Two days later, this question was systematically addressed at both the technical and classroom levels. On November 11th, at the 9th China Education Technology Innovation Development and Application Achievements Conference held in Wuxi, Jiangsu, Dr. Deng Shujun, Chief Scientist of Galaxy Smart Learning, gave a keynote speech titled "Let Every Child Learn, Learn Well, and Love Learning: Exploration and Practice of AI-Driven Scalable Personalized Education," detailing the technical architecture and the logic of implementing the company's product "Xiaolü AiXue" in school settings.
Galaxy Smart Learning was founded by Chen Lin in December 2024. The core team of Galaxy Smart Learning has previously worked in ByteDance's education business and Doubao's large model-related teams, and has continuously focused on "student self-directed learning" and intelligent education systems, accumulating experience in the development of self-directed learning products, curriculum research and development, and training in long-term memory, problem-solving, and explanation capabilities of large models. Among them, Dr. Deng Shujun, Chief Scientist, previously served as the head of Qinghua and Peking University Online School and the algorithm team at Dali Education, and participated in the development of Doubao AiXue and HeMa AiXue.
Although general-purpose large models possess strong common-sense capabilities, they may still produce hallucinations when dealing with rigorous subject logic. To meet educational scenario requirements, Galaxy Smart Learning adopts a "dual-core" technical architecture combining the Doubao large model with its self-developed intelligent learning engine, deeply integrating general intelligence capabilities with educational-specific algorithms.
In this architecture, the general-purpose large model primarily handles tasks such as language understanding, content generation, and basic reasoning. The team conducts specialized tuning for educational scenarios, including constructing problem-solving thought chains, aligning with Socratic-style questioning, and developing Agent capabilities based on long-term memory. By breaking down key steps in the problem-solving process, the model can pause at appropriate stages and pose thought-provoking questions, guiding students to gradually understand knowledge and identify cognitive blind spots.
Self-developed intelligent learning engine: based on learning science and cognitive measurement analysis, it integrates technologies such as Deep Knowledge Tracing (DKT), Item Response Theory (IRT), Cognitive Diagnostic Model (CDM), and Knowledge Space Theory (KST), analyzing students' responses, questions, and correction records to determine what knowledge they have already mastered, identify current weaknesses, and recommend suitable next steps for learning, providing a basis for learning situation diagnosis, ability assessment, learning difficulty matching, and subsequent task planning.
The two components complement each other: the general-purpose large model is responsible for knowledge explanation, prompting further questions, and process guidance, while the self-developed intelligent learning engine focuses on learning situation diagnosis and learning path determination. As Dr. Deng Shujun summarized in his speech, the former should "guide effectively," and the latter should "calculate accurately." Building on this, Galaxy Smart Learning further extends Agent capabilities to student learning planning and teacher instructional support.
The student-side planning agent relies on the large model's long-term memory, combining it with the student's long-term learning history to recommend the next learning task and explain the rationale. The teacher-side assistant agent extracts specific effort and progress clues from student learning records, providing references for teachers to offer personalized guidance.
The AI learning closed-loop further connects students' learning behaviors, teachers' instructional guidance, and school feedback. Students' learning performance feeds into the next round of learning situation analysis and task scheduling, allowing teachers to adjust their teaching strategies based on feedback, ensuring that AI capabilities continuously support real learning processes.
The core of this technical architecture lies in integrating subject knowledge, cognitive measurement, learning situation analysis, and large model interaction guidance, and continuously optimizing subsequent support through ongoing learning feedback.
From answering questions to guiding thinking, AI needs to go through a complete learning process.
For educational AI, technical capabilities ultimately need to be transformed into specific product interactions, implemented in understanding students' learning processes, and continuously validated in learning outcomes.
Currently, many AI education products still focus on Q&A, where students input questions and AI generates explanations, often ending the learning process at obtaining answers. However, true learning also includes understanding the problem, organizing thoughts, practicing applications, and verifying mastery. These steps collectively determine whether knowledge can truly be converted into students' own abilities.
In the design of Xiaoluaixue, the team emphasizes "letting technology go behind the scenes and letting students take the lead," and has designed specific functions around the complete learning process: AI provides support at appropriate times, allowing students to personally experience selection, attempts, errors, adjustments, and verification. In the process of mastering knowledge, students gradually learn to recognize and manage their own learning. The system supports students in asking multiple rounds of questions and follow-up inquiries around the learning content, connecting knowledge understanding, practice application, and outcome verification.
Dr. Deng Shujun gave an example in his sharing: when the height of a cylinder remains unchanged and the radius of the base is doubled, the volume should be four times the original, but some students might mistakenly think it's only doubled. To address such conceptual misunderstandings, Xiaoluaixue introduces dynamic concept experiments and Socratic prompting: students can intuitively observe the impact of radius changes on the cylinder's volume through a slider, while AI asks layered questions around key steps such as the volume formula and changes in base area, guiding students to derive conclusions independently.
"These prompts serve as a 'ladder' for students to understand knowledge, helping them find their first step and think independently at critical moments to answer step by step," said Dr. Deng. As understanding deepens, the system will appropriately withdraw support, allowing students to independently answer, explain their reasoning, and engage in variant practice to test whether they can answer correctly, explain clearly, and transfer their learning to new problems." The team hopes that by completing the process of overcoming difficult problems, students will gain the epiphany of "Oh, now I get it!" and the sense of accomplishment of "I did it!"
In cultivating students' interest and confidence in learning, the team combines deep knowledge tracking with the Zone of Proximal Development theory. By using models to predict the probability of students answering the next question correctly, the system matches learning tasks with moderate challenges. Dr. Deng Shujun introduced in his speech that the system selects questions with a predicted correct answer rate of about 70% to 80%, enabling students to achieve progress through effort.
" A small breakthrough achieved independently can become a 'spark' for children to rebuild their confidence," said Dr. Deng. The system records students' learning performance at different stages, showcasing specific improvements compared to the previous day or week, making minor efforts and breakthroughs visible. Learning teams encourage peers to motivate each other and tackle challenges together, while teaching assistants provide timely and targeted encouragement for teachers. "The team hopes that children will feel 'this step was taken by me,' gradually accumulating the confidence of 'I can learn,' thus being more willing to face the next unknown and take another step forward."
For school scenarios, single-point AI tools are insufficient to form a complete teaching solution. A student may quickly complete exercises with AI assistance, but if they haven't truly understood the knowledge, increased efficiency does not necessarily mean improved learning quality.
Xiaoluaixue positions itself as a human-AI co-education classroom solution for schools. By recording and analyzing students' learning processes, it provides personalized learning support and helps teachers understand students' knowledge mastery and individual differences. Teachers continue to take on roles such as designing instruction, explaining key concepts, guiding learning, and offering emotional support, adjusting teaching strategies based on AI feedback.
Teachers use the learning situation dashboard to understand students' learning progress, common questions, and correction status, conducting focused lectures, tiered guidance, and individual tutoring. At the school management level, the principal's end aggregates teaching process data, providing a basis for teaching observation, reflection, and management. Schools incorporate self-directed learning and practice, independent assessments, and targeted remedial learning into daily courses and schedules. Assessment results are then fed back into the next round of learning situation analysis and task assignment, forming a teaching closed-loop of "teacher-led, student-initiated, AI-supported."
The real challenge of AI education lies in how to scale personalized learning.
Personalized learning is not a new concept. For a long time, teaching according to individual capacity has been an educational goal. However, in traditional classrooms, teachers' time and energy are limited, making it difficult to continuously track each student's learning process, let alone provide timely and effective personalized guidance.
The possibilities brought by AI lie in reducing the cost of identifying student differences and providing personalized feedback through process data learning and intelligent analysis, making parts of the work that were previously highly dependent on teachers' personal experience more timely and sustainable. However, moving from individualized tutoring for a single student to large-scale application in schools still faces many challenges.
According to the introduction, in 2022, core members of Galaxy Smart Learning participated in the on-campus exploration of He Ma Ai Xue in Hefeng School in Fuyang, Anhui Province during the Da Li Education period. During long-term contact with frontline teachers and students, the team found that students' learning difficulties not only manifest in knowledge mastery but also include aspects such as learning methods and learning confidence. This frontline practical experience also influenced the team's later design of AI educational products, making them more focused on how students understand knowledge, master methods, and whether they can build confidence through continuous progress.
In the process of large-scale implementation, differences in teaching conditions, course progress, and student foundation exist among different schools. Whether teachers can integrate AI into daily teaching, whether the product can run stably, and whether learning feedback can truly be transformed into teaching actions all affect the actual application results.
Taking the Hongxing Senior High School in Nanyong, Guizhou Province as an example, the team gradually integrated functions such as recitation and spelling, question paper creation and distribution, online testing, AI grading, and learning situation feedback into the school's existing teaching types, such as morning reading, regular classes, time-limited practice, and evening study sessions. This helped teachers to timely grasp students' learning situations and conduct explanations for common problems. This continuous co-creation based on real classrooms also enabled the AI system to move beyond being a single learning tool and further enter the school's curriculum organization, teaching feedback, and daily management.
From technical solutions to real classrooms, educational AI is entering a deeper application stage. Currently, Xiaolu Ai Xue under Galaxy Smart Learning has been implemented in schools of different regions and types. According to Galaxy Smart Learning, the product has served and cooperated with more than 500 schools, covering 24 provincial administrative regions. In recent months, Galaxy Smart Learning has gradually disclosed case studies from schools such as Hefeng School in Fuyang, Anhui Province, Hui Longguan Yu Xin School affiliated with Capital Normal University, Hongxing Senior High School in Nanyong, Guizhou Province, Guo Ke Senior High School in Qingdao, Xincai Blue Sky Senior High School in Henan Province, and the Kuangao Experimental School in Siziwangqi, Inner Mongolia, showcasing the exploration of human-machine co-education models in different school contexts.
From knowledge acquisition to ability cultivation, the AI era requires a reevaluation of the value of learning.
Large models are changing the way knowledge is acquired. If the first stage of educational technology focused more on knowledge acquisition and learning efficiency, then the next change brought by AI might be a reevaluation of learning itself.
In the past, students spent a lot of time searching for materials, understanding concepts, and obtaining answers; now, AI can quickly provide information, explain knowledge, and even help complete complex tasks. As the threshold for knowledge acquisition continues to lower, the value of education also needs to shift from knowledge transmission to ability cultivation.
Therefore, the evaluation criteria for AI educational products also need to change: it cannot only focus on answer speed, number of exercises, and frequency of interaction, but also pay attention to whether students can truly understand knowledge, form effective learning methods, and gradually develop the ability to independently solve and actively explore problems, and build interest and confidence in learning through continuous progress.
This is the reason why Galaxy Smart Learning sets "learning, learning how to learn, and loving to learn" as its educational goals, and this goal is also related to the founder Chen Lin's decision to re-enter the education industry for entrepreneurship. In his article, Chen Lin reviewed his thoughts on education in the AI era. In his view, facing the technological changes brought by AI, education needs to help children develop the ability to deal with uncertainty and solve real-world problems. At the end of the article, he wrote: "This era has many people making AI stronger and stronger, while I want to help people become better and better..."
When AI enters the classroom, it ultimately points to human growth: enabling students, under the guidance of teachers and the support of technology, to gradually develop the ability to understand the world, make independent judgments, and solve problems, and to grow into individuals who can independently face the unknown world. And when technology no longer urgently proves how smart it is, but truly understands each child's cognitive blind spots, ignites their intrinsic motivation for self-exploration, and forms long-term collaboration with the school's educational practices, educational AI can truly realize its value.
This article is provided by Galaxy Smart Learning, and is authorized to be reprinted by Qubit. The views belong to the original author.
Copyright, no permission for any form of reprints or use, violators will be pursued.
Please fill in your name, gender, date of birth, contact information, and address.
The two components complement each other: the general-purpose large model is responsible for knowledge explanation, prompting further questions, and process guidance, while the self-developed intelligent learning engine focuses on learning situation diagnosis and learning path determination. As Dr. Deng Shujun summarized in his speech, the former should "guide effectively," and the latter should "calculate accurately." Building on this, Galaxy Smart Learning further extends Agent capabilities to student learning planning and teacher instructional support.
The student-side planning agent relies on the large model's long-term memory, combining it with the student's long-term learning history to recommend the next learning task and explain the rationale. The teacher-side assistant agent extracts specific effort and progress clues from student learning records, providing references for teachers to offer personalized guidance.
The AI learning closed-loop further connects students' learning behaviors, teachers' instructional guidance, and school feedback. Students' learning performance feeds into the next round of learning situation analysis and task scheduling, allowing teachers to adjust their teaching strategies based on feedback, ensuring that AI capabilities continuously support real learning processes.
The core of this technical architecture lies in integrating subject knowledge, cognitive measurement, learning situation analysis, and large model interaction guidance, and continuously optimizing subsequent support through ongoing learning feedback.
From answering questions to guiding thinking, AI needs to go through a complete learning process. For educational AI, technical capabilities ultimately need to be transformed into specific product interactions, implemented in understanding students' learning processes, and continuously validated in learning outcomes. Currently, many AI education products still focus on Q&A, where students input questions and AI generates explanations, often ending the learning process at obtaining answers. However, true learning also includes understanding the problem, organizing thoughts, practicing applications, and verifying mastery. These steps collectively determine whether knowledge can truly be converted into students' own abilities. In the design of Xiaoluaixue, the team emphasizes "letting technology go behind the scenes and letting students take the lead," and has designed specific functions around the complete learning process: AI provides support at appropriate times, allowing students to personally experience selection, attempts, errors, adjustments, and verification. In the process of mastering knowledge, students gradually learn to recognize and manage their own learning. The system supports students in asking multiple rounds of questions and follow-up inquiries around the learning content, connecting knowledge understanding, practice application, and outcome verification. Dr. Deng Shujun gave an example in his sharing: when the height of a cylinder remains unchanged and the radius of the base is doubled, the volume should be four times the original, but some students might mistakenly think it's only doubled. To address such conceptual misunderstandings, Xiaoluaixue introduces dynamic concept experiments and Socratic prompting: students can intuitively observe the impact of radius changes on the cylinder's volume through a slider, while AI asks layered questions around key steps such as the volume formula and changes in base area, guiding students to derive conclusions independently. "These prompts serve as a 'ladder' for students to understand knowledge, helping them find their first step and think independently at critical moments to answer step by step," said Dr. Deng. As understanding deepens, the system will appropriately withdraw support, allowing students to independently answer, explain their reasoning, and engage in variant practice to test whether they can answer correctly, explain clearly, and transfer their learning to new problems." The team hopes that by completing the process of overcoming difficult problems, students will gain the epiphany of "Oh, now I get it!" and the sense of accomplishment of "I did it!" In cultivating students' interest and confidence in learning, the team combines deep knowledge tracking with the Zone of Proximal Development theory. By using models to predict the probability of students answering the next question correctly, the system matches learning tasks with moderate challenges. Dr. Deng Shujun introduced in his speech that the system selects questions with a predicted correct answer rate of about 70% to 80%, enabling students to achieve progress through effort. " A small breakthrough achieved independently can become a 'spark' for children to rebuild their confidence," said Dr. Deng. The system records students' learning performance at different stages, showcasing specific improvements compared to the previous day or week, making minor efforts and breakthroughs visible. Learning teams encourage peers to motivate each other and tackle challenges together, while teaching assistants provide timely and targeted encouragement for teachers. "The team hopes that children will feel 'this step was taken by me,' gradually accumulating the confidence of 'I can learn,' thus being more willing to face the next unknown and take another step forward." For school scenarios, single-point AI tools are insufficient to form a complete teaching solution. A student may quickly complete exercises with AI assistance, but if they haven't truly understood the knowledge, increased efficiency does not necessarily mean improved learning quality. Xiaoluaixue positions itself as a human-AI co-education classroom solution for schools. By recording and analyzing students' learning processes, it provides personalized learning support and helps teachers understand students' knowledge mastery and individual differences. Teachers continue to take on roles such as designing instruction, explaining key concepts, guiding learning, and offering emotional support, adjusting teaching strategies based on AI feedback. Teachers use the learning situation dashboard to understand students' learning progress, common questions, and correction status, conducting focused lectures, tiered guidance, and individual tutoring. At the school management level, the principal's end aggregates teaching process data, providing a basis for teaching observation, reflection, and management. Schools incorporate self-directed learning and practice, independent assessments, and targeted remedial learning into daily courses and schedules. Assessment results are then fed back into the next round of learning situation analysis and task assignment, forming a teaching closed-loop of "teacher-led, student-initiated, AI-supported."The real challenge of AI education lies in how to scale personalized learning. Personalized learning is not a new concept. For a long time, teaching according to individual capacity has been an educational goal. However, in traditional classrooms, teachers' time and energy are limited, making it difficult to continuously track each student's learning process, let alone provide timely and effective personalized guidance.
The possibilities brought by AI lie in reducing the cost of identifying student differences and providing personalized feedback through process data learning and intelligent analysis, making parts of the work that were previously highly dependent on teachers' personal experience more timely and sustainable. However, moving from individualized tutoring for a single student to large-scale application in schools still faces many challenges.
According to the introduction, in 2022, core members of Galaxy Smart Learning participated in the on-campus exploration of He Ma Ai Xue in Hefeng School in Fuyang, Anhui Province during the Da Li Education period. During long-term contact with frontline teachers and students, the team found that students' learning difficulties not only manifest in knowledge mastery but also include aspects such as learning methods and learning confidence. This frontline practical experience also influenced the team's later design of AI educational products, making them more focused on how students understand knowledge, master methods, and whether they can build confidence through continuous progress.
In the process of large-scale implementation, differences in teaching conditions, course progress, and student foundation exist among different schools. Whether teachers can integrate AI into daily teaching, whether the product can run stably, and whether learning feedback can truly be transformed into teaching actions all affect the actual application results.
Taking the Hongxing Senior High School in Nanyong, Guizhou Province as an example, the team gradually integrated functions such as recitation and spelling, question paper creation and distribution, online testing, AI grading, and learning situation feedback into the school's existing teaching types, such as morning reading, regular classes, time-limited practice, and evening study sessions. This helped teachers to timely grasp students' learning situations and conduct explanations for common problems. This continuous co-creation based on real classrooms also enabled the AI system to move beyond being a single learning tool and further enter the school's curriculum organization, teaching feedback, and daily management.
From technical solutions to real classrooms, educational AI is entering a deeper application stage. Currently, Xiaolu Ai Xue under Galaxy Smart Learning has been implemented in schools of different regions and types. According to Galaxy Smart Learning, the product has served and cooperated with more than 500 schools, covering 24 provincial administrative regions. In recent months, Galaxy Smart Learning has gradually disclosed case studies from schools such as Hefeng School in Fuyang, Anhui Province, Hui Longguan Yu Xin School affiliated with Capital Normal University, Hongxing Senior High School in Nanyong, Guizhou Province, Guo Ke Senior High School in Qingdao, Xincai Blue Sky Senior High School in Henan Province, and the Kuangao Experimental School in Siziwangqi, Inner Mongolia, showcasing the exploration of human-machine co-education models in different school contexts.
From knowledge acquisition to ability cultivation, the AI era requires a reevaluation of the value of learning.
Large models are changing the way knowledge is acquired. If the first stage of educational technology focused more on knowledge acquisition and learning efficiency, then the next change brought by AI might be a reevaluation of learning itself.
In the past, students spent a lot of time searching for materials, understanding concepts, and obtaining answers; now, AI can quickly provide information, explain knowledge, and even help complete complex tasks. As the threshold for knowledge acquisition continues to lower, the value of education also needs to shift from knowledge transmission to ability cultivation.
Therefore, the evaluation criteria for AI educational products also need to change: it cannot only focus on answer speed, number of exercises, and frequency of interaction, but also pay attention to whether students can truly understand knowledge, form effective learning methods, and gradually develop the ability to independently solve and actively explore problems, and build interest and confidence in learning through continuous progress.
This is the reason why Galaxy Smart Learning sets "learning, learning how to learn, and loving to learn" as its educational goals, and this goal is also related to the founder Chen Lin's decision to re-enter the education industry for entrepreneurship. In his article, Chen Lin reviewed his thoughts on education in the AI era. In his view, facing the technological changes brought by AI, education needs to help children develop the ability to deal with uncertainty and solve real-world problems. At the end of the article, he wrote: "This era has many people making AI stronger and stronger, while I want to help people become better and better..."
When AI enters the classroom, it ultimately points to human growth: enabling students, under the guidance of teachers and the support of technology, to gradually develop the ability to understand the world, make independent judgments, and solve problems, and to grow into individuals who can independently face the unknown world. And when technology no longer urgently proves how smart it is, but truly understands each child's cognitive blind spots, ignites their intrinsic motivation for self-exploration, and forms long-term collaboration with the school's educational practices, educational AI can truly realize its value.
This article is provided by Galaxy Smart Learning, and is authorized to be reprinted by Qubit. The views belong to the original author.
