ForensicCPN: Towards Self-Evolving Forensic Skill for Image Forgery Detection and Localization

Multimodal large language models (MLLMs) with external forensic tools can combine semantic reasoning with specialized evidence for image forgery detection and localization. Existing methods, however, largely encode tool-use experience in model parameters or predefined workflows, rather than learning it as a directly revisable forensic skill. This limits the ability to directly update forensic experience when new attacks and tools emerge. We propose ForensicCPN, which treats the executable forensic skill as the learning target while keeping the MLLM and expert tools frozen. Inspired by Colored Petri Nets (CPNs), ForensicCPN represents forensic experience as an executable skill graph. The agent first proposes a hypothesis about the manipulation and acquires evidence to verify it. As evidence accumulates, the hypothesis is confirmed, revised, or resolved into a verdict, with the evolving context carried through the CPN. A forensic harness executes the skill by maintaining its evolving evidence state and enforcing graph-defined actions. The skill is self-evolving through both skill learning and test-time skill evolving: multiple agents distill labeled execution trajectories into validated revisions, while unresolved cases trigger label-free exploration whose reusable discoveries are consolidated back into the skill. Across seven conventional manipulation benchmarks, ForensicCPN achieves the strongest aggregate detection performance while maintaining competitive region-localization performance among the compared methods. On AI inpainting and global synthesis, test-time skill evolving further improves adaptation without parameter updates.