PerceptionDLM: Parallel Region Perception with Multimodal Diffusion Language Models

A research team has introduced PerceptionDLM, a multimodal diffusion language model designed to perform region perception tasks in parallel rather than sequentially, according to a paper submitted to arXiv on June 17, 2026 [1]. The model is built on PerceptionDLM-Base, a foundational system that the authors describe as achieving state-of-the-art performance among open-source diffusion multimodal large language models [1]. Most existing MLLMs rely on autoregressive generation, which processes regions one after another and limits efficiency for tasks requiring captions across multiple image areas [1]. PerceptionDLM instead uses efficient prompting and structured attention masking to enable simultaneous perception of multiple masked regions, generating descriptions in parallel at both the sequence and token levels [1]. The researchers report that this approach yields substantial speed improvements for multi-region perception tasks while maintaining competitive performance in region captioning [1]. To measure the parallelism property of visual perception, the team constructed a new benchmark called ParaDLC-Bench by scaling the existing DLC-Bench to include multiple region masks per image [1]. The benchmark allows joint evaluation of caption quality and inference efficiency [1]. The authors state they are the first to achieve parallel region captioning and perception by leveraging the advantages of diffusion language models [1]. Code, models, and datasets have been released alongside the paper [1]. arXiv, where the paper was posted, is an open-access repository of electronic preprints that are moderated but not peer-reviewed [6]. It was founded on August 14, 1991, and by the end of 2021 had surpassed two million articles [6]. As of November 2024, the submission rate stood at roughly 24,000 articles per month [6]. The repository serves fields including computer science, mathematics, and physics, and in some disciplines nearly all scientific papers are self-archived there before journal publication [6]. The paper appears under the Computer Vision and Pattern Recognition category and is accompanied by experimental browser features such as Bibliographic Explorer and CORE Recommender, which are part of the arXivLabs framework for community-contributed tools [4][5]. arXivLabs sets guidelines for collaborations between arXiv and third parties, requiring partners to share the repository's values of openness, community, excellence, and user data privacy [4].