Projects

HVGuard studies hateful video moderation, where harm is often conveyed jointly through speech, visuals, sarcasm, and pacing instead of a single explicit cue. The system combines transcripts and video frames in a multimodal LLM pipeline so the model can reason over cross-modal evidence rather than treating a clip as isolated text or isolated imagery.

This makes the approach more effective for implicit and context-dependent hate in real short-video settings. The project shows how structured multimodal reasoning can improve practical moderation for video platforms where harmful intent is often concealed behind humor, editing style, or audiovisual mismatch.

HMGuard focuses on harmful memes, a moderation problem where a small amount of text and imagery can hide hateful, harassing, or propagandistic intent. The framework uses multimodal large language models to reason about the relationship between the image, the overlaid text, and the broader social meaning conveyed by the meme.

The work treats meme moderation as an understanding problem rather than a pure classification task. That perspective is important for real-world moderation because harmful meaning often depends on cultural references, visual composition, and subtle multimodal cues that standard detectors miss.

UGCG-Guard targets promotional content used to lure users into unsafe user-generated content games. The system screens social posts, screenshots, and game-related imagery with large vision-language models to detect sexualized, exploitative, or otherwise illicit promotion before it spreads across the platform ecosystem.

The project is designed for the messy moderation setting around creator-driven games, where harmful content mixes platform-native slang, visual signals, and rapidly shifting promotion styles. It shows how LVLM-based moderation can protect vulnerable users, especially minors, in large UGC game communities.

NewWave studies hate speech that surges around breaking events, where static moderation policies and older classifiers quickly go stale. The paper introduces an LLM-based reasoning framework that captures the narratives, slogans, and contextual references tied to newly emerging hate waves triggered by real-world events.

Instead of assuming that the target classes are fixed, the project treats moderation as a continual adaptation problem. The result is a more responsive way to track event-driven abuse and update detection strategies without retraining a new model from scratch for every shift in online discourse.