ReadingResearchRadarInvestment framework
Sign in / Sign up
← Back to reading
NewsMaterial3 min readDisconfirmAI Editorial

Deng Shujun, Chief Scientist at Galaxy Zhixue (China): Large Models Can Solve Problems, But Cannot Explain Why Students Fail

In a speech, Deng Shujun, Chief Scientist of Galaxy Zhixue, proposed that the watershed moment for educational AI lies not in the ability to generate answers, but in understanding the learning process. The company employs a dual-core architecture combining the Doubao large model with a proprietary intelligent learning engine; the former handles heuristic guidance, while the latter uses deep knowledge tracking and cognitive diagnostic models to identify students' weak points, addressing the respective challenges of effective guidance and accurate calculation.

2026-11-12 18:45 GMT+8

Deng Shujun's core judgment is that while large models can rapidly generate problem-solving steps, they may not be able to answer why a student does not understand. As general-purpose models become infrastructure, the true watershed for educational AI lies in its ability to deeply learn the learning process—understanding disciplinary logic, identifying individual cognitive differences, and translating technology into verifiable learning outcomes in real classrooms. This judgment points to a specific architectural challenge: general-purpose large models may still suffer from hallucinations when facing rigorous disciplinary logic, and relying solely on language interaction cannot stably track each student's mastery of knowledge.

Galaxy Zhixue's solution is a dual-core architecture. The general-purpose large model (Doubao) handles language understanding, content generation, and basic reasoning; the team has performed specialized fine-tuning for educational scenarios, including constructing chains of thought for problem-solving, aligning Socratic-style heuristic guidance, and developing Agent capabilities based on long-term memory. The model pauses at key steps in solving problems and poses heuristic questions, guiding students to gradually understand knowledge and discover blind spots in their thinking. The proprietary intelligent learning engine, based on learning science and cognitive measurement analysis, integrates Deep Knowledge Tracking (DKT), Item Response Theory (IRT), Cognitive Diagnostic Models (CDM), and Knowledge Space Theory (KST). By analyzing students' responses, questions, and correction records, it determines what knowledge has been mastered, where current weaknesses lie, and what content is suitable for the next step. Deng Shujun summarized the relationship between the two as: the former must guide well, and the latter must calculate accurately.

Sign in / Sign up
ReadingResearchRadarInvestment framework

The functional logic of this division of labor is as follows: learning guidance requires language models to dynamically adjust questioning strategies in dialogue, pausing at appropriate stages to pose heuristic questions; academic diagnosis requires statistical models to continuously track students' knowledge states and perform probabilistic inference, providing a basis for matching learning difficulty and planning subsequent tasks. The two complement each other and are connected into a closed loop through Agent capabilities: the output of academic diagnosis determines the arrangement of the next task, while the large model decides how to explain and follow up. On this basis, Galaxy Zhixue further extends Agent capabilities to student-side learning planning and teacher-side teaching support.

At the product level, this architecture is reflected in specific interaction designs. Taking the volume of a cylinder as an example, a student might mistakenly believe that if the base radius doubles, the volume only doubles. The system does not directly give the answer; instead, it allows students to observe the impact of radius changes on volume via sliders, while the AI asks layered questions around the volume formula and changes in base area, guiding students to derive conclusions independently. As understanding deepens, the system removes assistance and tests transferability through independent answering and variant exercises. In terms of difficulty matching, the team combines deep knowledge tracking with the Zone of Proximal Development theory, selecting questions with a predicted correct-answer probability of approximately 70% to 80%, allowing students to make progress through effort and gradually build confidence.

It should be noted that the above technical architecture and effect descriptions come from the company's speeches and disclosures; there is currently no independently verified data on learning outcomes. The figure of over 500 schools served cumulatively, covering 24 provincial-level administrative regions, is also self-reported by the company. While the design logic of the dual-core architecture—separating cognitive measurement from language guidance—is technically checkable, whether its actual effectiveness surpasses single-model solutions still needs to be answered through controlled experiments.

Sources

  • qbitai.com量子位↗
  • infoq.cn↗
Continue researching →