ThinkFake: Reasoning in Multimodal Large Language Models for AI-Generated Image Detection

2025-09-24

AI Summary

Current AI-generated image detection faces three key challenges: low accuracy, poor interpretability, and weak generalization. To address these, we propose the first reasoning-based universal detection framework leveraging multimodal large language models (MLLMs). Our method integrates forged-image reasoning prompts with Group-wise Relative Policy Optimization (GRPO), a novel reinforcement learning strategy, to establish a structured and interpretable image forgery analysis pipeline. Crucially, it requires no dense pixel-level annotations and introduces explicit reasoning—previously absent in AI image detection—for the first time. This enables strong cross-dataset generalization and zero-shot transfer capability. On the GenImage benchmark, our approach surpasses state-of-the-art methods; in the LOKI challenge, it achieves significant zero-shot detection gains. Comprehensive evaluation confirms its effectiveness, robustness, and inherent interpretability.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Computer Vision: Large Vision Models
• Natural Language Processing: Fact-Checking / Misinformation Detection (NLP Focus)

Application Category

• Social Networks and Social Media: Generative AI / large language models and their impact on social systems
• Search and Retrieval-Augmented AI: Retrieval-Augmented Generation (RAG) and multi-modal RAG
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

Abstract

The increasing realism of AI-generated images has raised serious concerns about misinformation and privacy violations, highlighting the urgent need for accurate and interpretable detection methods. While existing approaches have made progress, most rely on binary classification without explanations or depend heavily on supervised fine-tuning, resulting in limited generalization. In this paper, we propose ThinkFake, a novel reasoning-based and generalizable framework for AI-generated image detection. Our method leverages a Multimodal Large Language Model (MLLM) equipped with a forgery reasoning prompt and is trained using Group Relative Policy Optimization (GRPO) reinforcement learning with carefully designed reward functions. This design enables the model to perform step-by-step reasoning and produce interpretable, structured outputs. We further introduce a structured detection pipeline to enhance reasoning quality and adaptability. Extensive experiments show that ThinkFake outperforms state-of-the-art methods on the GenImage benchmark and demonstrates strong zero-shot generalization on the challenging LOKI benchmark. These results validate our framework's effectiveness and robustness. Code will be released upon acceptance.

Problem

• Detecting AI-generated images to combat misinformation and privacy violations
• Overcoming limited generalization in existing binary classification methods
• Providing interpretable detection through reasoning-based structured outputs

Innovation

• Uses MLLM with forgery reasoning prompt for detection
• Employs GRPO reinforcement learning with reward functions
• Implements structured detection pipeline for reasoning quality