Overview

Content moderation aims to identify material such as hate speech, harassment, self-harm, and unsafe images or videos, then guide decisions about removal, review, or rewriting. Approaches range from text classifiers and multimodal vision-language models to policy-guided language models using retrieval, examples, or step-by-step reasoning; some systems also screen content as it is generated or streamed. Researchers evaluate these methods on labeled datasets and policy-specific benchmarks, including tests of multilingual content, memes, evolving rules, and adversarial changes to text or images. A central tension is that models must apply context-dependent policies consistently while avoiding unfair errors, resisting evasion, and meeting the cost and latency limits of real moderation workflows.

Latest papers

91 papers9 papers in the last four weeks, up 50% on the four weeks before. 0.1% of all new papers.

Aug 22, 2026

No One Model Catches Every Harm: Benchmarking Content Moderation Across Safety Scenarios

Large Language Models (LLMs) are increasingly deployed in real-world applications, yet they remain vulnerable to generating harmful content. From adversarial jailbreaks that bypass safety filters to implicit hate that evades detection, the range of risks these models pose continues to grow. While both specialized content moderators and general-purpose LLMs are being used as safety layers, the question of which model is best suited for which type of harmful content remains unanswered. We present the most comprehensive evaluation of LLM safety capabilities to date, systematically testing 53 models across 11 datasets that we organize into four distinct categories. Our evaluation under both prompt-only and prompt-response settings uncovers critical blind spots: large frontier models that lead on one category fall significantly behind smaller, specialized alternatives on others, and real-world conversational safety remains largely unsolved across all model families. These findings challenge the assumption that scale alone ensures safety, and provide the community with a structured framework for informed model selection.
Aug 18, 2026

Auditing Exposure to Harmful Content on TikTok using Multimodal Language Models: A Cross-National, Age-Stratified Study

Online video platforms can expose young users to harmful content, but independent audits remain difficult because video annotation is costly and moderation judgments vary across languages. We audit TikTok in France, Italy, and Sweden with sockpuppet accounts representing four age personas (13, 16, 19, 40), collecting 36,971 videos from passive For-You-page scrolling and active sessions that scroll, search for harm keywords, and scroll again. To scale annotation, we validate four multimodal LLMs against native-speaker labels on a 300-video reference set. Gemini 2.5 Flash with eight sampled frames plus text performs best (aggregate kappa = 0.42), at half the per-call cost of native-video upload, and we apply it to a 10% sample for approximately $50 in total API spend across both modalities. Keyword search returns 35-56% harmful content, a 1.5-7.5x increase over the scrolling baseline in ten of twelve country-age combinations; the spike is temporary and flattens the age differences observed in France and Sweden. Under passive scrolling, Italy has the highest harm rate at every age, with Italian age-19 reaching 48.6%. Overall, MLLM-based auditing offers a scalable approach for cross-national youth-safety audits, while provider safety filters (1.1% refusal rate) under-count the most explicit harms.
Aug 17, 2026

"If It Looks Like a User": Measuring Real-Time Moderation Effects via Social Media Simulation

Agent-based social media simulators offer a controlled environment to study content moderation, yet their value hinges on how faithfully they reproduce real platform dynamics. We develop a calibrated extension of SimSoM, an agent-based model of information diffusion on social networks, grounded in a real-world dataset of online vaccine discourse during the COVID-19 pandemic. Our approach replaces ad-hoc parametrisations with empirically fitted distributions, optimised via CMA-ES (Covariance Matrix Adaptation Evolution Strategy) and validated against real data across temporal, distributional, and structural dimensions. Using this validated simulator, we provide three key contributions. First, we show that the calibrated model reproduces key statistical signatures of the empirical data, including activity distributions, post/reshare ratios, and temporal patterns. Second, we apply established misinformation-spreader detection and prevention methods to both empirical and simulated data, progressively removing top-ranked users and showing that the resulting decline in low-quality content is consistent across the two. Third, comparing static (retroactive) and dynamic (in-simulation) moderation across 30 network realisations, we show that static evaluation significantly overestimates the effectiveness of user bans for the most effective detectors: when moderation is applied in real time, compensatory resharing by the remaining users dampens the expected reduction in low-quality content, so static estimates should be read as an upper bound. These findings highlight the necessity of simulation-based evaluation for content moderation policies and contribute a reusable, empirically grounded simulation framework.
Aug 16, 2026

VARM-Bench: Benchmarking Verifiable Structured Reasoning in Chinese Abusive Speech Moderation

The widespread circulation of abusive online content has increased the need for reliable moderation of Chinese social-media text. Existing Chinese benchmarks support label classification, fine-grained toxicity categorization, and target-aware extraction, but do not provide a unified representation for deterministically verifying the stated basis of a moderation decision. We introduce VARM-Bench, a benchmark for field-anchored chain-of-thought rationales in Chinese abusive-speech moderation. Each instance contains a concise natural-language rationale with explicit anchors for six decisions: target, target type, target explicitness, author stance, harmfulness label, and fine-grained category. Our deterministic protocol evaluates field correctness, target alignment, output validity, complete-record agreement, and hidden record errors conditioned on correct final decisions, without relying on an LLM judge. Under a common structured-output protocol, we evaluate language models across multiple model families using zero-shot prompting, taxonomy guidance, and structured CoT supervision, and analyze lexical-cue sensitivity and field-level errors. Results show that strong label-level performance can conceal substantial errors in complete moderation records. VARM-Bench provides an auditable and reproducible benchmark for evaluating verifiable moderation rationales in Chinese abusive-speech moderation.
Aug 10, 2026

Withholding the Completing Chunk: Exact Release-Boundary Equivalence for Production Streaming Guardrails

Streaming language-model output creates an enforcement boundary: a control that detects a prohibited pattern after releasing its completing chunk cannot recall it. We study a production policy in which each ordered family is the conjunction of two regular-language predicates. Incremental matching is classical. The problem is exact composition at release time across arbitrary chunk partitions, including end-of-prefix word boundaries that can change on extension. We define an ASCII-explicit policy grammar, compile each predicate to a persistent nondeterministic finite automaton (NFA), distinguish stable from provisional assertion state, apply document-order family priority, and check the decision before releasing each chunk. We show that the resulting monitor is release-boundary equivalent to an absorbing cumulative oracle for every policy in the declared grammar. Production Python and TypeScript implementations were evaluated on 101,653 partitioned cases; a public surrogate added 100,345 cases. Both campaigns produced zero oracle, cross-runtime, or intended-family mismatches. In a frozen neutral-output profile, the memoized incremental and native-regex cumulative slopes at 64-character chunks were 0.973 and 1.976. At 16,384 characters the incremental median was 30.2 ms versus 96.6 ms for native cumulative scanning at that chunk size. Native regex remained faster at 512-character chunks (12.4 versus 29.4 ms), exposing the constant-factor crossover rather than hiding it. A shared per-stream cache cap and 129-symbol alphabet bound optimization state; the campaign peaked at 364 of 4,096 without bypass. The result is policy conformance for a deterministic backstop, not evidence of semantic safety or policy completeness.
Aug 2, 2026

ArabicDialectSafety: A Dialect-Aware Benchmark for Arabic Content Safety Classification

We present ArabicDialectSafety, a human-curated Arabic safety dataset of 25,071 prompts covering six Arabic varieties: Modern Standard Arabic, Syrian, Egyptian, Algerian, Palestinian, and Moroccan. The dataset is annotated with dialect labels and seven fine-grained harm categories. We introduce a dual-task evaluation framework for binary safe/unsafe detection and granular harm classification across dialects. Benchmarking seven supervised and generative models, we find that fine-tuned MARBERTv2 achieves the strongest performance, with Macro-F1 scores of 0.95 for binary classification and 0.90 for granular classification, substantially outperforming prompted frontier LLMs, including Arabic-specialized models. Our analyses show that dialect conditioning is most effective when integrated at the representation level, while significant performance gaps remain for low-resource Maghrebi dialects. We further evaluate seven frontier LLMs as response generators on harmful dialectal Arabic prompts and observe unsafe generation rates below 5 percent across models. We release the dataset and code upon acceptance to support future research on dialect-aware Arabic safety evaluation. Warning: This paper contains examples of harmful and potentially offensive content included solely for research purposes.
Jul 30, 2026

Old Tricks, New Models: How Simple Image Transformations Break Modern AI-based Content Moderation

While automated content-moderation systems have become essential for screening harmful content at scale, conventional task-specific classifiers often provide limited policy cov- erage and contextual understanding. Recently, commercial multimodal moderation APIs built on large foundation models have been introduced with the promise of providing broader and more capable safety filters. In this work, we analyze whether this shift also yields more robust image moderation. We conduct a large-scale black-box evaluation on three established commercial image-moderation services and compare their robustness. By evaluating seven simple, model-agnostic image transformations across multiple providers, datasets, harm categories, perceptual-similarity constraints, and transformation intensities, we find that: (1) all three commercial services can be bypassed using inexpensive image transformations that require no gradients, surrogate models, or knowledge of the target system; (2) even fixed transformations such as color inversion and grayscale conversion induce unsafe-to-safe decision changes while preserving content that remains recognizable to humans; (3) their robustness varies substantially across datasets and harm categories, with multimodal content and self-harm exhibiting pronounced vulnerabilities. This yields the conclusion that replacing conventional moderation classifiers with foundation-model-based APIs does not, by itself, provide a reliable security boundary. Such systems must be evaluated under realistic transformations and deployed as one component of a layered moderation pipeline rather than as standalone safety filters.
Jul 28, 2026

Choosing Where and How to Moderate: End-to-End Trade-offs in Filter Placement and Response Rewriting

Content-moderation classifiers are usually evaluated in isolation, but deployment requires choosing where to intervene and what follows a flag. We evaluate these choices using two end-to-end customer-outcome metrics rather than component accuracy: Usefulness, the fraction of turns with a shown, non-harmful, relevant response, and Harmful Exposure, the fraction with a shown harmful response. Latency and error rates are diagnostics. We compare Input only, Response only, and Input + response hard blocking on a human-labelled product benchmark and public ToxicChat evaluation. At the evaluated operating points, Response only achieves the highest filter-only Usefulness in both settings, while Input + response achieves lower Harmful Exposure. Replacing Response only blocking with Response + rewrite recovers most blocked traffic and yields the same observed Harmful Exposure count as Response only blocking for the selected configuration; this equality is not an equivalence result. Probe routing substantially reduces conditional route-and-generation time relative to LLM routing at comparable measured outcomes. A focused output review shows how rewrites balance filter passage with usefulness by generalizing triggering language while retaining benign intent and safe redirection; some sensitive-domain outputs nevertheless omit potentially safety-relevant support information. These results support comparing moderation configurations under deployment-specific safety and latency constraints rather than applying a universal placement rule. Code and public artifacts are available at https://github.com/microsoft/mod-frontier
Jul 28, 2026

Shieldstral

We introduce Shieldstral, a 3B-parameter policy-adaptive multimodal safety classifier that matches or outperforms models nearly 7×\times its size on text safety benchmarks and sets a new state of the art on multimodal safety classification. Shieldstral formulates content moderation as a binary question-answering task. This simple formulation unifies diverse moderation tasks into a single yes/no problem, enabling heterogeneous safety datasets with divergent taxonomies to be consolidated under one training framework. We present the data construction recipe, covering curation and generation of approximately 54.1M samples and a fine-grained evaluation set to evaluate policy adaptability. Together, these enable a small adaptive model to match or outperform much larger models.
Jul 24, 2026

EVL-MCoT: Enhanced Vision-Language Multi-CoT for Harmful Meme Detection

MEMEs are widely used on the internet and often carry strong elements of sarcasm or irony. Understanding their hidden meanings typically requires a joint interpretation of text and vision. Existing methods focus on the dual-stream vision-language model to extract the visual and text simultaneously, which lacks background information and prior knowledge about the comprehensive explanation of MEME. One feasible option is to adopt chain-of-thought (CoT). However, the simple CoT approach lacks multi-perspective thinking, which may compromise the reliability of the resulting answers. Moreover, it often relies on shallow feature fusion, lacking the fusion of local details and fine-grained visual-prompt text alignment. This limitation prevents a deeper understanding of the intricate connections between the visual and the text. Herein, an enhanced vision-language multi-CoT (EVL-MCoT) approach is proposed to address these limitations. By promoting multi-CoT, EVL-MCoT enhances consistency and reduces bias in the decision-making process. Additionally, we design a prototype-guided and context-guided decoding framework, which incorporates visual prototypes to guide the fusion process and enables the model to align textual and visual information more precisely. We achieve promising results on the HatefulMemes and MultiOff datasets. The source code has been publicly released and is available at https://github.com/BGWH123/EVL-MCoT.
Jul 20, 2026

A Dual-Hypothesis Reasoning Framework for LLM Guardrails

We propose ARBITER, a novel LLM guardrail framework that introduces two key ideas: (i) dual-hypothesis reasoning, a reasoning method for LLM guardrails that explicitly considers both safe and unsafe interpretations of a prompt before making a safety decision, and (ii) multi-component supervised fine-tuning (MC-SFT), a structured training loss for reasoning-based guardrails that decomposes LLM outputs into logical components and weights them according to their importance. Existing reasoning-based guardrails often rely on expensive procedures, such as generating reasoning traces using larger or closed-source teacher models and applying full-parameter fine-tuning. In contrast, ARBITER uses a cost-effective self-generation strategy for reasoning traces and LoRA-based parameter-efficient fine-tuning while still achieving better performance than these expensive approaches. Additionally, ARBITER provides faithful evidence-phrase explanations for unsafe decisions, enabling a more transparent and interpretable guardrail method. Experiments on three safety moderation benchmarks show that ARBITER outperforms existing reasoning-based and non-reasoning guardrail baselines, with clear gains in out-of-domain evaluations.
Jul 12, 2026

Operational Evidence Gaps for LLMs in Fraud Detection and Trust-and-Safety Workflows

LLMs are now proposed for fraud detection, scam investigation, content moderation, and other trust-and-safety workflows. Much of the public literature still evaluates them as models, with less attention to their behavior as components in operational pipelines. This creates a practical evidence question: what would justify placing an LLM inside a live workflow with latency, cost, escalation, human-review, and adversarial-risk constraints? We address this question through a fraud-first survey of deployment evidence. We code 49 operationally relevant sources on LLM use in fraud detection, investigation support, content moderation, and cross-cutting robustness (18 fraud, 14 moderation, 17 cross-cutting), supplemented by 15 contextual references that establish the survey boundaries. These sources include systems, benchmarks, frameworks, and deployment-relevant surveys, not 49 production deployments. The main finding is an evidence imbalance. Fraud supplies the largest task-specific portion of the coded corpus. The moderation papers, however, include more explicit public evidence on latency, cost, governance, and fairness. Among the 18 fraud and investigation sources, none report clean per-decision latency, per-decision dollar cost, or calibration evidence; most report offline task performance, retrieval gains, or case-study accuracy instead. The survey contributes a role-and-evidence organizing frame, FORTE, for locating LLMs as classifiers, retrieval interfaces, explanation generators, reviewer assistants, agents, feature extractors, or escalation components. It also contributes a minimum deployment-evidence checklist covering latency budget, cost per decision, decision threshold, explanation integrity, and adversarial pressure. The resulting agenda identifies studies needed to support deployment claims for LLM-based fraud and trust-and-safety work.
Jul 8, 2026

R^3: Advertisement Compliance Rectification via Group-Relative Experience Extractor and Curriculum Reinforcement

Rigorous content moderation is crucial for online advertising but leads to millions of daily rejections. This scale renders manual rectification infeasible, particularly for video advertisements. However, existing safety-driven methods often suffer from aggressive over-editing, which compromises the advertiser's original semantic intent merely to satisfy compliance. In this work, we target the rectification of textual violations in video ads, covering both speech transcripts and on-screen text. We propose R^3, a novel framework designed to harmonize compliance with original semantic intent preservation. Our approach integrates three key innovations: (1) an experience-driven data synthesis framework that bootstraps high-quality supervision via a group-Relative compliance experience extractor; (2) a curriculum Reinforcement learning strategy with hierarchical rewards designed to enforce compliance while maximizing semantic consistency; and (3) a comprehensive video Rectification framework seamlessly integrating text recognition, rewriting, and re-rendering for industrial deployment. Extensive experiments on industrial datasets and online A/B testing demonstrate that R^3 significantly outperforms state-of-the-art baselines, achieving an optimal trade-off between violation rectification and intent preservation.
Jul 7, 2026

PolicyShiftGuard: Benchmarking and Improving Policy-Adaptive Image Guardrails

Image guardrails are typically trained and evaluated under a fixed safety policy, implicitly treating safety as an intrinsic property of an image. Real deployments are different: the same image may be allowed in one product, restricted in another, and newly disallowed when a policy boundary changes. We study policy-adaptive image guardrailing, where a model must decide whether an image violates the currently supplied policy and generalize to held-out policy definitions. We introduce PolicyShiftBench, a comprehensive benchmark with 2,000 policy-discriminative instances over 265 images, where each image is paired with 7.55 policy-conditioned prompts on average to test whether models adapt to the active policy rather than relying on image-level safety priors. We then propose PolicyShiftGuard, a compact policy-conditioned guardrail trained with a two-stage training recipe that combines Randomized Policy SFT (RP-SFT) with Boundary-Pair Policy Adaptation (BP-Adapt). BP-Adapt trains matched prompts for the same image and risk category using standard label supervision and a pairwise comparison loss that separates blocking policies from passing policies. Experiments show that existing VLMs and specialized guardrails remain brittle under policy shifts, while PolicyShiftGuard substantially improves policy-sensitive performance. The 7B model achieves SOTA performance of 76.9 Avg. F1 and 72.1 Avg. PSS on PolicyShiftBench, transfers well to UnSafeBench and SafeEditBench, and improves the latency-performance trade-off with a concise output format. Ablations confirm that matched pass/block boundary pairs are essential for stable policy adaptation.
Jun 25, 2026

Beyond Surface Forms: A Comprehensive, Mechanism-Oriented Taxonomy of Indirect Linguistic Encoding for LLM-Based Coded Language Detection

To avoid moderation and surveillance on social media, some users routinely invent indirect linguistic expressions (ILE) that camouflage sensitive meanings. Such disguised expressions surface as algospeak, euphemisms, and adversarial obfuscation, depending on intent and context, and often involve recurring encoding mechanisms. We propose a comprehensive, mechanism-oriented taxonomy of ILE that abstracts away from communicative goals and instead categorizes the underlying operations through which meaning is encoded and recovered. We evaluate the taxonomy by incorporating it into large language model (LLM) prompts and comparing it with four existing taxonomies and a no-taxonomy baseline, using 2,000 manually annotated TikTok and Bluesky posts. The proposed taxonomy attains the strongest document- and span-level performance across three LLMs. Compared with the best-performing benchmark, its document-level gains are +3.8 accuracy/+4.3 macro-F1 percentage points for GPT-5.4, +0.9/+1.3 for Sonnet 4.6, and +2.1/+2.2 for DeepSeek V4. While the direction of improvement is consistent, the magnitude is model-dependent. The span-level gains are mainly driven by higher recall, sometimes at the cost of precision. In an additional cross-lingual evaluation of 800 Chinese codedlanguage reviews, the unadapted taxonomy also achieves the highest document- and span-level F1 across the three LLMs. The empirical results reveal the importance of a comprehensive, mechanism-oriented taxonomy as a stable scaffold for detecting emerging coded language and a useful input to content moderation. Disclaimer: This paper contains content that may be profane, vulgar, or offensive.
Jun 18, 2026

CATCH-ME if you RAG: a dataset of Contextually Annotated multi-Turn Counterspeech against Hate and Misinformation Exchanges

Online hate speech and misinformation frequently overlap, yet NLP research has mainly treated them in isolation. While LLMs represent a scalable solution for assisting humans in the generation of counterspeech for both threats, zero-shot models frequently generate repetitive and vague responses, underscoring the need for high-quality examples to steer model generation. However, existing counterspeech datasets against the overlap of hate and misinformation are scarce and limited to single-turn English dialogues, while real-life interactions span across multiple turns and languages. To bridge this gap, we introduce the first large-scale, expert-curated, multilingual dataset of dialogues tackling the intersection of hate and misinformation. To ensure factual grounding, the dialogues are also anchored in verified external knowledge (i.e., fact-checking articles and NGO reports) and include document- and chunk-level span annotations, making it directly applicable for RAG systems. Covering five languages and targeting hate directed at seven marginalized groups, this novel resource enables the training and evaluation of more persuasive, factually grounded counterspeech models.
Jun 12, 2026

Selective Control under Noisy Perception: Governance Failures Hidden by Aggregate Metrics in Modular Networks

A content-moderation system can score well on every standard accuracy metric and still cause real harm, if its mistakes fall on the few users who connect otherwise separate communities. We show this in an agent-based model where N=240 learning agents on a community-structured network each post harmless, productive, or dangerous content, and a regulator removes or penalizes whatever a noisy classifier flags. Overall usefulness barely moves as the noise changes (one-way ANOVA, p=0.96): by aggregate measures, nothing looks wrong. The damage instead concentrates on these bridge users, whose useful posts are wrongly suppressed and whose dangerous posts are wrongly spared. A governance loss (L_gov) that prices these two mistakes separately from the cost of enforcement more than doubles under false-positive-heavy noise. Aggregate accuracy hides who is harmed, and the cheap quantity to audit is how many connections a user has (degree), a near-perfect proxy for the betweenness that defines a bridge (r=0.96).
Jun 12, 2026

BELLS-O: Evaluating the Operational Trade-offs of LLM Supervision Systems

LLM supervision systems, namely input/output moderation filters and jailbreak detectors, are the primary safeguard against misuse in deployed AI applications, yet existing benchmarks are often vendor-biased, omit cost and latency, and rarely compare specialized guardrails against repurposed generalist LLMs. We present BELLS-O (Benchmark for the Evaluation of LLM Supervision Systems, Operational), an independent operational benchmark that evaluates 33 systems from 17 developers: specialized guardrails (e.g., Llama Guard 4, ShieldGemma-2B, Lakera Guard) and frontier generalists repurposed as supervisors (e.g., GPT-5.4, Claude Sonnet 4.6, Grok 4.1 Fast), jointly on detection rate, false-positive rate, latency, and cost. We cover input/output moderation across 11 harm categories, jailbreak detection across 13 attack techniques, and prompt-injection detection, using in-house datasets built from handcrafted prompts, curated external samples, and quality-controlled synthetic generation, with synthetic moderation samples paraphrased to suppress generator fingerprints. Content-moderation runs are repeated five times and all rates reported with 95% intervals. The Pareto frontier reveals use-case-dependent tradeoffs. On content moderation, specialized supervisors are operationally ahead: top systems match the best generalist LLMs on detection (93-96%, no significant difference) at low false-positive rates while running ~5x faster; cost depends more on serving setup than on class. On jailbreak detection, generalist LLMs achieve higher median detection on three of four datasets and lower false-positive rates, but at higher latency and, for closed frontier models, 10-100x higher cost. On prompt injection, the two best dedicated detectors lead. We release the framework, an interactive leaderboard, and four datasets as a vendor-neutral basis for selecting safeguards under deployment constraints.
Jun 11, 2026

Mod-Guide: An LLM-based Content Moderation Feedback System to Address Insensitive Speech toward Indigenous Ethnic and Religious Minority Communities

Language operates as a mechanism of both marginalization and resistance, especially for minority communities navigating insensitive and harmful speech online. As content moderation increasingly depends on large language models (LLMs), concerns arise about whether these systems can recognize culturally insensitive speech-language that disregards or marginalizes the cultural and religious perspectives of historically underrepresented communities, often through implicit erasure, misrepresentation, or normative framing, rather than overt hostility. Focusing on Bangladesh's Hindu and Chakma communities -- the country's largest religious and Indigenous ethnic minorities, respectively -- this paper investigates the epistemic limits of LLM-based moderation systems and explores methods for incorporating minority perspectives. We co-created a culturally grounded corpus of insensitive speech with community members and integrated their narratives into moderation pipelines using retrieval augmented generation (RAG). Our tool, Mod-Guide, improves LLM sensitivity to minority viewpoints by leveraging contextual cues derived from lived experience. Through mixed-method evaluations involving both minority and majority participants, we demonstrate that RAG-enhanced moderation responses are more contextually accurate and perceived differently across ethnic lines. This work advances research in human-computer interaction, AI ethics, and social computing by foregrounding restorative justice and hermeneutical inclusion in the design of content moderation systems.
Jun 4, 2026

Forgive or forget: Understanding the context of hate in audio retrieval systems

Handling toxic retrieval in text-to-audio systems is challenging due to contextual dependencies. Existing strategies (e.g., rephrasing, summarization) risk altering intent or omitting details. We propose a post hoc causal debiasing framework with a sentiment-controlled mediator to preserve semantic relevance while suppressing harmful speech. Our approach is model-agnostic and integrates seamlessly with existing retrieval pipelines. We introduce two variants: Forgive, which re-ranks and filters toxic audio via logit adjustment, and Forget, which generates counterfactual toxic prompts to mitigate harmful retrievals. Experiments show consistent toxicity reduction with minimal loss in retrieval accuracy, improving both safety and reliability.
Jun 4, 2026

UNIVID: Unified Vision-Language Model for Video Moderation

Global-scale video moderation faces a dual challenge: the need for fine-grained multi-modal reasoning and the demand for interpretable outputs to support downstream enforcement. Traditional moderation systems often rely on fragmented black-box classifiers that are difficult to maintain and lack transparency. In this paper, we present UNIVID, a UNIfied VIsion-language model for video moDeration. Unlike standard classification models, UNIVID generates policy-aware captions that serve as an interpretable intermediate representation, enabling human-verifiable decisions and multi-task reusability. While existing open-source and commercial VLMs often suffer from safety-guardrail refusals and lack fine-grained policy alignment, we develop a specialized training data recipe that combines expert human-refined labels with synthetic data to align the model with our safety guidelines. By integrating UNIVID as the core captioner, we design a novel end-to-end video moderation system that reduces violation leakage by 42.7% and overkill rate by 37.0% relatively. Meanwhile, by replacing over 1,000 policy-specific models with a single UNIVID backbone, we recycled extensive computation resources while reducing engineering maintenance overhead. To our knowledge, this is one of the first reports of a high-efficiency captioning VLM successfully supporting industrial-scale moderation and cross-functional business.
Jun 4, 2026

When Surface Form Changes Moderation Decisions: A Paired Study of Code-Mixed Workflow Instability

Hate moderation is often evaluated as classification on clean English inputs, but deployed systems must route content to actions such as ALLOW, FLAG, or REVIEW. We study how this workflow changes under code-mixed inputs using a paired evaluation setting where the same underlying content is expressed as clean English and Tamil-English code-mix. Under thresholds tuned on clean English development data, code-mixed inputs produce substantial action instability, with a paired clean- to-code-mix decision flip rate of 0.265. The main workflow effects are increased review burden and increased false-flagging of non-hateful content: review rate rises from 0.138 to 0.297 and non-hate false-flag rate rises from 0.069 to 0.104. Tamil-only inputs show stronger degradation overall, suggesting a broader language-coverage limitation rather than the same code-mixed instability pattern. A simple disagreement-based deferral rule reduces automatic errors on stressed inputs, but only by increasing review load. These results show that workflow-level evaluation reveals moderation failures that standard classification summaries can miss.
Jun 1, 2026

The Ghost Annotator: a Framework to Explore Human Label Variation in Content Moderation through Conformal Prediction

Current research primarily focuses on model performance, while comparatively less attention has been devoted to uncertainty estimation, particularly in settings where LLMs are increasingly used to generate annotated data. We introduce a framework combining conformal prediction with Collaborative Filtering-style annotators' representation to model LLM behavior in relation to human annotators and to analyze patterns of agreement and disagreement. Using Non-Conformity Scores, we introduce the Ghost Prediction metric and the Ghost Annotator representation to quantify cases in which model predictions diverge from all available human annotations. We compute cosine similarity measures to explore differences in model behavior across sociodemographic axes. We evaluated four LLMs of different size and families across four content moderation datasets. Our finding shows that while we find that all models uncertainty increases with annotator disagreement, larger models tend to be more confident in the classification of texts that are not aligned with any human annotation. Finally, the Ghost Annotator framework reveals a consistent and robust pattern of demographic misalignment, suggesting a structural bias likely rooted in pretraining corpora.
Jun 1, 2026

FigSIM: A Dataset for Fine-grained Suicide Severity and Figurative Language in Suicide Memes

Suicide memes are memes used to express suicide-related thoughts or comment on suicide-related issues. Suicide memes are increasingly common on social media, yet remain poorly understood and potentially harmful. There is an urgent need to better understand their characteristics and to develop appropriate content moderation strategies that limits users' exposure to potentially harmful content. Currently, the absence of annotated datasets of suicide memes remains a key barrier to developing and evaluating automated moderation approaches. In this paper, we introduce FigSIM, the first dataset designed for fine-grained analysis of suicide memes. The dataset consists of 1049 memes, each annotated for (1) fine-grained suicide severity levels, (2) figurative phenomena (e.g., metaphors), and (3) suicide-related content (e.g., suicide method depiction). We benchmark 16 unimodal and multimodal models across three tasks: figurative language, suicide severity, and suicide-related content detection. Overall, FigSIM demonstrates that suicide memes pose unique challenges for both modeling and content moderation. Analysis revealed biases, such as underprediction of higher suicide severity levels, especially for figurative memes. The dataset (including splits used for analyses) is publicly available. Content Warning: This paper contains suicide-related content that may be triggering.
Jun 1, 2026

SentGuard: Sentence-Level Streaming Guardrails for Large Language Models

Large language models increasingly stream long, reasoning-intensive responses in real time, making when to moderate as critical as whether to moderate. Existing guardrails fall into two unsatisfactory extremes: response-level methods delay intervention until the full output is generated, whereas token-level methods act on incomplete semantics, often producing unstable decisions and excessive guard invocations. To address this challenge, we propose SentGuard, a sentence-level streaming guardrail that operates in parallel with generation. A lightweight waiting buffer groups streamed tokens into sentence chunks and releases only verified chunks to the user, introducing a small offset that enables SentGuard to assess the current prefix while the target LLM decodes subsequent content. To support this, we construct StreamSafe, a benchmark with structured per-sentence annotations across 8 harm categories, capturing the evolution of safety risks across both reasoning and response segments. We further train SentGuard with a coarse-to-fine objective to detect unsafe intent as soon as it emerges at sentence boundaries. Experiments on 5 safety benchmarks show that SentGuard outperforms existing baselines, detecting 90.5% of unsafe cases within two sentences while maintaining a low streaming false-positive rate of 7.41%.
May 29, 2026

TSM-Bench: Detecting LLM-Generated Text in Real-World Wikipedia Editing Practices

Automatically detecting machine-generated text (MGT) is critical to maintaining the knowledge integrity of user-generated content (UGC) platforms such as Wikipedia. Existing detection benchmarks primarily focus on generic text generation tasks (e.g., ``Write an article about machine learning.''). However, editors frequently employ LLMs for specific writing tasks (e.g., summarisation). These task-specific MGT instances tend to resemble human-written text more closely due to their constrained task formulation and contextual conditioning. In this work, we show that a range of SOTA MGT detectors struggle to identify task-specific MGT reflecting real-world editing on Wikipedia. We introduce TSM-Bench, a multilingual, multi-generator, and multi-task benchmark for evaluating MGT detectors on common, real-world Wikipedia editing tasks. Our findings demonstrate that (i) average detection accuracy drops by 10--40% compared to prior benchmarks, and (ii) a generalisation asymmetry exists: fine-tuning on task-specific data enables generalisation to generic data -- even across domains -- but not vice versa. We demonstrate that models fine-tuned exclusively on generic MGT overfit to superficial artefacts of machine generation. Our results suggest that, in contrast to prior benchmarks, most detectors remain unreliable for automated detection in real-world contexts such as UGC platforms. TSM-Bench therefore provides a critical foundation for developing and evaluating future models.
May 27, 2026

Robust and Efficient Guardrails with Latent Reasoning

Maintaining the safety of large language models (LLMs) is crucial as they are increasingly deployed in real-world applications. Existing safety guardrails typically rely on single-pass classification or, more recently, distilled reasoning. Reasoning-based guardrails significantly outperform classification-only baselines, but they incur substantial query latency and token overhead that make them impractical for highthroughput deployment. To address this challenge, we propose COLAGUARD, a guardrail model that transfers multi-step safety reasoning into a continuous latent space through a stage-wise training curriculum, enabling direct hidden-state propagation at inference. Evaluated on ten prompt- and response-moderation settings spanning eight safety benchmarks, COLAGUARD improves macro-F1 by 8.24 points over Llama Guard 3 and matches our explicit reasoning baseline, GuardReasoner, in macroF1 while delivering a 12.9X speedup and 22.4X reduction in token usage. Our results suggest that latent reasoning offers a practical alternative to explicit rationale generation for deployable guardrails, jointly improving safety robustness and inference efficiency rather than treating them as competing objectives.
May 25, 2026

AI Content Moderation in Therapy Conversations

Large language models (LLMs) are increasingly being used for emotional support. They are also being developed for formal therapy purposes. However, LLMs like ChaptGPT or Llama are often developed with content moderation guardrails that prevent them from discussing sensitive subjects with users for both liability and safety purposes, and this inability to broach these subjects may affect their capacity as therapists. In this study, we perform an algorithm audit on three state-of-the-art moderation systems (OpenAI's moderation endpoint, Meta's Llama Guard, and Google's Shield Gemma) to investigate the extent to which these systems flag the content of real-life therapy sessions as undesirable. Our results raise implications for the limitations that users and organizations may encounter when designing LLMs to play the part of a therapist.
May 23, 2026

Appearance-Invariant Detection of Suggestive Motion via Laban Movement Descriptors

Content moderation in online multiplayer 3D virtual environments is increasingly automated, yet detection has focused on images, video, and audio, leaving suggestive motion a blind spot. We present a motion-only classification pipeline that detects suggestive and explicit movement from SMPL skeleton trajectories using Laban Movement Analysis (LMA) descriptors. On a dataset spanning everyday, artistic, suggestive, and explicit movement (17+ hours of video), a logistic regression trained on 61-feature LMA descriptors reaches 68% binary SFW/NSFW accuracy (70% random forest) under a leak-free evaluation protocol. At this level, our descriptor performs comparably to a learned video model trained on the same motion re-rendered as appearance-free video, a gray figure with no clothing, skin, or scene. The indirectness (tortuosity) of each joint's trajectory, measured as the ratio of the joint's path length to its net displacement, peaks at the suggestive tier, showing that the Direct-to-Indirect polarity of Laban's Space factor provides an interpretable marker of the shift from functional to suggestive motion. Ultimately, Laban-based kinematic descriptors offer a lightweight, interpretable approach to suggestive-motion detection: every decision decomposes into named, theory-grounded features. Because the classifier operates on pose trajectories alone, moderation can run directly on avatar poses in virtual environments, with no appearance data.
May 22, 2026

Improving Labeling Consistency with Detailed Constitutional Definitions and AI-Driven Evaluation

Many automated labeling pipelines classify inputs into categories defined by a written specification, content moderation being a prominent use case. Simple category definitions are not detailed enough for labelers to produce the accurate, consistent golden labels these pipelines require. One solution is to write a prescriptive definition that settles enough real boundary cases that labelers cannot disagree with the written interpretation. In practice, definitions at that level of detail exceed what a human annotator can hold in working memory, so annotators fall back on intuition and the labels drift from the written rules, regressing on accuracy and consistency. We propose and demonstrate the efficacy of an AI-driven workflow in which AI helps write a per-category constitution that defines the label in enough detail to cover edge cases, and a frontier LLM interprets it on each input to produce the golden label more consistently and accurately than humans reading the same document. We evaluate on three content moderation categories (harassment, hate speech, non-violent crime) and show that the approach reduces cross-model inconsistency by up to 57x compared to paragraph definitions, with cross-model disagreement diagnosing specification gaps and the human responsible for high-level decisions about what each category should mean rather than individual labeling calls. For the safety evaluation, we introduce a dual-axis formulation scoring intent and content independently over the full conversation, so downstream consumers can act on either axis or both.