Fake-HR1: Rethinking reasoning of vision language model for synthetic image detection

Recent studies show Chain-of-Thought (CoT) reasoning can improve synthetic image detection, but lengthy reasoning incurs substantial resource overhead. Fake-HR1 proposes a large-scale hybrid-reasoning model that adaptively determines whether reasoning is necessary, based on generative detection task characteristics. It uses a two-stage training framework: Hybrid Fine-Tuning (HFT) for cold-start initialization, followed by online reinforcement learning with Hybrid-Reasoning Grouped Policy Optimization (HGRPO) to implicitly learn reasoning mode selection. Experimental results demonstrate Fake-HR1 adaptively performs reasoning, surpassing existing LLMs in reasoning ability and generative detection performance, while significantly improving response efficiency.