Harmful online content, including hate speech, fraud, and phishing, is increasingly disseminated in obfuscated forms designed to evade detection. This creates an urgent need for accurate and efficient real-time de-obfuscation methods to protect users and maintain trust. Existing obfuscation detection methods rely on large auto-regressive models and byte-level fallback tokenizers, which are hindered by slow inference speeds and face difficulties in handling graphemes with multiple code points and out-of-vocabulary (OOV) processing. This study proposes Bidirectionally Aligned Next-Token Denoising ( BIND ), which integrates character-level token alignment with a novel attention technique to enable precise and efficient corrections at fixed positions. Experiments conducted on a public dataset of obfuscated harmful text demonstrate that BIND outperforms existing methods. BIND has shown strong robustness against various text-based visual, phonetic, and semantic perturbations, proving particularly resilient against emojis and other OOV elements. This research highlights how a task-specific small language model can outperform larger ones, offering a practical solution for real-time harmful content mitigation and contributing to the development of a safer and more responsible web.
论文检索
输入标题、作者或关键词，从 954 篇学术成果中精准定位
Previous studies on visual customization primarily rely on the objective alignment between various control signals (e.g., language, layout and canny) and the edited images, which largely ignore the subjective emotional contents, and more importantly lack general-purpose foundation models for affective visual customization. With this in mind, this paper proposes an LLM-centric Affective Visual Customization (L-AVC) task, which focuses on generating images within modifying their subjective emotions via Multimodal LLM. Further, this paper contends that how to make the model efficiently align emotion conversion in semantic (named inter-emotion semantic conversion) and how to precisely retain emotion-agnostic contents (named exter-emotion semantic retaining) are rather important and challenging in this L-AVC task. To this end, this paper proposes an Efficient and Precise Emotion Manipulating (EPEM) approach for editing subjective emotions in images. Specifically, an Efficient Inter-emotion Converting (EIC) module is tailored to make the LLM efficiently align emotion conversion in semantics before and after editing, followed by a Precise Extra-emotion Retaining (PER) module to precisely retain the emotion-agnostic contents. Comprehensive experimental evaluations on our constructed L-AVC dataset demonstrate the great advantage of the proposed EPEM approach to the L-AVC task over several state-of-the-art baselines. This justifies the importance of emotion information for L-AVC and the effectiveness of EPEM in efficiently and precisely manipulating such information.
IConMoE: Modeling Intents of Misinformation using Concept Activation Vector-based Mixture of Experts
Understanding the intent behind misinformation helps reveal distinct patterns and underlying motives, enabling more accurate and efficient detection. However, existing approaches often struggle to effectively model the interplay of multiple coexisting intents within claims, and typically require additional LLM inferences to infer intent for each sample. In this paper, we introduce an intent-informed concept activation vector-based mixture-of-experts framework, iConMoE, which models latent intents in an unsupervised manner and predicts the veracity of claims in an end-to-end fashion, without the need for additional LLM inferences. Evaluated on three benchmark datasets, our method consistently outperforms both baseline and state-of-the-art models. We also conduct comprehensive ablation studies and error analyses to highlight the contributions of individual components and identify key challenges. In addition, we provide a qualitative analysis of the learned intents across samples to offer deeper insights into the model's reasoning.
Detecting harmful memes is crucial for safeguarding the integrity and harmony of online environments, yet existing detection methods are often resource-intensive, inflexible, and lacking explainability, limiting their applicability in assisting real-world web content moderation. We propose U-CoT+, a resource-efficient framework that prioritizes accessibility, flexibility and transparency in harmful meme detection by fully harnessing the capabilities of lightweight unimodal large language models (LLMs). Instead of directly prompting or fine-tuning large multimodal models (LMMs) as black-box classifiers, we avoid immediate reasoning over complex visual inputs but decouple meme content recognition from meme harmfulness analysis through a high-fidelity meme-to-text pipeline, which collaborates lightweight LMMs and LLMs to convert multimodal memes into natural language descriptions that preserve critical visual information, thus enabling text-only LLMs to ''see'' memes by ''reading''. Grounded in textual inputs, we further guide unimodal LLMs' reasoning under zero-shot Chain-of-Thoughts (CoT) prompting with targeted, interpretable, context-aware, and easily obtained human-crafted guidelines, thus providing accountable step-by-step rationales, while enabling flexible and efficient adaptation to diverse sociocultural criteria of harmfulness. Extensive experiments on seven benchmark datasets show that U-CoT+ achieves performance comparable to resource-intensive baselines, highlighting its effectiveness and potential as a scalable, explainable, and low-resource solution to support harmful meme detection.
Skin-toned emojis are crucial for fostering personal identity and social inclusion in online communication. As AI models, particularly Large Language Models (LLMs), increasingly mediate interactions on web platforms, the risk that these systems perpetuate societal biases through their representation of such symbols is a significant concern. This paper presents the first large-scale comparative study of bias in skin-toned emoji representations across two distinct model classes. We systematically evaluate dedicated emoji embedding models (emoji2vec, emoji-sw2v) against four modern LLMs (Llama, Gemma, Qwen, and Mistral). Our analysis first reveals a critical performance gap: while LLMs demonstrate robust support for skin tone modifiers, widely-used specialized emoji models exhibit severe deficiencies. More importantly, a multi-faceted investigation into semantic consistency, representational similarity, sentiment polarity, and core biases uncovers systemic disparities. We find evidence of skewed sentiment and inconsistent meanings associated with emojis across different skin tones, highlighting latent biases within these foundational models. Our findings underscore the urgent need for developers and platforms to audit and mitigate these representational harms, ensuring that AI's role on the web promotes genuine equity rather than reinforcing societal biases.
The dominant paradigm for interacting with the Web has long relied on menu-driven navigation and complex graphical structures that can impose a high cognitive load on users. Recent advances in Large Language Models (LLMs) enable flexible NL understanding, yet purely conversational systems lack the stability and visual affordances required for complex tasks. This paper introduces Mirage, a framework that bridges these paradigms by making graphical user interface (GUI) components interpretable and orchestrable by LLMs, enabling on-the-fly construction of simplified UIs tailored to user requests. Developers declare application functions as reusable components with semantic descriptions and typed schemas; Mirage maps user goals, expressed in natural language, to the most relevant components and renders them as interactive interfaces. We implemented Mirage as a Next.js framework and validated it in a user study comparing the new paradigm to the navigation of a traditional website. Results show that Mirage significantly reduces task completion time, lowers cognitive workload, and achieves higher usability without compromising accuracy. Our approach also highlights new design directions toward supporting on-the-fly building of goal-driven interfaces.
A central feature of many deliberative processes, such as citizens' assemblies and deliberative polls, is the opportunity for participants to engage directly with experts. While participants are typically invited to propose questions for expert panels, only a limited number can be selected due to time constraints. This raises the challenge of how to choose a small set of questions that best represent the interests of all participants. We introduce an auditing framework for measuring the level of representation provided by a slate of questions, based on the social choice concept known as justified representation (JR). We present the first algorithms for auditing JR in the general utility setting, with our most efficient algorithm achieving a runtime of O(mnłog n), where n is the number of participants and m is the number of proposed questions. We apply our auditing methods to historical deliberations, comparing the representativeness of (a) the actual questions posed to the expert panel (chosen by a moderator), (b) participants' questions chosen via integer linear programming, (c) summary questions generated by large language models (LLMs). Our results highlight both the promise and current limitations of LLMs in supporting deliberative processes. By integrating our methods into an online deliberation platform that has been used for over hundreds of deliberations across more than 50 countries, we make it easy for practitioners to audit and improve representation in future deliberations.
Since the first online ad in 1994, advertising has grown into a vast ecosystem delivering billions of ads daily. Advertisements are everywhere on the Web: search engines promote results, most websites display ads, some require users to accept ads as a condition for access, video streaming services fund their infrastructure through an increasing volume of ads, and much of the gaming industry has adopted ad-based revenue models. In exchange for free access to a wide range of content, web users sacrifice their privacy and pay with personal data to enable targeted marketing, a trade-off widely studied in literature. We argue that the ad ecosystem imposes an additional, overlooked, cost on web users: energy consumption. In this paper, we present a large-scale analysis of the client-side energy consumption of ads, including their associated tracking mechanisms. We design a robust methodology aimed at realistically modeling user behavior and monitoring CPU activity. Through measurements of 724,994 website visits across diverse devices, we study the energy implications of consenting to tracking and of blocking ads through browser and network-based software. We find that consent via cookie banners increases energy consumption by a median 2.57% across all websites and devices. We also show that websites relying on real-time bidding increase median energy consumption by 33.98%. Finally, although ad-blocking solutions consume energy due to their filtering processes, we observe a median reduction of 9.62% in client-side energy consumption when using uBlock Origin.
Visual data from the Web power image classifiers, which underpin web services, including recommendation and moderation. However, Web data contain spurious correlations and social biases, and neural networks tend to learn biases in data. This can reinforce unfairness in web services and the web data, leading to a vicious cycle. In image classification, networks learn bias attributes for a specific class when most images contain the same attribute for a given class. Hence, training a fair and debiased classifier from a biased dataset demands handling imbalance between a majority of images with bias attributes (bias-aligned samples) and a minority without (bias-conflict samples). In this work, we introduce BiasEdit, a modular framework that automatically detects bias attributes from the original dataset and edits them to construct a debiased dataset. Specifically, BiasEdit first detects unknown bias attributes via statistical dependence and mutual information analysis of visual–linguistic representations, and then explicitly edits those attributes using text-guided image editing to generate realistic bias-conflict samples. Unlike prior works that assume known bias attributes or rely on synthetic mixing, our method operates without manual annotations and leverages off-the-shelf vision–language and editing models. BiasEdit addresses a fundamental challenge in Web-sourced visual AI, mitigating dataset-induced bias and achieving state-of-the-art debiasing performance even when training data are fully biased.
Single-expert fake news detectors, such as Graph Neural Networks (GNNs) and Large Language Models (LLMs), increasingly struggle to counter the diversifying camouflage tactics of modern adversaries, which range from semantic (e.g., mimicking writing styles) to structural (e.g., manipulating propagation paths). To address this, existing methods attempt to build a hybrid model by sequentially incorporating GNNs and LLMs; however, such hybridization blurs the distinction between experts and prevents critical cross-validation. In addition, existing methods rely heavily on vast labeled data, which is costly to acquire, particularly for fake news samples. In this paper, we propose a Semi-supervised Mixture of Experts framework for Fake news detection, namely S2MOE-F. The core idea of S2MOE-F is to establish a robust defense against multifaceted camouflage by cross-validating the complementary judgments of two independent experts, GNN and LLM. On the one hand, S2MOE-F drives experts' judgments by using a One-Class Classification (OCC) objective, which constrains true news within a compact hypersphere and identifies samples outside this boundary as fake, reducing reliance on scarce fake news labels. On the other hand, S2MOE-F generates high-confidence pseudo-labels based on consensus or divergence between experts to exploit abundant unlabeled data. In addition, we propose a novel reinforcement learning (RL)-based routing policy that dynamically determines the dominant expert for input samples without explicit supervision. Finally, we design a disentangled masked Transformer to ensure experts' specialization by reducing inter-expert redundancy. Extensive experiments on real-world datasets sourced from Web platforms and social media demonstrate the superior performance of S2MOE-F.
Doxing refers to the disclosure of personal information without consent, has evolved from sporadic acts of online vigilantism into a structured and commodified practice. In Chinese cyberspace, this shift has produced Doxing-as-a-Service (DaaS), a commercial model in which personal data is retrieved, organized, and traded as on-demand products. This industrialization of privacy violation lowers the barriers to doxing and amplifies its social harms, posing new challenges for building a responsible and safe web. Yet little is known about how DaaS operates or sustains itself, motivating our systematic, data-driven examination of its ecosystem and practices. This paper provides the first systematic study of the Chinese DaaS ecosystem. Analyzing 25,972 messages and 13.22 million subscriber links from 100 major channels on Telegram, we demystify its organization, operations, and user engagement. We find that the DaaS ecosystem operates through a three-tier supply chain linking data providers, service operators, and end users. Operators sustain illicit businesses through six major service categories, persistent advertising, and crypto-based payments, while users interact via specialized group spaces that enable real-time matching, large-scale identity exposure, and community-driven fraud mitigation. Our findings reveal a mature, resilient underground data market operating within mainstream messaging platforms, highlighting new challenges for online privacy and exposing critical vulnerabilities in platform governance and content moderation. This study provides empirical evidence for developing effective regulatory frameworks and accountability mechanisms to mitigate commodified online harms on encrypted messaging services.
Large-scale diffusion models have demonstrated remarkable success across a variety of domains. These models not only exhibit exceptional performance in their primary tasks but also adapt well to downstream applications through the 'pre-train & fine-tune paradigm'. However, the potential misuse of diffusion models for generating unsafe content has raised significant concerns regarding their governance and regulation, necessitating robust unsafe output prevention strategies. Despite the urgent demand for mitigation techniques, a significant challenge persists: once a model is distributed for local deployment or fine-tuning, the model provider and third-party regulators relinquish control over the model's behavior. To address this challenge, we propose a tripartite interactive regulatory scheme (Trireme) to enforce unsafe output prevention for diffusion models. Trireme involves a third-party regulator to monitor the results of the intermediate steps, throughout both the fine-tuning and the sampling phase of the diffusion model. To regulate model behavior, Trireme introduces regulatory signatures and adds additional regulatory layers into vanilla diffusion models. These two components are trained to be tightly coupled with each other, allowing the regulator to control the sampling process of diffusion models with different types of signatures. Through extensive experiments, we demonstrate that the proposed regulatory scheme surpasses existing approaches in terms of preventing the generation of inappropriate content. Moreover, it proves effective even when confronted with multiple adversarial attacks.
Reliable behavior control is central to deploying Large Language Models (LLMs) on the web. Activation steering offers a tuning-free route to align attributes (e.g., truthfulness) that ensure trustworthy generation. Prevailing approaches rely on coarse heuristics and lack a principled account of where to steer and how strongly to intervene. To this end, we propose Position-wise Injection with eXact Estimated Levels (PIXEL), a position-wise activation steering framework that, in contrast to prior work, learns a property-aligned subspace from dual views (tail-averaged and end-token) and selects intervention strength via a constrained geometric objective with a closed-form solution, thereby adapting to token-level sensitivity without global hyperparameter tuning. PIXEL further performs sample-level orthogonal residual calibration to refine the global attribute direction and employs a lightweight position-scanning routine to identify receptive injection sites. We additionally provide representation-level guarantees for the minimal-intervention rule, supporting reliable alignment. Across diverse models and evaluation paradigms, PIXEL consistently improves attribute alignment while preserving model general capabilities, offering a practical and principled method for LLMs' controllable generation. Our code is available at https://anonymous.4open.science/r/PIXEL-Adaptive-Steering-95DC
This paper presents a large-scale experimental audit of YouTube's user control mechanisms for managing unwanted video recommendations. Drawing on crowdsourced data from 22,722 participants using a custom-built browser extension, we analyzed over 567 million recommendations over six months. The extension introduced a ''Stop Recommending'' button overlaying recommended videos, which—depending on randomized assignment—triggered one of four native feedback signals to YouTube (e.g., ''Dislike,'' ''Not Interested,'' ''Don't Recommend Channel,'' ''Remove from History'') or no signal at all in the control group. This design allowed us to assess the effectiveness of different user controls through actual user behavior and downstream changes in recommendations. Using a machine learning model trained to estimate video similarity, we quantified how often unwanted content reappeared after user feedback. We find that YouTube's feedback mechanisms are largely ineffective: even the most ''definitive'' controls prevented fewer than half of similar recommendations. Since unwanted recommendations are relatively rare, our large-scale approach was essential to detect these effects. These findings reveal a substantial gap between user expectations and platform behavior. We conclude with design and policy recommendations to enhance user agency, transparency, and researcher access for platform accountability.
Implicit hate speech often hides harmful intent behind innocuous wording, metaphors, or hostile tone, making it difficult for detectors that rely on surface cues. We observe that large language models (LLMs) frequently exhibit pseudo-reasoning that shows over-sensitivity to spurious cues while missing the latent semantic units that actually realize hateful intent. We call these units Latent Hate Components (LHCs) and argue they should be the anchors of inference. We propose DuPL, a Dual-Process argumentation framework that centers detection on LHCs. DuPL separates (i) Mining of LHCs via a high-recall Critical Miner and a Confusion Judge that early-exits on clear cases, from (ii) Deliberation of LHCs via component-wise argumentation and a final Integrative Decision. Across IHC, SBIC, and ToxiGen, DuPL consistently outperforms prompt-learning baselines, improving accuracy by +8.36, +4.93, and +3.78 percentage points and macro-F1 by +7.18, +5.00, and +3.99 percentage points, respectively. DuPL also lowers both false positive rate and false negative rate in most settings, indicating balanced mitigation of the two failure modes common in LLM detectors. By explicitly mining and deliberating over LHCs, DuPL turns opaque, uncontrolled reasoning into structured argumentation, yielding more accurate and interpretable decisions for web moderation.
Forecasting state evolution of network systems, such as the spread of information on social networks, is significant for effective policy interventions and resource management. However, the underlying propagation dynamics constantly shift with new topics or events, which are modeled as changing coefficients of the underlying dynamics. Deep learning models struggle to adapt to these out-of-distribution shifts without extensive new data and retraining. To address this, we present Zero-Shot Forecasting of Network Dynamics through Weight Flow Matching (FNFM), a generative, coefficient-conditioned framework that generates dynamic model weights for an unseen target coefficient, enabling zero-shot forecasting. Our framework utilizes a Variational Encoder to summarize the forecaster weights trained in observed environments into compact latent tokens. A Conditional Flow Matching (CFM) module then learns a continuous transport from a simple Gaussian distribution to the empirical distribution of these weights, conditioned on the dynamical coefficients. This process is instantaneous at test time and requires no gradient-based optimization. Across varied dynamical coefficients, empirical results indicate that FNFM yields more reliable zero-shot accuracy than baseline methods, particularly under pronounced coefficient shift.
Space-based Parameter Evolving with Lightweight Optimization for Graph Adaptation to Evolving Shifts
Adapting graph neural networks to evolving domain shifts presents a fundamental challenge: how to acquire new knowledge while preventing catastrophic forgetting. Existing continual learning methods often rely on memory replay or complex regularization schemes, incurring significant computational overhead. We propose STEM (State-based Parameter Evolving with Lightweight Optimization), a replay-free framework that transforms continual adaptation into controlled parameter space evolution via a controller-worker architecture. At its core is a Test-time Evolving State Space (TESS) controller with a selective gating mechanism that recursively updates its hidden state by integrating compact summaries of the current graph domain. Unlike traditional linear state space models, TESS enables nonlinear, input-dependent state transitions that capture temporal dynamics of domain evolution. A lightweight parameter generator decodes this evolving state into domain-specific adapter parameters injected into a frozen base GNN. We employ unsupervised Information Maximization and parameter space stability regularization that penalizes adapter changes across time steps, with theoretical guarantees of forgetting mitigation and stable convergent adaptation. Extensive experiments validate that our method achieves state-of-the-art performance while maintaining minimal computational overhead. The code is available at https://github.com/miaomiao1220/stem
Graph generation underlies many critical applications, from social network modeling to knowledge graph reasoning. Across these diverse domains, many graphs are rationale–dominated : a small, semantically meaningful subgraph determines the property of interest, while the remaining edges contribute largely noisy variation. Despite the significance of this inherent structure, existing generative methods often fail to preserve these task–critical substructures. We introduce RARD (Rationale-first blockwise AutoRegressive Diffusion), a topology-guided framework that learns to separate and prioritize the rationale. RARD employs a persistent-homology-based learning objective to discover an optimal graph filtration, an edge ordering that explicitly separates rationale from noise. Building upon this learned filtration, RARD generates graphs blockwise: it adds filtration-aligned blocks autoregressively and refines each new block with a shared local discrete diffusion module, ensuring the rationale appears early while peripheral structure is added later. We provide theoretical analysis showing that maximizing the topological gap yields rationale-first ordering and collapses to a two-level filtration. Comprehensive experiments across seven benchmarks demonstrate that RARD achieves state-of-the-art performance on widely used metrics.
Temporal Betweenness Centrality (TBC) measures how often a node appears on optimal temporal paths, reflecting its importance in temporal networks. However, exact computation is highly expensive, and real-world TBC distributions are extremely imbalanced, causing learning-based models to overfit to zero-centrality nodes and fail to identify truly central nodes. Existing graph neural networks (GNNs) either ignore temporal dependencies or cannot handle such extreme imbalance. To address these issues, we propose CLGNN, a scalable and inductive contrastive learning-based GNN for accurate TBC prediction. CLGNN preserves temporal path validity through an instance graph and encodes structural, path-time aware dependencies via dual aggregation. To mitigate imbalance, a stability-based clustering-guided contrastive module separates nodes of different centrality levels in representation space, while a regression head estimates TBC values. Extensive experiments on diverse benchmarks demonstrate that CLGNN is scalable, generalizable, and effective.
Graph Neural Networks (GNNs) have achieved remarkable success across various graph-based tasks but remain highly sensitive to distribution shifts. In this work, we focus on a prevalent yet underexplored phenomenon in graph generalization, Minimal Shift Flip (MSF)—where test samples that slightly deviate from the training distribution are abruptly misclassified. To interpret this phenomenon, we revisit MSF through the lens of Sharpness-Aware Minimization (SAM), which characterizes the local stability and sharpness of the loss landscape while providing a theoretical foundation for modeling generalization error. To quantify loss sharpness, we introduce the concept of Local Robust Radius, measuring the smallest perturbation required to flip a prediction and establishing a theoretical link between local stability and generalization. Building on this perspective, we further observe a continual decrease in the robust radius during training, indicating weakened local stability and an increasingly sharp loss landscape that gives rise to MSF. To jointly solve the MSF phenomenon and the intractability of radius, we develop an energy-based formulation that is theoretically proven to be monotonically correlated with the robust radius, offering a tractable and principled objective for modeling flatness and stability. Building on these insights, we propose an energy-driven generative augmentation framework (E2A) that leverages energy-guided latent perturbations to generate pseudo-OOD samples and enhance model generalization. Extensive experiments across multiple benchmarks demonstrate that E2A consistently improves graph OOD generalization, outperforming state-of-the-art baselines. Code is available at https://github.com/anders1123/E2A