Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

AgentFoX: LLM Agent-Guided Fusion with eXplainability for AI-Generated Image Detection

Yangxin Yu, Yue Zhou, Bin Li, Kaiqing Lin, Haodong Li, Jiangqun Ni, Bo Cao

Abstract

The increasing realism of AI-Generated Images (AIGI) has created an urgent need for forensic tools capable of reliably distinguishing synthetic content from authentic imagery. Existing detectors are typically tailored to specific forgery artifacts--such as frequency-domain patterns or semantic inconsistencies--leading to specialized performance and, at times, conflicting judgments. To address these limitations, we present \textbf{AgentFoX}, a Large Language Model-driven framework that redefines AIGI detection as a dynamic, multi-phase analytical process. Our approach employs a quick-integration fusion mechanism guided by a curated knowledge base comprising calibrated Expert Profiles and contextual Clustering Profiles. During inference, the agent begins with high-level semantic assessment, then transitions to fine-grained, context-aware synthesis of signal-level expert evidence, resolving contradictions through structured reasoning. Instead of returning a coarse binary output, AgentFoX produces a detailed, human-readable forensic report that substantiates its verdict, enhancing interpretability and trustworthiness for real-world deployment. Beyond providing a novel detection solution, this work introduces a scalable agentic paradigm that facilitates intelligent integration of future and evolving forensic tools.

AgentFoX: LLM Agent-Guided Fusion with eXplainability for AI-Generated Image Detection

Abstract

The increasing realism of AI-Generated Images (AIGI) has created an urgent need for forensic tools capable of reliably distinguishing synthetic content from authentic imagery. Existing detectors are typically tailored to specific forgery artifacts--such as frequency-domain patterns or semantic inconsistencies--leading to specialized performance and, at times, conflicting judgments. To address these limitations, we present \textbf{AgentFoX}, a Large Language Model-driven framework that redefines AIGI detection as a dynamic, multi-phase analytical process. Our approach employs a quick-integration fusion mechanism guided by a curated knowledge base comprising calibrated Expert Profiles and contextual Clustering Profiles. During inference, the agent begins with high-level semantic assessment, then transitions to fine-grained, context-aware synthesis of signal-level expert evidence, resolving contradictions through structured reasoning. Instead of returning a coarse binary output, AgentFoX produces a detailed, human-readable forensic report that substantiates its verdict, enhancing interpretability and trustworthiness for real-world deployment. Beyond providing a novel detection solution, this work introduces a scalable agentic paradigm that facilitates intelligent integration of future and evolving forensic tools.
Paper Structure (24 sections, 10 equations, 11 figures, 8 tables)

This paper contains 24 sections, 10 equations, 11 figures, 8 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Signal-level Expert Models in AIGI Detection
  4. MLLM-based AIGI Detection
  5. Method
  6. Profile Investigation
  7. Agentic Inference
  8. Quick Integration
  9. Experiments
  10. X-Fuse: High-Conflict Benchmark
  11. Robustness to Expert Disagreement
  12. In-the-Wild Evaluation
  13. Ablation Study
  14. Conclusion
  15. Extended Benchmark Evaluation
  16. ...and 9 more sections

Figures (11)

  • Figure 1: Consensus and Disagreement Landscape Among AIGI Detectors. Samples from seven publicly available datasets(Sect. \ref{['sec:x-fuse']}). (Left) The UpSet plot quantitatively analyzes the performance of four expert detectors. (Right) Visual examples showcase these challenging cases of disagreement. Different experts fail on different images, creating conflicting signals.
  • Figure 2: Overview of the AgentFoX framework. Left: Profile Investigation constructs Expert Profiles and Clustering Profiles. Right: Agentic Inference leverages these profiles through a Toolkit to conduct multi‑stage forensic reasoning and compile an authenticity report.
  • Figure 3: Figure 3 showcases the agent's reasoning flow. After an ambiguous semantic analysis (Stage 1), it identifies conflicting expert opinions (Stage 2). During conflict resolution (Stage 3), the agent arbitrates by consulting both Expert and Clustering Profiles. Recognizing the cluster's unreliability, it down-weights this context and trusts the globally more reliable experts. This traceable adjudication culminates in the final report (Stage 4).
  • Figure 4: Sample sizes and cross-domain distribution for the decision--label--generator grouping. Some cells did not meet the target quota of $7\times 15$ due to limited global availability.
  • Figure 5: Accuracy and F1-Score of the proposed agentic aggregation method across 16 expert-decision combinations with varying conflict intensities. The upper panel shows performance metrics (blue solid circles and purple dashed squares); the lower panel shows the corresponding signal-level expert composition.
  • ...and 6 more figures