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.
Ye Leng, Junjie Chu, Yiting Qu +3cs.CV cs.AI cs.CR
Picture books and comics have long been used to disseminate hateful narratives because they are easily understood even by children, as exemplified by the notorious Nazi propaganda picture book \emph{Der Giftpilz}. Recently, frontier text-to-image (T2I) systems such as Gemini and GPT-Image have enabled conversational generation with consistent characters and scenes across turns, making hateful visual stories, namely ordered image groups that collectively convey hateful narratives, cheap and scalable to produce. Although prior work has studied hateful content generation by T2I systems, it focuses on individual images, leaving group-level hateful meaning largely unexplored. We aim to address the gap. Concretely, we introduce \texttt{HatefulStoryPrompts}, comprising 330 multi-turn configurations from 55 hateful stories across two languages and three visual styles, and evaluate five frontier models over 4,950 attempts. Every model completes over 80\% of the stories, with the strongest reaching 99.0\%. We further evaluate existing moderation systems on \texttt{HatefulVisualStory}, a human-labeled dataset of 969 hateful image sets and 990 benign controls, and find that they frequently miss group-level hateful meaning: dedicated safety models achieve at most 34.9\% recall, while a strong vision-language model reaches 67.5\%. Finally, we propose complementary proactive and post-generation defenses. An interaction-aware monitor achieves 97.3\% recall for prompt-only sessions and 92.6\% when the user supplies the first image, while post-generation methods jointly analyzing completed image groups reach 80.2\%. Our work shows that, as image generation evolves from isolated outputs to coherent visual narratives, safety must evolve accordingly, from per-image moderation to stateful reasoning over interactions and image relationships.
A vision-language AI assistant returns its answer as a stream of generated tokens. Therefore, a safety guard that watches that answer has to keep up with the stream and stop a harmful reply before a user reads it. Recent vision-language guardrails instead generate a chain of thought before they issue a verdict. They believe that step-by-step reasoning yields a safer guard. This design makes the guard heavy and slow, since the model must decode many tokens for harmfulness detection. We pose the question of whether a vision-language guard really needs to reason in order to screen a response. We answer with a guard that has no chain. ResponseGuard reads a harmful verdict from a single pooled representation of the request, the response, and the image in one forward pass. Across a standard multimodal guardrail benchmark, our 2B ResponseGuard outperforms a recent 3B reasoning-based vision-language guard on response harmfulness detection, without any reasoning and at about 150 times lower time cost. On request harmfulness the reasoning guard retains an overall lead, and the remaining gap on both tracks sits on the image-only cells. We observe that the gap may stem from the frozen vision encoders that both designs use rather than from the missing chain. We have also found the reasoning guard directs almost none of its verdict attention to the image. Based on a single-pass detection, ResponseGuard can screen an answer sentence by sentence as it streams and stop a harmful answer before it finishes. For guarding the response of a vision-language model, a calibrated single-pass label may provide a sufficient safety signal. We fully release all source code, trained models, and datasets at https://github.com/ndb796/ResponseGuard.
While Large Language Models (LLMs) demonstrate remarkable performance on zero-shot annotation tasks, they often struggle with the specialized conventions of gold-standard benchmarks. We propose the systematic reuse and refinement of annotation guidelines as an alignment mechanism, introducing an iterative moderation framework that simulates the early phases of annotation projects. We evaluate three hypotheses: (1) the efficacy of guideline integration, (2) the advantage of reasoning optimized models, and (3) the viability of moderation under minimal supervision. Testing across biomedical NER tasks (NCBI Disease, BC5CDR, BioRED) with three LLM families (GPT, Gemini, DeepSeek), our results empirically confirm all three hypotheses. While the iterative moderation framework shows good potential in effectively refining guidelines, our analysis also reveals substantial room for improvement.