Technical Summary: ForeAgent – Hindsight-Driven Self-Refining Forensics Agent for AI-Generated Image Detection

1. Problem Statement

The rapid advancement of generative models (GANs, Diffusion Models, Autoregressive Models) has narrowed the perceptual gap between authentic and AI-generated images, posing significant challenges to digital forensics and copyright protection. Existing detection methods face two primary limitations:

1. Insufficient Sensitivity: Feature-based detectors often rely on statistical pattern matching (e.g., up-sampling traces) and lack high-level semantic reasoning, making them unable to assess intrinsic logical consistency. Conversely, Multimodal Large Language Models (MLLMs) possess strong general visual understanding but exhibit limited sensitivity to the fine-grained, subtle artifacts characteristic of AI-generated imagery.
2. Static Supervision & High Cost: Current MLLM-based approaches heavily depend on static synthetic instruction data generated by frontier models (e.g., GPT-4o, Pixtral-124B) or costly human annotations. This reliance limits flexibility and prevents the models from engaging in iterative self-refinement or learning from their own misclassified samples.

2. Methodology: ForeAgent

The authors propose ForeAgent, an agentic forensic framework designed for iterative self-evolution. The framework operates through two core mechanisms: a Perception–Verdict Architecture and a Hindsight-Driven Self-Refining Strategy.

A. Perception–Verdict Architecture

Unlike prior methods that modify backbone architectures or train task-specific heads, ForeAgent preserves the MLLM (specifically Qwen3-VL-8B) in its original form, utilizing it as a centralized reasoning engine. It aggregates multi-view cues through three complementary dimensions:

1. Semantic Cues: The original image is provided to capture high-level visual inconsistencies (e.g., unnatural object compositions, implausible lighting).
2. Frequency-Domain Cues: A wavelet transform (specifically the diagonal detail sub-band, cD) is applied to the image to expose spectral anomalies imperceptible in the pixel domain.
3. Spatial Cues: An external, pre-trained NPR (Neighboring Pixel Relationships) detector is invoked as a callable tool to extract structural traces from up-sampling operations.

The MLLM receives the original image and the wavelet-transformed image as dual-image inputs, along with the structured inference results from the NPR tool. A dedicated Verdict Module then synthesizes these signals to produce a coherent, logically grounded decision, including a binary label, confidence score, and an interpretable reasoning trace.

B. Hindsight-Driven Self-Refining

To enable continual improvement without external frontier models or human annotation, ForeAgent employs a Sampling–Reflection–Evolution paradigm:

1. Sampling: The agent performs inference rollouts on training shards. Trajectories are categorized into: (a) Incorrect Predictions, (b) Low-Quality Reasoning (correct label but vague logic), and (c) High-Quality Reasoning.
2. Reflection: Guided by ground-truth labels as "hindsight," the agent re-examines failure cases (groups a and b). It regenerates improved reasoning traces, either correcting the verdict or refining the rationale to be more evidence-grounded.
3. Dual-Expert Quality Gating (DEG): To ensure rigor, regenerated samples are evaluated by two independent MLLMs (ForeAgent itself and Qwen3-VL-Plus). A sample is accepted into the high-quality training set only if both experts agree it meets strict criteria (logic consistency, evidence sufficiency, causal clarity).
4. Evolution: The curated high-quality samples are used for Supervised Fine-Tuning (SFT) via LoRA. The model parameters are updated only if the new checkpoint outperforms the previous best on a test set, ensuring monotonic improvement.

3. Key Contributions

1. ForeAgent Framework: An agentic multi-view forensic framework that unifies semantic, spatial, and frequency-domain cues under a Perception–Verdict architecture without modifying the MLLM backbone.
2. Self-Refining Paradigm: A Hindsight-Driven Self-Refining strategy that converts failure cases into high-value supervision signals through a Sampling–Reflection–Evolution loop with dual-expert quality gating, enabling continuous self-evolution.
3. State-of-the-Art Performance: Extensive experiments demonstrating that ForeAgent achieves superior detection accuracy and reasoning quality compared to existing feature-based detectors, other MLLM-based methods, and proprietary models like GPT-5.

4. Experimental Results

The paper evaluates ForeAgent on two primary benchmarks:

• AIGCDetectBenchmark: Across 16 diverse generators, ForeAgent achieves a mean accuracy of 93.3%, outperforming the strongest baselines, AIGI-Holmes (93.2%) and AIDE (92.8%). It shows particular robustness on modern diffusion-based generators (e.g., 96.0% on Midjourney, 98.9% on DALLE2).
• Chameleon Benchmark: ForeAgent achieves an overall accuracy of 82.18%, surpassing AIDE by 16.41% and outperforming GPT-5 (81.91%). Notably, it maintains high accuracy on real images (96.69%) while significantly improving fake detection (61.86%) compared to other open-source MLLMs.
• Reasoning Quality: Evaluated on 2,000 samples using an external judge (Gemini-3-flash-preview), ForeAgent achieved the highest average reasoning quality score (5.35 on a 0-10 scale), outperforming GPT-5 (4.15) and GPT-5-mini (4.81). It demonstrated superior performance in logical consistency (75.6%) and causal clarity (76.1%).

Ablation Studies:

• Wavelet Components: The diagonal detail sub-band (cD) was found to be the most discriminative frequency-domain cue.
• NPR Integration: Adding the NPR spatial expert provided complementary local statistical cues, improving overall accuracy.
• Dual-Expert Gating: Using two independent models for quality control significantly reduced bias and improved both accuracy and reasoning quality compared to single-expert gating.
• Self-Evolution vs. GPT-Synthesis: Training on self-refined data yielded better balance between real/fake detection and higher reasoning quality compared to training on data synthesized by GPT-5, which tended to over-detect fakes and hallucinate.

5. Significance and Claims

The paper claims that ForeAgent addresses the critical limitations of current AI-generated image detection by moving away from static, externally supervised models toward an iterative, self-evolving agent. Its significance lies in:

• Autonomy: It reduces reliance on costly human annotations or proprietary frontier models for training data generation.
• Interpretability: It provides logically grounded, evidence-aware reasoning traces that are more consistent and causally clear than those produced by GPT-5.
• Robustness: The hindsight-driven mechanism allows the model to continuously adapt to new failure modes, transforming errors into learning signals and progressively sharpening detection capabilities.

The authors acknowledge limitations, noting reduced performance on specific sources with distinct characteristics (e.g., WFIR) and the current reliance on a fixed set of multi-view features. Future work is proposed to explore adaptive feature selection, reinforcement learning for real-time adaptation, and extension to fine-grained manipulation localization.