模拟会犯真实错误的学生，帮助 AI 导师更快学习

A new system from Microsoft and the University of Illinois builds realistic replicas of individual students from limited data. These replicas provide rapid feedback when getting it from real students would be too expensive and slow, helping researchers improve AI tutors. AI tutors work best when they adapt to each student's strengths and weaknesses. But finding out which guidance works for each student takes time and money because it requires real learners. Training an AI tutor with a large, diverse group of students is "prohibitively expensive and time-consuming," the authors write in their paper . As a result, improvements to these tutors have lagged behind advances in AI models. The researchers propose using digital replicas of students to provide quick feedback in their place. Their system, called StudentSim, builds a separate replica for each student, even when very few records of that person's work are available. Existing approaches handle only one of two necessary skills, according to the researchers. Some models learn from real student data and reliably reproduce a student's behavior, but they can't use a tutor's explanations. Others are language models prompted to act as students. They readily follow the tutor's hints but fail to match the abilities of the student they're supposed to mimic. StudentSim turns both skills into measurable goals. It measures how closely a replica matches a student's answers, including typical mistakes, and how readily it revises an answer after the tutor helps. Tutor training needs both a realistic starting point and a simulated student that responds to instruction. The researchers' biggest obstacle is a lack of data. In the English writing dataset, the median student has written just three essays, and more than two-thirds have written five or fewer. Training a replica directly on so few examples fails, the researchers say, because the model overfits to those examples. StudentSim instead trains in two stages. First, a base model learns from the pooled data of all students in a subject. It learns common mistakes and how students revise their answers after a tutor's hint. The researchers then tailor that model to an individual student using the few records available for that person. Across all subjects, the system uses Alibaba's Qwen3-4B-Instruct language model as its base. The researchers tested the method on 60 students across chess, English as a foreign language, and math. They used public datasets containing records from real learners. StudentSim outperforms the larger GPT-5.4 language model in all three subjects when GPT-5.4 is prompted to act as a student. In chess, StudentSim correctly predicts a player's next move about twice as often and almost always follows corrective guidance. GPT-5.4 and specialized chess models fall behind. Each existing method has a different weakness, the researchers say. GPT-5.4 follows hints but doesn't reproduce a particular student's mistakes. Chess models match a player's behavior but can't understand verbal hints and ignore them. In one position, three real players chose three different moves. StudentSim reproduced each player's choice, while a chess model predicted the same most likely move for all three. GPT-5.4 got all three wrong. As another proof of concept, the researchers used a student replica to improve a chess tutor. Professional chess players evaluated three versions, one without this training, one trained with GPT-5.4 as the student, and one trained with StudentSim.