---
title: "HarmTrace: Anchor-Calibrated Decoupled Optimization for Fine-Grained Target Identification in Harmful Memes"
source: "arXiv AI/CL/LG（论文）"
source_tier: "t2"
url: "https://arxiv.org/abs/2608.16622v1"
permalink: "https://feed.astralor.com/items/itm_a1aa56da100fb077f52c0c9c"
published: "2026-08-17T14:21:41.000Z"
score: 81
channel: "ai"
tags: ["arXiv","HarmTrace","JRA","Meme3W","MLLMs","Harmful Content Detection"]
---

# HarmTrace: Anchor-Calibrated Decoupled Optimization for Fine-Grained Target Identification in Harmful Memes

## AI commentary

> This paper is worth reading because it pushes detection from binary classification toward accountable identification; for content moderation practitioners, locating the attacked target is far more useful than merely assigning a label. The JRA metric also exposes model weaknesses clearly, so I strongly recommend it to researchers in multimodal safety.

## At a glance

Existing harmful meme detection often treats the task as image-text harmfulness classification, so a model may correctly predict harmfulness while misidentifying the attacked target or supporting evidence. This work extends harmful meme detection with fine-grained target identification, requiring the model to output target category, target entity, textual mention, and visual region for harmful memes. The authors introduce Meme3W, which unifies multiple public harmful meme datasets with human-verified annotations, and propose Joint Record Accuracy (JRA), a strict record-level metric requiring the harmfulness label and all target-identification fields to be jointly correct. Experiments with representative multimodal large language models reveal a substantial gap between harmfulness accuracy and JRA. To narrow this gap, they propose HarmTrace, an anchor-calibrated decoupled optimization framework. HarmTrace strengthens target-entity supervision via entity-aware supervised fine-tuning and applies Conditional Target-identification Policy Optimization (CTPO) to decouple harmfulness and target-identification advantages, using a Virtual Positive Anchor (VPA) for normalization. On the Qwen3-VL-8B backbone, JRA improves from 17.58% to 52.51%. Code is publicly available.

## Key points

- Introduces fine-grained target identification requiring category, entity, text mention, and visual region outputs
- Builds Meme3W, a unified dataset with human-verified annotations from multiple public harmful meme datasets
- Proposes JRA, a strict record-level metric requiring joint correctness of harmfulness and all target fields
- HarmTrace boosts JRA from 17.58% to 52.51% on Qwen3-VL-8B

## Full text

Multimodal harmful meme detection is typically formulated as image--text harmfulness classification. A model may correctly predict harmfulness while misidentifying the attacked target or its supporting evidence. We therefore extend harmful meme detection with fine-grained target identification, asking what type of target is attacked, who is targeted, and where the target appears in the meme. The model predicts harmfulness for every meme and, for harmful memes, outputs the target category, target entity, textual mention, and visual region. To support this task, we introduce Meme3W, which unifies multiple public harmful meme datasets and provides human-verified annotations for harmful instances. We further introduce Joint Record Accuracy (JRA), a strict record-level metric requiring the harmfulness label and all target-identification fields to be jointly correct. Experiments with representative multimodal large language models reveal a substantial gap between harmfulness accuracy and JRA. To narrow this gap, we propose HarmTrace, an anchor-calibrated decoupled optimization framework. HarmTrace strengthens target-entity supervision through entity-aware supervised fine-tuning. It then applies Conditional Target-identification Policy Optimization (CTPO) to decouple harmfulness and target-identification advantages, restricting target-identification optimization to label-correct responses for harmful examples. CTPO uses a Virtual Positive Anchor (VPA) as a fully correct reference for target-identification advantage normalization. HarmTrace improves both JRA and harmfulness accuracy across the evaluated backbones, with JRA on the Qwen3-VL-8B backbone increasing from 17.58\% to 52.51\%. Our code is publicly available at https://github.com/llly1234/HarmTrace-for-Harmful-Memes.

---

Verify against the original before citing; curated by Astral Feed: https://feed.astralor.com/items/itm_a1aa56da100fb077f52c0c9c