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.
The Challenge of Multimodal Metaphor
Detecting harmful memes is a "hard mode" problem for artificial intelligence. Unlike standard text-based hate speech, memes are deeply metaphorical; their meaning emerges from the subtle interplay between an image and text, often requiring specific cultural context to understand.
Current state-of-the-art methods typically rely on a single Large Language Model (LLM) to reason about a meme. However, memes are products of social consensus. A single model might suffer from "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.
As illustrated above, a single MLLM might fail to see the harmful intent by focusing on surface-level features. To truly understand a meme, the system must simulate how different people from various backgrounds might collide and aggregate their viewpoints to reach a consensus.
Methodology: The SGoTRE Framework
The authors propose the Social Graph of Thought Reasoning Enhancement (SGoTRE) framework, which moves beyond single-opinion analysis through three main phases:
1. Diverse User Opinion Acquisition
Instead of a single prompt, the system employs a multi-agent simulation. It assigns various "personas" (derived from real-world social media data, including diverse occupations and interests) to different agents. Each agent generates a four-step Chain of Thought (CoT):
- Content Interpretation: What is literally depicted?
- Image-Text Relationship: How do the visual and textual elements relate?
- Stance Identification: What is the attitude toward the subjects?
- Metaphor Analysis: What is the deeper, implied meaning?
2. User Opinion Consensus Modeling
The system aggregates these diverse chains into a Social Graph of Thought (SGoT). Using hierarchical clustering, it identifies "consensus nodes"—points where multiple agents agree—and links these nodes across the four reasoning steps. This graph structure captures both the diversity of initial opinions and the eventual group consensus, providing a structural expression of social cognition.
3. Social Cognitive Reasoning Distillation
Running dozens of agents for every meme is computationally expensive and slow for real-time applications. To address this, the authors "distill" the knowledge from the complex SGoT (the teacher) into a student model, SGoT-R1. This student model is trained to internalize the collective logic, sequentially outputting the reasoning graph before providing the final "harmful/not harmful" label.
Experimental Results
The SGoT-R1 model was evaluated on three major public benchmarks: the Facebook Hateful Memes (FHM) dataset, Harm-C (COVID-19 related), and Harm-P (U.S. politics related).
Key findings include:
- State-of-the-Art Performance: SGoT-R1 achieved an 8.9% improvement in accuracy over previous SOTA models on the FHM dataset, rising from approximately 75.6% to 84.5%.
- Zero-Shot Superiority: In zero-shot settings, SGoT-enhanced versions consistently outperformed heavyweights like GPT-4o and Gemini 1.5 Flash.
- Efficiency: By internalizing the reasoning logic, the 7B version of SGoT-R1 processes memes in roughly 1.62 seconds, compared to over 15 seconds for a full multi-agent SGoT logic run through a base model like GPT-4o.
Discussion and Insights
The success of SGoT-R1 highlights the "wisdom of the simulated crowd." By forcing the model to consider multiple perspectives during the training phase, it learns to look past surface-level humor to identify hidden malice. Furthermore, representing reasoning as a graph rather than a flat string allows the model to maintain logical continuity, seeing how specific interpretations lead to specific stances and, ultimately, harmful metaphors.
While the initial cost of generating multi-agent data is high, the resulting distilled model offers a scalable and interpretable solution for social media moderation. Future work may involve expanding the diversity of agent personas and applying this "Social Graph of Thought" approach to other subjective tasks like sentiment analysis or complex social reasoning.
Summary updated