Recently, a research paper by Newcapec AI Division titled “GroundVTS: Visual Token Sampling in Multimodal Large Language Models for Video Temporal Grounding” has been accepted by CVPR 2026. This research focuses on fine-grained temporal understanding in video large language models and proposes a text-question-guided visual token sampling method, offering a new technical approach for video content understanding and key event localization.
What is CVPR?
CVPR (IEEE/CVF Conference on Computer Vision and Pattern Recognition) is a highly influential international conference in the field of computer vision and pattern recognition. Officially defined as the flagship annual conference in this field, it is also classified as a Category A conference in Artificial Intelligence by the China Computer Federation (CCF), enjoying high academic recognition and industry influence.
This acceptance of the paper by CVPR 2026 reflects Newcapec’s continuous investment and technical accumulation in multimodal intelligence, video content understanding, and other directions. It also demonstrates the strategic achievements of the team’s dual-driver approach of “cutting-edge technology research + industrial application deployment.”
Paper Focus: Video Large Language Models – Why Aren’t They Yet “Proficient at Watching”?
Although current multimodal large language models possess basic video understanding capabilities, their performance in tasks requiring “precise capture” – such as “accurately identifying the time segment in which an event described by a sentence occurs” – remains less than ideal. This type of task is commonly referred to in academia as Video Temporal Grounding.
The key issue is that most existing video large language models process video content using uniform sampling, allocating attention equally regardless of the importance of different segments. While simple, this approach may miss crucial clues when truly key actions occur only in a few moments. Moreover, when input contains a large amount of irrelevant footage, it can be easily distracted, affecting judgment accuracy. Enabling models to learn to “target their focus” on truly useful frames is precisely the starting point of this research.
Core Innovation: GroundVTS – Enabling Large Language Models to Focus on Key Segments Based on Queries
To address the above issues, Newcapec AI Division proposed the GroundVTS architecture, which no longer rigidly performs uniform sampling but dynamically screens video information based on user queries.
How does GroundVTS work?
The core of GroundVTS is the Visual Token Sampling (VTS) module. It first evaluates the relevance between different visual tokens in the video and the text question, then selectively retains high-value information, forming a non-uniform sampling approach better suited to temporal understanding requirements. Additionally, the paper designs a three-stage progressive optimization strategy to stably adapt this new sampling mechanism to existing video large language models.
The goal is not merely to compress input or reduce computation, but more importantly to teach the large language model where to focus and what to ignore when facing a specific question.
Experimental Results: Leading Performance Across Multiple Tasks, Balancing Efficiency and Accuracy
Overall, GroundVTS achieves leading performance on mainstream datasets such as Charades-STA, ActivityNet-Captions, and QVHighlights. On two key tasks – Moment Retrieval and Highlight Detection – it surpasses comparable baseline models (e.g., Qwen2.5VL-7B, InternVL3.5-8B) by over 10 percentage points, with some metrics improving by tens of points. Compared to existing representative methods, it achieves an mIoU improvement of up to 7.7 points and an mAP improvement of up to 12.0 points, fully validating its effectiveness and competitiveness in fine-grained video content understanding.
Beyond higher accuracy, GroundVTS maintains strong performance and stability even with lower visual token budgets. Using only half the visual token budget, GroundVTS still outperforms the uniform sampling baseline under full budget. Even under more aggressive compression settings, the advantages remain significant. This indicates that GroundVTS not only improves localization accuracy but also greatly enhances the efficiency of video information utilization. It sees more accurately while also seeing more efficiently.
Deployment Outlook: Empowering Smart Campuses and Industry Applications
Newcapec is always committed to deeply integrating cutting-edge AI technologies with real-world scenario needs. The fine-grained video understanding capability represented by GroundVTS has broad application prospects in smart campuses and more industry scenarios, making technology truly visible and usable:
Campus Safety Scenarios: Quickly locate specific abnormal events in surveillance video, improving event retrieval and emergency response efficiency.
Teaching Resource Retrieval: Help teachers and students precisely locate knowledge point explanation segments within massive instructional videos.
Practical Training Scenarios: In skills training, precisely locate operational error moments and extract key operation segments, aiding training process review and evaluation, making teaching more targeted and enhancing learning outcomes.
The acceptance of this research by CVPR 2026 represents the research accumulation of Newcapec AI Division in multimodal video understanding and further solidifies the technical foundation for the large-scale deployment of related capabilities in education and more industry scenarios.
From top-tier conference paper publications, to core technology breakthroughs, to exploration of scenario deployments across thousands of industries, Newcapec AI Division continues to deepen its exploration in key directions such as large language models, intelligent agents, and multimodal understanding and generation. It accelerates the deep coupling of cutting-edge technologies with industry application scenarios, enabling AI capabilities to better understand scenarios and effectively serve the digital transformation of education and industry intelligence.
Stay tuned for the latest progress from Newcapec AI Division and witness every step of technology empowering the future.
About Newcapec AI Division: As the core engine driving the company’s AI technology innovation, product development, and scenario deployment, Newcapec AI Division closely focuses on the core needs of smart campuses and industry digitalization, continuously promoting the integrated development of cutting-edge technologies with business scenarios. On one hand, it focuses on key technology directions such as multimodal large language models, natural language processing, computer vision, and intelligent agent collaboration to build a solid core technology foundation. On the other hand, targeting core scenarios such as campus management, services, and teaching, it drives product development and deployment of solutions including campus AI assistants, AI middleware, data agents, intelligent teaching equipment, and smart terminals, advancing AI capabilities from technology research toward usable, deployable, and sustainably operable product systems.
