Open Category Submission

254 – Beyond Answers: LLM-Powered Tutoring System in Physics Education for Deep Learning and Precise Understanding

The integration of artificial intelligence (AI) in education has shown significant promise, yet the effective personalization of learning, particularly in physics education, remains a challenge. Physics-STAR pioneers a structured LLM framework that bridges the AI personalization gap in physics education through cognitive diagnosis and adaptive pathways. Through structured intervention and cognitive diagnosis, Physics-STAR solves the pain point of ‘focusing on answers but not thinking’ in the existing education technology products, while offering scalable solutions for STEM and beyond.

The framework of Physics-STAR, drawing on the Situation-Task-Action-Result (STAR) approach, provides an adaptable structure for LLM input in three steps of personalized tutoring. Based on LLM, we propose the Physics-STAR framework, which consists of three steps:
1. Knowledge Explanation: This step involves explaining the concepts, basic formulas, and application scenarios of specific knowledge to students.
2. Error Analysis: Through the communication with the student, this step analyzes the reasons behind student’s mistakes on a certain questions.
3. Review Suggestion: Personalized review suggestions are provided based on the analysis of errors. Following this, related questions are presented to the students for testing. If they pass the test, it indicates that the knowledge point has been adequately understood; otherwise, the three-step process is repeated.
The structure of prompts in each step consists of four sections: a situation, a task, an action, and a result. Students are required to engage in question-and-answer sessions with the personalized tutoring system according to the prompts provided.

Physics-STAR leverages a structured STAR framework to transform generic LLMs into domain-specific cognitive tutors, employing iterative problem-solving grounded in real-world contexts to bridge AI's "personalization gap." This approach pioneers scalable STEM education solutions that prioritize conceptual mastery over rote answers, demonstrating transformative potential for various disciplinary in AI era.

At present, the contradiction between the standardized and synchronous teaching characteristics of the mainstream class teaching system and the personalized needs of contemporary education is becoming increasingly prominent. In traditional classrooms, teachers often need to deal with super-large classes of 30-50 students, and this collectivized teaching model leads to three levels of structural dilemmas: first, the cognitive basis of students presents a gradient distribution; Second, teachers can only use the fragmented time to carry out limited problem diagnosis; Third, personalized counseling resources are seriously scarce. This contradiction between scale and individuality makes the educational ideal of "teaching students according to their aptitude" stay at the theoretical level for a long time.

The intelligent education system based on large language models provides a breakthrough solution to solve this problem. Through core technologies such as knowledge graph construction, cognitive diagnosis modeling, and conversational reasoning engine, this kind of system shows three core advantages: first, it has the ability to provide full-time service, which can respond to students' questions in real time and accurately locate knowledge blind spots through conversational interaction; Secondly, relying on the deep neural network with tens of billions of parameters, the system can automatically generate stepwise problem-solving ideas, and help students establish a complete cognitive framework through chain-of-thought guidance. In addition, the multi-modal interaction capability allows it to flexibly switch between teaching strategies such as text explanation, graphic presentation, and case analogy according to the differences in learning styles. More importantly, the dynamic student portrait formed by the system through continuous learning can predict potential learning disabilities and intervene in advance.

When large language models are deeply integrated with brain science and educational psychology, the future of education will truly realize the historic leap from "standardized training" to "precise empowerment".

Based on the technical closed-loop of "knowledge modeling-cognitive analysis-strategy generation", the scheme deconstructs subject knowledge into a computable and traceable conceptual network by integrating the knowledge graph system in the field of education. At the same time, it is equipped with the ability to think chain reasoning, and dynamically constructs students' cognitive maps in the process of dialogue.

The core technology implementation of the solution includes three key levels: in the interaction layer, the system supports multiple interaction methods to capture the students' problem-solving process and thinking trajectory in real time; At the analysis layer, based on the improved cognitive diagnosis model, the system can accurately identify students' knowledge breakpoints and ability shortcomings, and generate adaptive learning paths with reference to the theory of the nearest development zone. At the executive level, the system dynamically adjusts teaching strategies based on real-time feedback from students, providing step-by-step scaffolding support and implementing Socratic interrogation training.

The theoretical framework of educational psychology is deeply integrated: the cognitive load theory is used to optimize the way of knowledge presentation, and the information processing efficiency is improved through segmented teaching and working memory reinforcement training. Affective computing technology is used to analyze the intonation and micro-expression features of speech, and when frustration is detected, the encouragement strategy is automatically switched or the teacher's manual intervention mechanism is activated.

In order to verify the effectiveness of the program, a strict pre- and post-test control design was adopted, the pre-test determined the baseline level of students through diagnostic evaluation, and the post-test established a multi-dimensional evaluation model from the three dimensions of knowledge mastery, thought transfer and learning motivation value, and the effect attribution was carried out by using item response theory and covariance analysis.

We select Stream 1, which is blank here.

This scheme focuses on solving the disadvantages of "one-size-fits-all" in traditional teaching by constructing a dynamic learning portrait and intelligent guidance system. According to the learning characteristics of physics, the R&D team has carried out the teaching adaptability transformation of the basic large language model: on the one hand, the analysis accuracy of the model on important and difficult knowledge is improved by injecting the middle school physics syllabus and typical cognitive misunderstanding datasets; On the other hand, the interactive mechanism of "progressive guidance" is innovatively designed, which transforms the original mode of direct output of answers into a heuristic dialogue process.

When students ask a question such as "how to understand Newton's third law?", the system simulates the guidance strategy of a senior teacher: first, through real-life examples (such as the interaction between the paddle and water during rowing) to build perceptual cognition, then use the force analysis diagram to clarify the relationship between action and reaction forces, and finally guide students to apply the law independently in different situations. In this process, the system will capture students' blind spots in understanding in real time, and dynamically adjust the level of detail of the explanation—adding analogies for those who have difficulty understanding, and extending the relevant knowledge of the history of physics for those who grasp it quickly. This hierarchical and progressive guidance method effectively cultivates students' scientific thinking chain of "observing phenomena, refining laws, and transferring applications".

This kind of "scaffolding" guidance not only reduces the burden of repetitive explanations, but also significantly improves students' independent learning ability and autonomy.

This solution builds an intelligent education ecosystem that is both scalable and adaptable. At the level of system scalability, the three-level collaborative architecture of "cloud-edge-end" is adopted: the core large language model and knowledge graph engine are deployed in the cloud, which is responsible for global resource scheduling and complex inference tasks; The regional education node is equipped with edge computing units to carry localized knowledge bases and copies of commonly used models. The terminal device retains a lightweight interactive interface and supports seamless cross-platform switching.

In view of the potential bottleneck problem, the system implements a multi-dimensional breakthrough strategy: in terms of computing resource optimization, the knowledge distillation technology is used to construct a "teacher-student" dual model system; Deploy a hardware acceleration solution to improve the inference speed. Establish a hierarchical response mechanism to distinguish between real-time interaction and in-depth analysis of task flows, ensuring that core functions always run smoothly. At the data management level, an incremental learning framework is introduced to support the iteration of the knowledge base without interrupting services, and realize the collaborative evolution of regionalized models, which not only protects data privacy but also improves update efficiency.

In order to ensure long-term user participation, the system builds a dynamic growth ecosystem: the adaptive engine continuously tracks the learning trajectory and intelligently adjusts the difficulty coefficient and knowledge density of the questions based on recent performance; The 3D knowledge graph visualization system transforms the abstract learning progress into a virtual space that can be explored interactively, and generates personalized learning reports on a regular basis. The community-based function module builds a virtual seminar room based on the similarity of knowledge defects to support collaborative problem solving, and sets up a multi-dimensional achievement system to stimulate continuous exploration.

This project systematically solves the problem of unequal educational opportunities through technological innovation and mechanism design, and strives to build an inclusive intelligent education ecology. In order to promote educational fairness, we have built an AI teaching template sharing platform, a dynamic balance algorithm to ensure that rural students receive 80% of the same tutoring resources in cities, and public welfare computing power quotas give priority to supporting weak schools. At the level of fairness, the knowledge recommendation threshold is set to prevent the gap from being exacerbated by advanced education, the regional compensation algorithm automatically tilts high-quality resources to poor areas, and the anonymous learning mode eliminates the interference of family background. In terms of inclusiveness, it supports input in 7 minority languages, the cultural adaptation module automatically replaces example question scenarios, and the cognitive recognition system matches multiple learning styles. The social impact assessment focuses on five dimensions: education equity (the ratio of urban to rural tutoring hours), learning effectiveness (the rate of progress of undergraduates), resource accessibility (the use of low-level equipment), social inclusion (the use rate of special groups), and sustainable impact (three-year academic tracking curve). Demand response captures emerging demands in tens of millions of dialogues in real time through a dynamic perception system, analyzes the new education policy with a policy semantic engine, and determines optimization priorities based on quarterly "pain point crowdfunding". Establish a two-way channel in the curriculum reform experimental area, support teachers to visually adjust AI teaching logic, and transform high-frequency requirements into prototypes within two weeks. The system promotes education equity from resource equality to equal development opportunities, optimizes human capital through technology empowerment, and ultimately promotes social inclusion.

Personal Information Collection Statement (PICS):
1. The personal data collected in this form will be used for activity-organizing, record keeping and reporting only. The collected personal data will be purged within 6 years after the event.
2. Please note that it is obligatory to provide the personal data required.
3. Your personal data collected will be kept by the LTTC and will not be transferred to outside parties.
4. You have the right to request access to and correction of information held by us about you. If you wish to access or correct your personal data, please contact our staff at lttc@eduhk.hk.
5. The University’s Privacy Policy Statement can be access at https://www.eduhk.hk/en/privacy-policy.

• I have read and agree to the competition rules and privacy policy.