BusterX++: Towards Unified Cross-Modal AI-Generated Content Detection and Explanation with MLLM

Researchers have introduced BusterX++, a unified multimodal large language model designed to detect and explain AI-generated content across both images and videos, according to a paper posted to arXiv [1]. The model employs a single-stage reinforcement learning strategy that its creators say matches or exceeds conventional two-stage training approaches [2]. The system, detailed in a submission last revised in June 2026, addresses what its authors describe as a growing risk of multimodal visual misinformation driven by advances in generative AI [2]. Existing detection methods typically handle image and video forensics as separate problems, leaving potential cross-modal benefits untapped [2]. BusterX++ is paired with a new evaluation suite called GenBuster-Bench++, a difficulty-aligned benchmark containing balanced image and video samples from recent generation models and real-world scenarios [2]. The paper’s central technical finding concerns the post-training paradigm commonly used for large language models. Many pipelines follow a supervised fine-tuning stage before applying reinforcement learning, a sequence often abbreviated as SFT→RL [2]. The BusterX++ team revisited this convention and found that a single-stage, pure RL approach—driven only by sparse outcome rewards—consistently matched or surpassed a strong SFT+RL baseline in both unified and single-modality settings [2]. The authors attribute this result to policy entropy. Supervised fine-tuning, they argue, imposes lower policy entropy, which narrows the policy search space and limits exploratory freedom [2]. In contrast, the single-stage pure RL method maintains higher policy entropy throughout training, which the researchers say unlocks the spontaneous emergence of cross-modal capability transfer between image and video forensics [2]. The paper reports state-of-the-art performance on the GenBuster-Bench++ benchmark [2]. The work was submitted by Haiquan Wen and colleagues, with the initial version posted in July 2025 and the fourth revision uploaded on June 15, 2026 [1]. The submission history shows the manuscript size shrinking from 27,225 KB in the first version to 4,031 KB in the final revision [1]. The paper appears on arXiv under the Computer Vision and Pattern Recognition category [1]. Broader context underscores the relevance of such detection tools. The rapid improvement of image and video synthesis models has made it easier to produce convincing fake media, a challenge that has drawn attention from both academic and industry labs [2]. The BusterX++ paper is indexed on Hugging Face’s paper pages, a platform that links research artifacts such as models, datasets, and interactive demos to academic manuscripts [4]. Hugging Face and arXiv have also collaborated to embed community-built demos directly alongside paper abstracts, allowing readers to test models without writing code [5]. The release of BusterX++ comes amid a period of intense activity in open-weight model development. Companies such as DeepSeek and Alibaba’s Qwen have distributed large language models under permissive licenses, fueling rapid experimentation across the research community [7][9]. DeepSeek, a Chinese AI firm founded in 2023, gained attention for training competitive models at reported costs far below those of larger rivals, using techniques such as mixture-of-experts layers [7]. Alibaba Cloud’s Qwen family, meanwhile, offers models under Apache 2.0 and other licenses, further expanding the pool of openly available foundation models [9].