From November 4th to 9th, the EMNLP (Conference on Empirical Methods in Natural Language Processing), one of the three top international conferences in the field of natural language processing, was held in Suzhou. The artificial intelligence team of CVTE achieved a breakthrough: a research paper on "Model Pruning" technology, led by CVTE's AI team, was accepted for the main conference forum. With a total of 8,174 submissions this year and a main conference acceptance rate of only 22.16%, the selection of this paper marking a significant validation of CVTE's technical prowess in on-device AI algorithms by the international academic community .
EMNLP is one of the three top-tier international conferences in the field of natural language processing, alongside ACL and NAACL. The conference focuses on cutting-edge empirical research methods in natural language processing and attracts leading global scholars each year. Papers accepted at EMNLP represent the highest level of innovation in the field and carry significant academic authority and industrial impact.
Research Achievements Reach the Forefront of International Academia
Building on the observation that different modules of large models have varying impacts on model accuracy, CVTE's AI research team has proposed an adaptive pruning method. This approach reformulates the pruning task as a combinatorial optimization problem. Through rigorous mathematical derivation, the team successfully reduced the time complexity of the solution from exponential to polynomial time. Experimental validation demonstrates that the method not only achieves strong performance without fine-tuning but also surpasses current state-of-the-art results when fine-tuning is applied. The related technical paper has been accepted by the main conference of EMNLP, a top-tier international NLP conference, earning recognition from the global academic community for this methodological advancement.
Large-scale models are now being deployed across various end-user products, with industry efforts focused on enhancing both model accuracy and inference speed to improve the user experience across multiple dimensions. "Model pruning" stands out for its unique advantage of delivering speed improvements while maintaining precision.
The CVTE AI technology team shared insights into their development process: "During the development process, we observed that geometric shape recognition tasks incurred approximately 15 seconds of inference latency on edge devices. After a series of optimizations, we reduced this to around 3 seconds. However, we were still not satisfied and aimed to bring it below 2 seconds. This is why we initiated research into pruning methods, with the goal of pruning 50% of the parameters while maintaining accuracy, in order to reduce the inference delay."
It is reported that this method belongs to the lightweight technology for large models and will be applied to business scenarios related to large models, such as geometric shape recognition and voice assistants. While ensuring the recognition accuracy, it can achieve an inference speedup of over 1.5x, thereby enhancing the user experience.
From R&D to Practical Application: Promoting "AI+" Empowerment Across Various Industries
CVTE's profound technical expertise stems from sustained investment in research and development. In August this year, the R&D team of Seewo, CVTE's education technology brand, achieved notable success at the prestigious international speech event, Interspeech MLC-SLM (Multilingual Spoken Language Model) Challenge, securing third place in Track two and eighth place in Track one, demonstrating outstanding R&D strength and technical accumulation in the field of speech language models. Currently, the award-winning research results have been widely integrated into Seewo's full range of AI products and solutions, significantly enhancing the brand's language recognition capabilities. For instance, AI-powered teaching aids developed by educators can now more accurately recognize and respond to voice interactions, helping to create more immersive situational classrooms.
Currently, CVTE is promoting extensive AI applications across diverse scenarios: Leveraging its "Three Institutes and One Station" R&D system, CVTE has established closed-loop R&D capabilities in key robotics technologies such as motion control, autonomous navigation, and multi-modal perception. In the "AI + Enterprise Services" sector, several MAXHUB conference products have received Microsoft Teams Rooms certification. Within education, its AI education solutions enable functionalities like AI-assisted lesson preparation, AI-driven classroom interaction recognition, and automated post-class observation report generation.
Particularly in "AI + Education", Seewo has constructed a "1+N+N" AI technology system, continuously fostering deeper integration of AI technologies into educational scenarios. As of the end of June 2025, the Seewo Classroom Intelligent Feedback System had established 19 key application demonstration zones nationwide, covering over 3,000 schools and deployed in more than 7,000 classrooms, generating over 360,000 intelligent classroom feedback reports. Furthermore, the number of activated users for Seewo's AI Lesson Preparation tool has surpassed 600,000.
*1.5 times refers to the R&D data.
*Data sourced from CVTE's 2025 Interim Report.
return
