Perception, Verdict, and Evolution: Hindsight-Driven Self-Refining Forensics Agent for AI-Generated Image Detection

As generative models produce increasingly realistic images, traditional forensic methods struggle to detect subtle artifacts. While Multimodal Large Language Models (MLLMs) offer potential for explainable detection, they often lack sensitivity to fine-grained forensic traces and rely on costly, static supervision. This paper asks: Can an agentic framework, capable of self-reflection and iterative evolution, improve the detection of AI-generated images without relying on external, high-cost annotations?

The authors introduce ForeAgent, which employs a two-part architecture. First, the Perception-Verdict architecture aggregates three types of cues: semantic (original image), frequency-domain (wavelet transforms), and spatial (via an external NPR detector). These are fused by an MLLM to produce a logically grounded verdict. Second, the framework uses a Hindsight-Driven Self-Refining strategy based on a Sampling-Reflection-Evolution paradigm. During training, the agent identifies failure cases and uses ground-truth labels to reflect on and regenerate higher-quality reasoning traces. These traces are filtered through a dual-expert quality gating module, ensuring that only high-quality samples are used for subsequent fine-tuning.

ForeAgent achieves state-of-the-art performance on major benchmarks, reaching 82.18% accuracy on the Chameleon dataset—a 16.41% improvement over the AIDE baseline—and 93.3% mean accuracy on the AIGCDetect-Benchmark across 16 different generators. Furthermore, qualitative evaluations indicate that ForeAgent produces more consistent and causally grounded reasoning compared to frontier models like GPT-5 and GPT-5-mini, demonstrating the effectiveness of its iterative self-evolution process.

This work provides a scalable path for improving forensic tools by reducing reliance on expensive human-labeled data. By enabling models to learn from their own mistakes through a structured reflection process, ForeAgent offers a more flexible and robust approach to the ongoing arms race between generative models and detection systems.