Recently, multiple companies in the education sector have announced the integration of their products with the DeepSeek large model. In the field of educational technology, the application of AI large models is transitioning from concept to implementation.

   Do the technical characteristics of DeepSeek naturally fit educational scenarios? What are the opportunities and challenges as the education industry accelerates the application of AI? On February 11, Xueersi Education Group CTO Tian Mi was interviewed by Cover News, where he revealed that Xueersi has also reached a deep cooperation with the DeepSeek large model. Xueersi plans to develop vertical large models more suited for educational scenarios based on DeepSeek's technical capabilities and intends to launch smart educational hardware products equipped with the "Deep Thinking Mode" within February.

   The advantages that DeepSeek has gained through its in-depth research on data and algorithms have greatly boosted the confidence of educational companies in the development of large model technologies. Tian Mi is optimistic about the industry's development against this background: He believes that as large model technologies like DeepSeek continue to mature, the education industry will face unprecedented opportunities.

Intelligent Tutoring Breaks Traditional Learning Tutoring Models

From Problem-Solving to Cognitive Inspiration

   "Intelligent tutoring is not just about providing answers; it helps children understand the logic of problem-solving and develop cognitive abilities through AI’s deep thinking process," Tian stressed when asked why they chose to integrate the DeepSeek large model. He emphasized its breakthroughs in fields such as logical reasoning, code generation, and multi-turn dialogue. These capabilities are highly compatible with the needs for problem-solving, teaching, and personalized Q&A in educational scenarios. Specifically, DeepSeek-R1’s mathematical reasoning skills are now comparable to the current level of the MathGPT model.

   Taking mathematics learning as an example, traditional online education tools often only provide standard answers and fixed problem-solving steps. After integrating DeepSeek, smart educational hardware products such as learning machines and practice machines will be able to showcase AI’s deep thinking process while answering questions. This can turn the standard problem-solving process into a perceivable cognitive path, allowing children to receive more comprehensive inspiration and guidance from AI’s thinking process. This "cognitive visualization" approach can help students better understand the logic behind the problems.

   Furthermore, DeepSeek’s optimization lowers the complexity of prompt engineering. Its "deep thinking" ability can automatically decompose user needs, combine online searches, and generate structured answers. This means that students do not need to precisely describe their questions for the AI to understand their intentions and provide targeted answers. This upgrade in interaction will greatly enhance learning efficiency and user experience. "We will also conduct secondary training on the DeepSeek foundation using Xueersi's proprietary educational data to create vertical large models better suited for educational scenarios," Tian emphasized.

The background of applying generative artificial intelligence to education  

Educational enterprises continue to break through barriers and overcome obstacles.

   Although AI large models show great potential in educational scenarios, the "hallucination output" problem (i.e., generating inaccurate or off-syllabus content) remains a significant industry challenge. Tian acknowledged that this is also a technical challenge Xueersi has focused on tackling during implementation.

   "Through optimizing training data and upgrading reasoning strategies, we ensure that the content output by the large models is both accurate and adheres to educational principles," Tian stated. For example, Xueersi developed a specialized LaTeX legality detection tool to transform formulas into correct text formats, ensuring the accuracy of mathematical reasoning. Additionally, the team uses algorithms similar to Q* to evaluate multiple candidate steps generated at each reasoning stage and select the optimal path extension, thereby enhancing the accuracy of the results.

   Besides the hallucination output issue, balancing "personalized recommendations" with the ethical conflict of "information cocoons" presents another challenge in applying generative AI to specific scenarios. Tian also introduced solutions for this, such as in the "MathGPT Anytime Ask" application, the AI does not directly provide answers but guides students through questioning and dialogue to gradually understand problem-solving methods. This approach not only avoids the formation of information cocoons but also cultivates students' independent thinking skills.

   "Additionally, as educational products targeting minors, it is crucial to establish a strict content filtering system and alignment mechanism to ensure that the AI outputs content suitable for minors' physical and cognitive development capabilities," he pointed out.

   AI large models open a new window for traditional education. But it still requires time and patience to be implemented. As AI large models gradually become standard in educational hardware, the industry’s competitive dimensions are shifting from a dual focus on content and hardware to a deep integration of AI and scenarios. Tian believes that future educational technology competition will increasingly focus on deep cultivation of data and scenarios. He commented that currently, the high starting point of the DeepSeek base model narrows the technical gap within the industry and generally reduces costs for related applications. However, vertical education still requires deep data and scenario cultivation. This is also why new functions based on DeepSeek's technology integration will be launched successively in February, as internal repeated testing and feedback from real users are needed.

   Although AI opens a door to assisting education and brings more possibilities to educational tutoring, Tian stressed that AI’s value lies in assisting teachers and will not completely replace human teachers. "The care, encouragement, and motivation provided by teachers are irreplaceable."

   According to iMedia Research data, the market size of Chinese education smart hardware reached 80.7 billion yuan in 2023, a 29.53% year-on-year growth. It is projected to exceed 100 billion yuan by 2025. The "2024 Artificial Intelligence + Education Industry Development Report" released by iMedia also predicts that the AI contribution rate to the online education market is expected to increase from approximately 7% in 2023 to around 16% by 2027. Tian also envisions the long-term value of AI empowering education: AI large models are opening a new window for traditional education, but their implementation requires time and patience.

   "As an industry practitioner, I deeply feel the changes brought by large models. In fact, traditional AI technology has already brought many changes to the education industry, reflected in various stages of preparation, teaching, practice, evaluation, and management. However, the tolerance rate for large models in the education industry is high, and parents provide considerable feedback on issues including content being off-syllabus and multiple solutions to problems," he added. Despite the challenges, Tian remains confident in the AI-empowered education industry: "I believe that as large model technologies like DeepSeek continue to mature, the education industry will face unprecedented opportunities."

Original article from: Innovator Cover News https://www.thecover.cn/newsinterview/7925647.html