Yujia Hu
Publications
Structured Visual Narratives Undermine Safety Alignment in Multimodal Large Language Models
Multimodal Large Language Models (MLLMs) extend text-only LLMs with visual reasoning, but also introduce new safety failure modes under visually grounded instructions. We study comic-template jailbreaks that embed harmful goals inside simple three-panel visual narratives and prompt the model to role-play and "complete the comic." Building on JailbreakBench and JailbreakV, we introduce ComicJailbreak, a comic-based jailbreak benchmark with 1,167 attack instances spanning 10 harm categories and 5 task setups. Across 15 state-of-the-art MLLMs (six commercial and nine open-source), comic-based attacks achieve success rates comparable to strong rule-based jailbreaks and substantially outperform plain-text and random-image baselines, with ensemble success rates exceeding 90% on several commercial models. Then, with the existing defense methodologies, we show that these methods are effective against the harmful comics, they will induce a high refusal rate when prompted with benign prompts. Finally, using automatic judging and targeted human evaluation, we show that current safety evaluators can be unreliable on sensitive but non-harmful content. Our findings highlight the need for safety alignment robust to narrative-driven multimodal jailbreaks.
SEAHateCheck: Functional Tests for Detecting Hate Speech in Low-Resource Languages of Southeast Asia
Hate speech detection relies heavily on linguistic resources, which are primarily available in high-resource languages such as English and Chinese, creating barriers for researchers and platforms developing tools for low-resource languages in Southeast Asia, where diverse socio-linguistic contexts complicate online hate moderation. To address this, we introduce SEAHateCheck, a pioneering dataset tailored to Indonesia, Thailand, the Philippines, and Vietnam, covering Indonesian, Tagalog, Thai, and Vietnamese. Building on HateCheck's functional testing framework and refining SGHateCheck's methods, SEAHateCheck provides culturally relevant test cases, augmented by large language models and validated by local experts for accuracy. Experiments with state-of-the-art and multilingual models revealed limitations in detecting hate speech in specific low-resource languages. In particular, Tagalog test cases showed the lowest model accuracy, likely due to linguistic complexity and limited training data. In contrast, slang-based functional tests proved the hardest, as models struggled with culturally nuanced expressions. The diagnostic insights of SEAHateCheck further exposed model weaknesses in implicit hate detection and models' struggles with counter-speech expression. As the first functional test suite for these Southeast Asian languages, this work equips researchers with a robust benchmark, advancing the development of practical, culturally attuned hate speech detection tools for inclusive online content moderation.
HateXScore: A Metric Suite for Evaluating Reasoning Quality in Hate Speech Explanations
Hateful speech detection is a key component of content moderation, yet current evaluation frameworks rarely assess why a text is deemed hateful. We introduce \textsf{HateXScore}, a four-component metric suite designed to evaluate the reasoning quality of model explanations. It assesses (i) conclusion explicitness, (ii) faithfulness and causal grounding of quoted spans, (iii) protected group identification (policy-configurable), and (iv) logical consistency among these elements. Evaluated on six diverse hate speech datasets, \textsf{HateXScore} is intended as a diagnostic complement to reveal interpretability failures and annotation inconsistencies that are invisible to standard metrics like Accuracy or F1. Moreover, human evaluation shows strong agreement with \textsf{HateXScore}, validating it as a practical tool for trustworthy and transparent moderation. \textcolor{red}{Disclaimer: This paper contains sensitive content that may be disturbing to some readers.}