Research map / Large language models / Reasoning and chain of thought
SGoT-R1: Social Graph of Thought Reasoning-Enhanced Multimodal Large Language Model for Harmful Meme Detection
Xiuxian Wang, Yuting Su, Wenhui Li, Xiaowen Wang, Zhuojun Li, Anan Liu
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes SGoT-R1, a reasoning-enhanced multimodal LLM for harmful meme detection that builds on social graph-of-thought style reasoning (abstract truncated; details beyond the abstract are not available).
Problem
Harmful meme detection requires understanding metaphorical multimodal content; prior LLM-reasoning-based methods reason only over the meme itself, without considering broader social context.
Approach
Extends reasoning for harmful meme detection beyond the meme content to social context via a social graph of thought, as suggested by the title; full method details unavailable from the truncated abstract.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add SGoT as input in zero-shot inference (ablation) | GPT-4o · FHM | accuracy: 68.8 -> 72.1 | Improved |
| add SGoT as input in zero-shot inference (ablation) | GPT-4o · Harm-P | accuracy: 63.1 -> 80.6 | Improved |
| add SGoT as input in zero-shot inference (ablation) | Gemini 1.5 Flash · FHM | accuracy: 60.2 -> 67.8 | Improved |
| add SGoT as input in zero-shot inference (ablation) | Gemini 1.5 Flash · Harm-P | accuracy: 61.6 -> 78.2 | Improved |
| add SGoT as input in zero-shot inference (ablation) | GPT-4o · Harm-C | accuracy: 67.5 -> 73.2 | Improved |
| distill SGoT reasoning into the model instead of generating SGoT with base model at inference (ablation) | GPT-4o w/ SGoT vs SGoT-R1-7B · FHM | inference time: 15.3s -> 1.62s | Improved |
| add SGoTRE training instead of plain SFT (ablation) | Qwen2VL-7B · FHM | accuracy: 78.6 -> 84.5 | Improved |
| add SGoTRE training instead of plain SFT (ablation) | Qwen2VL-2B · FHM | accuracy: 76.2 -> 79.4 | Improved |
| add SGoTRE training instead of plain SFT (ablation) | Qwen2VL-2B · Harm-P | accuracy: 80.3 -> 89.0 | Improved |
| add SGoTRE training instead of plain SFT (ablation) | Qwen2VL-7B · Harm-C | accuracy: 85.9 -> 89.0 | Improved |
| add SFT to zero-shot Qwen2VL-7B (ablation) | Qwen2VL-7B · FHM | accuracy: 63.8 -> 78.6 | Improved |
| scale SGoT-R1 from 2B to 7B | SGoT-R1 · FHM | accuracy: 79.4 -> 84.5 | Improved |
Limitations
- Only the truncated abstract was available; full method and results unknown
Closest papers by meaning
- Towards Comprehensive Detection of Chinese Harmful Memes (NeurIPS 2024 · similarity 0.84)
- Hateful Meme Detection through Context-Sensitive Prompting and Fine-Grained Labeling (Student Abstract) (AAAI 2025 · similarity 0.83)
- TRACE: Textual Relevance Augmentation and Contextual Encoding for Multimodal Hate Detection (AAAI 2026 · similarity 0.83)
- Multi-Granular Multimodal Clue Fusion for Meme Understanding (AAAI 2025 · similarity 0.83)
- Multi-Modal Discussion Transformer: Integrating Text, Images and Graph Transformers to Detect Hate Speech on Social Media (AAAI 2024 · similarity 0.83)
- A Multi-Task Benchmark for Abusive Language Detection in Low-Resource Settings (NeurIPS 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
