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

TL;DR

SGoT-R1 introduces a framework that detects harmful memes by simulating "group consensus" rather than relying on a single AI's opinion. By distilling the collective reasoning of multiple simulated personas into a single multimodal model, it achieves a massive 8.9% accuracy boost on challenging benchmarks while remaining efficient for real-world use.

What Problem Does This Paper Address?

Detecting harmful memes is a "hard mode" problem for AI. Unlike standard hate speech, memes are deeply metaphorical; their meaning emerges from the interplay between an image and text, often requiring specific cultural context to understand.

Current state-of-the-art methods typically ask a single Large Language Model (LLM) to reason about a meme. However, the authors argue that memes are products of social consensus. A single model might have a "cognitive bias"—for example, misinterpreting a meme as "cute" because it features puppies, while a human group would recognize it as a disparaging metaphor for a specific demographic. To truly understand a meme, you need to simulate how different people from various backgrounds might collide and aggregate their viewpoints.

Methods

The authors propose the Social Graph of Thought Reasoning Enhancement (SGoTRE) framework, which consists of three main phases:

  1. Diverse User Opinion Acquisition: Instead of one prompt, the system uses a multi-agent simulation. It assigns various "personas" (based on real-world Twitter data, including occupations and interests) to different agents. Each agent generates a four-step Chain of Thought (CoT):
  2. Content Interpretation: What is literally there?
  3. Image-Text Relationship: How do they relate?
  4. Stance Identification: What is the attitude toward the subjects?
  5. Metaphor Analysis: What is the deeper, implied meaning?
  6. User Opinion Consensus Modeling: The system takes these diverse chains and builds a Social Graph of Thought (SGoT). It uses hierarchical clustering to find "consensus nodes"—points where multiple agents agree—and links these nodes across the four reasoning steps. This graph captures both the diversity of opinions and the eventual group consensus.
  7. Social Cognitive Reasoning Distillation: Running dozens of agents for every meme is too slow and expensive. To fix this, the authors "distill" the teacher (the complex graph) into a student model (SGoT-R1). The student is trained to take a meme and sequentially output the reasoning graph before providing the final "harmful/not harmful" label.

Datasets

The model was evaluated on three major public benchmarks:

  • FHM (Facebook Hateful Memes): A widely used, challenging dataset designed to foil simple classifiers.
  • Harm-C: Focused on memes related to COVID-19.
  • Harm-P: Focused on memes related to U.S. politics.

Key Insights and Ideas

  • The Wisdom of the (Simulated) Crowd: The paper proves that "group thinking" in AI leads to more robust results. By forcing the model to consider multiple perspectives during training, it learns to look past surface-level humor to find hidden malice.
  • Structure Matters: Representing reasoning as a graph rather than a flat string of text allows the model to capture "logical continuity." It sees how a specific interpretation of an image leads to a specific stance, and eventually, a harmful metaphor.
  • Efficiency via Distillation: One of the biggest "aha" moments is that you don't need to run a multi-agent system at inference time. By training the model to internalize that collective logic, SGoT-R1 is significantly faster than multi-agent approaches while maintaining their accuracy.

Key Results

SGoT-R1 set new state-of-the-art (SOTA) records across the board:

  • FHM Dataset: Achieved an 8.9% improvement in accuracy over previous SOTA models (rising from ~75.6% to 84.5%).
  • Zero-Shot Superiority: When compared to heavyweights like GPT-4o and Gemini 1.5 Flash in zero-shot settings, the SGoT-enhanced versions consistently outperformed the base models by significant margins.
  • Speed: SGoT-R1 (7B version) takes about 1.62 seconds per meme, whereas running the full SGoT logic through a base model like GPT-4o takes over 15 seconds.

Limitations and Future Work

While the results are impressive, the paper notes a few areas for growth:

  • Initial Cost: The "teacher" phase requires generating a massive amount of multi-agent data, which is computationally expensive upfront.
  • Persona Dependency: The diversity of the "group consensus" is only as good as the personas injected into the agents. If the initial user pool is biased, the consensus might be too.
  • Future Directions: The authors suggest that this "Social Graph of Thought" approach could be applied to other subjective tasks beyond memes, such as sentiment analysis or complex social media moderation.

Summary updated

Abstract

Internet memes serve as widely distributed multimodal social content that conveys complex ideas through metaphorical expressions, often containing harmful implications that make accurate harmful meme detection an important problem. Reasoning knowledge extracted from large language models plays a crucial role in recent advances in harmful meme detection. However, these methods only perform reasoning analysis on memes from a single opinion, ignoring that memes are essentially products of group consensus, where their true meaning interpretation highly depends on the collision and aggregation process of diverse user viewpoints. To address this problem, we propose a Social Graph of Thought Reasoning Enhancement (SGoTRE) framework for harmful meme detection. The SGoTRE contains three key steps: First, through multi-agent simulation technology, we obtain diverse chains of thought that represent the parsing logic of users from different backgrounds toward memes, authentically restoring the diversity characteristics of group cognition. Second, we construct a Social Graph of Thought (SGoT) that effectively integrates multi-chain reasoning processes and structurally expresses the consensus and diversity of viewpoints among users. Finally, we utilize the SGoT for cognitive distillation, internalizing multi-opinion reasoning logic into a single multimodal large model SGoT-R1 to achieve efficient and interpretable harmful meme detection. Experimental results show that SGoT-R1 significantly improves detection performance on mainstream datasets. Particularly on the most challenging FHM dataset, SGoT-R1 achieves an 8.9% improvement over state-of-the-art models.
BibTeX
@inproceedings{wang2026sgot,
  author = {Xiuxian Wang and Yuting Su and Wenhui Li and Xiaowen Wang and Zhuojun Li and Anan Liu},
  title = {{SGoT-R1: Social Graph of Thought Reasoning-Enhanced Multimodal Large Language Model for Harmful Meme Detection}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2026},
  doi = {10.1609/aaai.v40i31.39868},
}

Loading related content...

Figure