FACT: A Forensic Agent with Compiled Tool-Use Trajectories for AI-Generated Image Detection

Authors: Jiaoyang Chen, Bin Hu, Jingyu Hu, Kun Zhou, Qin Zhang, Zhengzhe Liu

Published: 2026-09-05 04:39:49+00:00

AI Summary

FACT (Forensic Agent with Compiled Tool-use Trajectories) addresses the open-world problem of AI-generated image detection by learning an image-conditioned tool-use policy for forensic analysis. Instead of relying on a fixed detector, FACT dynamically decides which forensic tools to use, interprets the evidence, and stops when sufficient evidence is gathered. This agentic approach, trained through an Evolve–Distill–Refine pipeline, significantly outperforms existing methods across various benchmarks, including those with unseen generators, deepfakes, and manipulated images.

Abstract

AI-generated image detection is increasingly open-world: new image generators produce highly realistic images that make visual artifacts harder to identify. Existing detectors usually rely on a fixed set of forensic cues, so a detector that works well for one generator family may fail on another. We introduce FACT (Forensic Agent with Compiled Tool-use Trajectories), which learns an image-conditioned tool-use policy for forensic analysis. Instead of applying a fixed detector, FACT decides which forensic tools to call, interprets the returned evidence, and stops when sufficient evidence has been collected. FACT follows an Evolve--Distill--Refine pipeline: it evolves an execution-verified forensic skill, compiles the skill into action--observation tool-use trajectories, distills them into a compact agent, and refines the policy with cost-aware GRPO. Across two internal and four public benchmarks, FACT achieves the best performance among all compared methods, including on recent unseen generators, deepfakes, and manipulated images.

Key findings

FACT consistently achieves state-of-the-art performance across six diverse benchmarks, outperforming individual expert detectors, majority voting, and other published baselines. Its ability to generalize to new, unseen generators, deepfakes, and manipulated images demonstrates its robustness and adaptability in open-world settings. The Evolve–Distill–Refine pipeline, especially the cost-aware GRPO, effectively optimizes evidence acquisition while maintaining high detection accuracy.

Approach

FACT learns an image-conditioned tool-use policy through an Evolve–Distill–Refine pipeline. It first evolves an execution-verified forensic skill using a general AI model (GPT-5.1), then compiles this skill into action–observation tool-use trajectories. These trajectories are distilled into a compact agent (Qwen3-VL-8B-Instruct) via supervised fine-tuning, and finally, the policy is refined with cost-aware GRPO to optimize tool selection and stopping decisions.

Datasets

STD3K, NewGen-900, Chameleon, LOKI, AIGCDetectionBenchmark, HydraFake

Model(s)

GPT-5.1 (for skill evolution and trajectory synthesis), Qwen3-VL-8B-Instruct (as the compact forensic agent)

Author countries

Hong Kong, China

← Previous