Adapting Reinforcement Learning with Chain-of-Thought Supervision for Explainable Detection of Hateful and Propagandistic Memes

2026-06-13 · Source: Artificial Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Natural Language Processing · Depth: Expert, quick

Summary

A new reinforcement learning-based post-training method enhances thinking-based multimodal large language models (MLLMs) for explainable detection of hateful and propagandistic memes. This approach, utilizing task-specific rewards and Group Relative Policy Optimization (GRPO), improves both classification performance and reference-based explanation quality. Researchers conducted an empirical study of off-the-shelf MLLMs on English and Arabic benchmarks, extending existing meme datasets with weakly supervised Chain-of-Thought (CoT) rationales and multi-LLM propaganda annotations. The method introduces a GRPO-based objective with thinking-length regularization, jointly optimizing accuracy and explanation quality, and explores self-supervised GRPO on unlabeled memes. Experiments on Hateful Memes and ArMeme benchmarks show FHM accuracy improvements up to +2.1% (from 79.9% to 82.0%) and ArMeme macro-F1 up to +7.6 points (from 0.536 to 0.612). The system also generates natural-language explanations, offering more balanced per-class performance on ArMeme compared to sequence-classification baselines.

Key takeaway

For Machine Learning Engineers developing content moderation systems, this research indicates that integrating reinforcement learning with Chain-of-Thought supervision can significantly boost MLLM accuracy and explanation quality for hateful memes. You should consider adapting GRPO-based post-training methods to fine-tune your multimodal models, especially when explainability is critical. This approach offers balanced per-class performance, which is vital for robust detection across diverse harmful content.

Key insights

Reinforcement learning with CoT supervision significantly improves MLLM performance and explainability for hateful meme detection.

Principles

• Memes require multimodal understanding.
• CoT rationales enhance MLLM explainability.
• RL can optimize classification and explanation.

Method

A GRPO-based objective with thinking-length regularization jointly optimizes MLLM classification accuracy and explanation quality, using task-specific rewards and weakly supervised CoT rationales. Self-supervised GRPO with pseudo-labels is also explored.

In practice

• Extend meme datasets with CoT rationales.
• Apply GRPO for MLLM fine-tuning.
• Use self-supervised GRPO on unlabeled data.

Topics

• Reinforcement Learning
• Chain-of-Thought
• Multimodal LLMs
• Hateful Meme Detection
• Explainable AI
• Group Relative Policy Optimization

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• Think When Needed: Adaptive Reasoning-Driven Multimodal Embeddings with a Dual-LoRA Architecture — Adaptive reasoning in multimodal embeddings improves both efficiency and retrieval quality by selectively generating Chain-of-Thought.
• Reinforcement Learning Improves LLM Accuracy and Reasoning in Disease Classification from Radiology Reports — Reinforcement learning with GRPO enhances LLM accuracy and reasoning in medical text classification without explicit reasoning supervision.
• Structured Role-Aware Policy Optimization for Multimodal Reasoning — SRPO improves LVLM multimodal reasoning by assigning role-aware, token-level credit for perception and reasoning.
• Socially Responsible and Explainable Automated Fact-Checking and Hate Speech Detection — Explainable NLP methods and high-quality datasets can jointly optimize performance and interpretability for Portuguese hate speech and fact-checking.
• I Know What You Meme, Even If it Emerged Today: Understanding Evolving Memes through Open-World Knowledge Acquisition — Dynamic meme understanding requires open-world knowledge acquisition to overcome outdated parametric model knowledge.
• LLMs Are Not Enough for Multimodal Fake News Detection: Why Global Label Propagation Helps — LLMs are better as weak annotators than direct classifiers for multimodal fake news detection.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.