Technical Summary: ClueAegis

Problem Statement

The rapid advancement of generative artificial intelligence has produced synthetic images that are increasingly realistic, challenging the reliability of existing detection methods. Current approaches predominantly rely on unified end-to-end classification or monolithic reasoning pipelines. These methods often fail to model structured forensic reasoning and struggle to handle heterogeneous visual evidence. Forensic cues—such as lighting inconsistencies, frequency artifacts, OCR errors, and anatomical anomalies—exhibit fundamentally different distributions. Forcing these diverse cues into a single, undifferentiated reasoning space can lead to feature entanglement and suboptimal evidence extraction. Furthermore, existing datasets typically organize data by generator type rather than by structured forensic heuristics, hindering the learning of decomposed cognitive skills.

Methodology

The authors propose ClueAegis, a framework that re-frames synthetic image detection as a Heuristic-to-Reasoning cognitive skill learning process, inspired by the dual-process theory of human cognition (System 1: heuristic perception; System 2: analytical reasoning).

1. ClueAegis-Bench

To support this paradigm, the authors introduce ClueAegis-Bench, a large-scale benchmark (120K images) that decomposes detection into 12 explicitly annotated forensic cognitive skills. Unlike binary benchmarks, each image is labeled with the specific skill most effective at exposing its forgery. The 12 skills include:

• Lighting/Shadow/Physical Consistency: Evaluating illumination, shadow coherence, and physical laws.
• Common Sense/Functional Inconsistency: Detecting semantic or functional impossibilities.
• OCR/Region/Transformation Consistency: Analyzing text, local regions, and post-processing artifacts.
• Human/Animal Anatomy: Checking structural plausibility.
• Frequency/Pixel Consistency: Examining spectral and low-level pixel distributions.

2. The ClueAegis Framework

ClueAegis operates as a two-stage agentic framework:

• Stage I: Heuristic-Guided Skill Selection (System 1): The model acts as a perception-driven agent. It extracts perceptual forensic clues from the input image and selects the most relevant forensic skill from a predefined library. This transforms cognitive understanding into an actionable plan.
• Stage II: Skill-Guided Forensic Reasoning (System 2): Based on the selected skill, the model performs evidence-driven reasoning. It dynamically invokes corresponding analysis tools (e.g., OCR engines, lighting analyzers, frequency spectrum tools) and feature transformation operators. The reasoning process generates a trajectory that integrates external tool outputs to make a final authenticity prediction.

3. Training Strategy

The model is trained using a two-stage approach:

1. Supervised Fine-Tuning (SFT): The model learns to map forensic evidence to the correct skill and perform skill-conditioned reasoning.
2. Group Relative Policy Optimization (GRPO): A preference optimization step is applied to improve reasoning consistency, skill selection accuracy, and format compliance. The reward function balances answer correctness, skill selection consistency, and adherence to the two-stage reasoning structure.

Key Contributions

• Task Re-formulation: The paper shifts synthetic image detection from binary classification to a skill-driven, structured reasoning problem.
• ClueAegis-Bench: A novel benchmark with skill-oriented annotations that organizes heterogeneous artifacts into a unified skill space, enabling structured evaluation beyond simple real/fake classification.
• Skill-Driven Classification Space: The definition of 12 forensic skills corresponding to different reasoning capabilities, forming an adaptive space for forensic analysis.
• Two-Stage Forensic Reasoning: A heuristic-to-reasoning framework that separates global perceptual understanding (skill selection) from localized, evidence-guided reasoning.

Experimental Results

Extensive experiments demonstrate that ClueAegis achieves state-of-the-art (SOTA) performance across multiple benchmarks:

• ClueAegis-Bench: Outperforms previous SOTA methods by 6.99% in accuracy, achieving an average of 99.01%.
• AIGI-Now (Generalization): Shows strong generalization to recent generative models (e.g., FLUX, gpt4o), surpassing the previous SOTA by 7.88%.
• GenImage: Achieves the best overall performance on the large-scale GenImage benchmark, improving accuracy by 0.67% over existing methods.
• Robustness: The method demonstrates superior stability under common image perturbations (JPEG compression, Gaussian blur, resizing) compared to baselines, attributed to the skill-conditioned framework operating on abstract forensic cues rather than fragile pixel-level patterns.
• Ablation Studies: Results confirm that larger backbone models improve performance and that the full pipeline (Skill selection + SFT + GRPO) is necessary for optimal results.

Significance and Claims

The paper claims that ClueAegis offers a more explainable alternative to conventional end-to-end detectors by providing transparent reasoning trajectories and structured forensic evidence. By bridging perception, skill selection, and forensic reasoning, the framework addresses the limitations of feature entanglement in current methods. The authors posit that this skill-aware approach promotes trustworthy synthetic media detection by enabling adaptive routing and evidence verification across diverse forgeries.

The authors acknowledge limitations, noting that the framework relies on a predefined set of skills which may not cover all emerging manipulation patterns, and that performance is partially dependent on the underlying vision-language backbone's capability. Additionally, the structured reasoning pipeline introduces computational overhead compared to direct binary classification.