Perception, Verdict, and Evolution: Hindsight-Driven Self-Refining Forensics Agent for AI-Generated Image Detection
Abstract Overview
This paper presents ForeAgent, an agentic framework for detecting AI-generated images that combines a Perception-Verdict architecture with a hindsight-driven self-refinement loop. The system aggregates semantic cues from the original image, frequency-domain cues from wavelet transforms, and spatial cues from an external NPR detector, then uses an MLLM to produce an interpretable final verdict. During training, it iteratively reflects on misclassifications and weak reasoning traces, filters regenerated samples through a dual-expert quality-gating process, and fine-tunes itself on the curated data. The experiments evaluate both detection accuracy and reasoning quality on AIGCDetectBenchmark and Chameleon.
Novelty
The distinctive aspect of the work is the combination of multi-view forensic evidence with an iterative self-evolution mechanism that uses hindsight from ground-truth labels to regenerate and curate better reasoning traces. Unlike approaches that rely on static synthetic supervision or modify the backbone architecture, ForeAgent keeps the MLLM backbone intact and improves it through sampling, reflection, dual-expert gating, and repeated fine-tuning.
Results
ForeAgent reaches 82.18% overall accuracy on the Chameleon benchmark, outperforming AIDE by 16.41 percentage points, and achieves 93.3% mean accuracy on AIGCDetectBenchmark across 16 generators. It also obtains the highest reported reasoning-quality score in the paper's comparison (5.35), exceeding GPT-5-mini (4.81) and GPT-5 (4.15). The paper also notes a weaker result on WFIR, where several CNN-based detectors remain stronger.
Key Points
- ForeAgent fuses semantic, frequency-domain, and spatial forensic cues and uses an MLLM as a verdict module to generate evidence-grounded detection decisions.
- Its hindsight-driven self-refining training loop turns errors and weak rationales into new supervision through reflection and dual-expert quality gating, reducing reliance on external frontier models or human annotation.
- Empirically, the method reports state-of-the-art performance on Chameleon and strong cross-generator generalization on AIGCDetectBenchmark, while still showing limitations on some sources such as WFIR.