Online memes are frequently used to spread unsafe content, including disinformation and propaganda. Their meaning shifts quickly in response to news events and relies heavily on contextual clues, which makes it difficult for platforms to assess whether a given meme is harmful or benign.
The paper introduces MemeTrace, a multimodal LLM-based agent built specifically for safety-aware provenance analysis. Rather than simply classifying a meme in isolation, the agent is designed to trace where a meme came from and how it has evolved, giving moderators more information about its intent and trajectory.
The abstract does not include experimental results or comparisons to other methods, so the reported effectiveness of MemeTrace is not yet clear from this source. Still, the work points toward a more context-sensitive approach to meme moderation, one that treats memes as dynamic artifacts rather than static images.