New HarmTrace framework boosts accuracy in identifying harmful meme targets

Researchers have developed HarmTrace, a novel framework designed to improve the accuracy of identifying targets within harmful memes. This system addresses the limitation where models can correctly classify a meme as harmful but fail to pinpoint the specific target or supporting evidence. HarmTrace enhances target-entity supervision through entity-aware fine-tuning and employs Conditional Target-identification Policy Optimization (CTPO) to decouple harmfulness and target-identification performance. By using a Virtual Positive Anchor (VPA) for normalization, HarmTrace significantly boosts both overall harmfulness accuracy and fine-grained target identification, as demonstrated by a substantial increase in Joint Record Accuracy (JRA) on the Qwen3-VL-8B model. AI

IMPACT Enhances AI's ability to precisely identify targets in harmful memes, improving content moderation and safety.

RANK_REASON The cluster contains a research paper detailing a new method for a specific AI task. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• arXiv
• Conditional Target-identification Policy Optimization (CTPO)
• HarmTrace
• Hugging Face
• Joint Record Accuracy (JRA)
• Meme3W
• Qwen3 VL 8B
• Virtual Positive Anchor (VPA)

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →