the economy. An empirical phenomenon is that platforms, such as Amazon, Apple, and DoorDash, also enter their own marketplaces, imitating successful products developed by third-party sellers. We formulate a Stackelberg model, where the platform acts as the leader by committing to an entry policy: when will it enter and compete on a product? We study this model through a theoretical and computational framework. We begin with a single seller, and consider different kinds of policies for entry. We characterize the seller's optimal explore-exploit strategy via a Gittins-index policy, and give an algorithm to compute the platform's optimal entry policy. We then consider multiple sellers, to account for competition and information spillover. Here, the Gittins-index characterization fails, and we employ deep reinforcement learning to examine seller equilibrium behavior. Our findings highlight the incentives that drive platform entry and seller innovation, consistent with empirical evidence from markets such as Amazon and Google Play, with implications for regulatory efforts to preserve innovation and market diversity.
In online advertising, the inherent complexity and dynamic nature of advertising environments necessitate the use of auto-bidding services to assist advertisers in bid optimization. The complexity escalates in multi-channel scenarios, where effective allocation of budgets and constraints across channels with distinct behavioral patterns becomes critical for optimizing return on investment. Current approaches predominantly employ either optimization-based strategies or reinforcement learning (RL) techniques. However, optimization-based methods lack the flexibility to adapt to dynamic market conditions, while RL-based approaches struggle to capture essential historical dependencies and observational patterns within the constraints of Markov Decision Process (MDP) frameworks. To address these limitations, we propose AHBid, an Adaptable Hierarchical Bidding framework that integrates generative planning with real-time control. The framework employs a high-level generative planner utilizing diffusion models to dynamically allocate budgets and constraints through effective capture of historical context and temporal patterns. We introduce a constraint enforcement mechanism to ensure compliance with specified constraints, complemented by a trajectory refinement mechanism that enhances adaptability to environmental changes through historical data utilization. The system further incorporates a control-based bidding algorithm that synergistically combines historical knowledge with real-time information, significantly improving both adaptability and operational efficacy. Extensive experiments are conducted using both large-scale offline datasets and online A/B tests, demonstrating the effectiveness of AHBid by yielding a 13.57% increase in overall return compared to existing baselines.
The growing scale of ad auctions on online advertising platforms has intensified competition, making manual bidding impractical and necessitating auto-bidding to help advertisers achieve their economic goals. Current auto-bidding methods have evolved to use offline reinforcement learning or generative methods to optimize bidding strategies, but they can sometimes behave counterintuitively due to the black-box training manner and limited mode coverage of datasets, leading to challenges in understanding task status and generalization in dynamic ad environments. Large language models (LLMs) offer a promising solution by leveraging prior human knowledge and reasoning abilities to improve auto-bidding performance. However, directly applying LLMs to auto-bidding faces difficulties due to the need for precise actions in competitive auctions and the lack of specialized auto-bidding knowledge, which can lead to hallucinations and suboptimal decisions. To address these challenges, we propose a hierarchical Large auto-Bidding Model (LBM) to leverage the reasoning capabilities of LLMs for developing a superior auto-bidding strategy. This includes a high-level LBM-Think model for reasoning and a low-level LBM-Act model for action generation. Specifically, we propose a dual embedding mechanism to efficiently fuse two modalities, including language and numerical inputs, for language-guided training of the LBM-Act; then, we propose an offline reinforcement fine-tuning technique termed GQPO for mitigating the LLM-Think's hallucinations and enhancing decision-making performance without simulation or real-world rollout like previous multi-turn LLM-based methods. Experiments demonstrate the superiority of a generative backbone based on our LBM, especially in an efficient training manner and generalization ability.
Optimizing the advertiser's cumulative value of winning impressions under budget constraints poses a complex challenge in online advertising, under the paradigm of AI-Generated Bidding (AIGB). Advertisers often have personalized objectives but limited historical interaction data, resulting in few-shot scenarios where traditional reinforcement learning (RL) methods struggle to perform effectively. Large Language Models (LLMs) offer a promising alternative for AIGB by leveraging their in-context learning capabilities to generalize from limited data. However, they lack the numerical precision required for fine-grained optimization. To address this limitation, we introduce GRPO-Adaptive, an efficient LLM post-training strategy that enhances both reasoning and numerical precision by dynamically updating the reference policy during training. Built upon this foundation, we further propose DARA, a novel dual-phase framework that decomposes the decision-making process into two stages: a few-shot reasoner that generates initial plans via in-context prompting, and a fine-grained optimizer that refines these plans using feedback-driven reasoning. This separation allows DARA to combine LLMs' in-context learning strengths with precise adaptability required by AIGB tasks. Extensive experiments on both real-world and synthetic data environments demonstrate that our approach consistently outperforms existing baselines in terms of cumulative advertiser value under budget constraints.
AI is no longer just a component of Web systems; it is increasingly shaping the experiences users have online. From search and recommendation to conversational and generative interfaces, AI is redefining how people interact with content at Web scale. In this keynote, I reflect on how recent advances in AI, including deep learning and generative models, are reshaping the design space of Web technologies and the systems that support them. Drawing on insights from developing AI-driven systems at Spotify, I discuss how search and recommendation are evolving into interactive, intent-aware, and user-controllable experiences that support exploration and discovery. The talk highlights emerging system paradigms, deployment challenges, and open research questions around building such experiences at scale, and reflects on the implications for the design of future Web systems and interactions.
We present an AI-powered interactive multimodal system that enriches digital image accessibility for blind and visually impaired (BVI) users. Our demonstration showcases two application domains: (1) an educational digital book, and (2) an online shopping interface. In both use-cases, users can virtually feel material textures (leather, wood, etc.) and engage in voice-driven inquiry about images. The system integrates state-of-the-art AI components – including voice-to-voice conversational agents, vision models for object segmentation, and a custom 16-actuator vibrotactile display – to provide multimodal feedback (haptic vibrations, spoken descriptions, and audio cues). The result is an inclusive technology with significant societal benefit, empowering BVI users to learn and shop more independently through natural multimodal interactions.
Post-training for Large Language Models (LLMs) can be mainly categorized into offline Supervised Fine-Tuning (SFT) for knowledge acquisition and online Reinforcement Fine-Tuning (RFT) for adaptive refinement. Current state-of-the-art approaches typically employ a sequential cold-start pipeline (SFT-Then-RFT). However, we argue that this disjoint transition imposes an “alignment tax", leading to catastrophic forgetting and reward hacking during the unregularized exploration phase. In this work, we advocate for Joint Online-Offline Fine-Tuning as a superior paradigm that breaks the convention of restricting offline data to SFT and online data to RFT. By integrating full offline response generation with online rollouts—particularly within the realm of Reinforcement Learning with Verifiable Rewards (RLVR)—this approach mitigates the limitations of isolated training phases. We provide the first comprehensive survey focusing specifically on the synchronization of data provenance. We introduce a novel taxonomy for related works, analyze their theoretical advantages in balancing stability with plasticity, and outline a roadmap for next-generation post-training frameworks.
Recommenders are significantly shaping online information consumption. While effective at personalizing content, these systems increasingly face criticism for propagating irrelevant, unwanted, and even harmful recommendations. Such content degrades user satisfaction and contributes to significant societal issues, including misinformation, radicalization, and erosion of user trust. Although platforms offer mechanisms to mitigate exposure to undesired content, these mechanisms are often insufficiently effective and slow to adapt to users' feedback. This paper introduces an intuitive, model-agnostic, and distribution-free method that uses conformal risk control to provably bound unwanted content in personalized recommendations by leveraging simple binary feedback on items. We also address a limitation of traditional conformal risk control approaches, i.e., the fact that the recommender can provide a smaller set of recommended items, by leveraging implicit feedback on consumed items to expand the recommendation set while ensuring robust risk mitigation. Our experimental evaluation on data coming from a popular online video-sharing platform demonstrates that our approach ensures an effective and controllable reduction of unwanted recommendations with minimal effort.
Chlor-alkali production is a large-scale industrial process whose operating conditions and equipment states evolve over time. Its process optimization requires ongoing trade-offs among conflicting objectives such as product yield, energy consumption, and equipment life. Existing optimization approaches are typically static and must be re-optimized after environmental changes, limiting their real-world applicability. In this work, we model the problem as a dynamic multi-objective sequential decision-making problem that continuously tracks a time-varying Pareto set under changing conditions. We propose MORL-CA, a multi-objective reinforcement learning framework that integrates offline pretraining on historical data with constrained online policy refinement. MORL-CA introduces a state-aware adaptive objective weighting mechanism within a multi-critic actor-critic architecture, enabling localized Pareto-improving policy updates while satisfying operational and safety constraints. Extensive experiments in an environment conducted from real chlor-alkali data demonstrate that MORL-CA achieves superior Pareto solution quality and smoother adaptation to dynamics compared with state-of-the-art multi-objective optimizers and MORL baselines.
Motivated by the problem of assigning mediators to cases in the Kenyan judicial system, we study an online resource allocation problem where incoming tasks (cases) must be immediately assigned to available, capacity-constrained resources (mediators). The resources differ in their quality, which may need to be learned. In addition, resources can only be assigned to a subset of tasks that overlaps to varying degrees with the subset of tasks other resources can be assigned to. The objective is to maximize task completion while satisfying soft capacity constraints across all the resources. The scale of the real-world problem poses substantial challenges, since there are over 2000 mediators, and a multitude of combinations of geographic locations (87) and case types (12) that each mediator is qualified to work on. Together, these features—unknown quality of new resources (newly onboarded mediators), soft capacity constraints (due to the mandate to assign cases without delay), and high-dimensional state space—make existing scheduling and resource allocation algorithms either inapplicable or inefficient. We formalize the problem in a tractable manner, using a quadratic program formulation for assignment and a multi-agent bandit style framework for learning. We demonstrate the key properties and advantages of our new algorithm, SMaRT (Selecting Mediators that are Right for the Task), compared with baselines on some stylized instances of the mediator allocation problem. We then turn to considering its application to real-world data on cases and mediators from the Kenyan Judiciary. SMaRT outperforms baselines and allows for controlling the tradeoff between the strictness of the capacity constraints and overall case resolution rates, both in situations where mediator quality is known beforehand and when the problem is bandit-like in that learning is part of the problem definition. On the strength of these results, we plan to conduct a randomized controlled trial in which we deploy SMaRT in the Judiciary's mediation management system.
Online learning systems provide accessible learning opportunities for blind or low-vision students. To support access to complex scientific materials, the speech models used in these systems need to deliver accurate scientific formula verbalization. While recent large speech language models (LSLMs) provide remarkable low-latency streaming capabilities, their potential for scientific formula verbalization remains underexplored. In this paper, we propose Formula-Speech, the first end-to-end LSLM designed for scientific formula verbalization. Specifically, we construct two high-quality scientific formula datasets with educational experts to align speech models with scientific formula verbalization patterns. We then adopt a lightweight and effective two-stage training framework, combining supervised fine-tuning for basic formula-to-speech alignment with reinforcement learning guided by a custom reward function to optimize for human-preferred verbalization. Experimental results show that our model significantly improves the verbalization performance of LSLMs and achieves state-of-the-art results across multiple scientific domains. Our code and datasets are available at https://github.com/ai4ed/FormulaSpeech.