Synthesis: Teacher-centered versus student-centered LLM agents in physics learning — Wang et al. (2026) build two prompt-engineered agents on the Coze platform (both powered by DeepSeek R1 at temperature 0.3) and compare them in a two-question conceptual physics task. The teacher-centered agent answers as an authoritative expert working from a bounded textbook knowledge source; the student-centered agent is configured as an empathic teacher with pedagogical content knowledge, diagnosing Misconceptions about AI, explaining from the student's perspective, and closing each round by checking comprehension. The student-centered condition produced higher post-test scores (9.67 vs. 7.93), lower extraneous cognitive load (8.33 vs. 10.76) and higher germane load (14.27 vs. 12.62), stronger flow experience (43.73 vs. 37.97), and higher empathy perception (21.27 vs. 18.24) — evidence that agent role design, not just model capability, shapes learning and affect in AI-supported science learning.

Overview

Large language models answer many educational questions fluently but still struggle to produce accurate, coherent explanations for physics problems, which demand conceptual reasoning and multistep analysis. Prompt engineering is the practical lever: role, skill, and constraint instructions steer what the model does. Prior work has mainly tested expert agents, leaving other role designs — notably a student-centered agent — empirically unexamined. This study compares a teacher-centered and a student-centered agent on four outcomes: learning performance, cognitive load, flow experience, and empathy perception.

The two agents

Both agents were built on Coze with chain-of-thought guidance and the same foundation model, so the comparison isolates the prompt design.

• Teacher-centered agent — an experienced physics teacher with deep subject knowledge. Its skills are analyzing the question context, providing solutions, and offering supplementary examples. Its constraints: responses must be accurate and credible, grounded in a bounded knowledge source of Ministry of Education-approved electronic textbooks (with explicit acknowledgment when a question falls outside that boundary), explained from an instructor's perspective, and concise and structured with no extraneous content. The authors note these constraints were intended to reduce hallucination without guaranteeing accuracy.
• Student-centered agent — a physics teacher with empathic capability, configured around pedagogical content knowledge and specifically knowledge of students' understanding: the ability to diagnose comprehension levels and common misconceptions. Its scripted moves are "You have this question because…" (analyzing the cause of the misconception), "This question involves…" (naming the relevant concept), and "This phenomenon is similar to…, also because…" (transferring to a new real-world situation). Explanations come from the student's perspective, interaction is prioritized, and the style is deliberately friendly.

Method

• Participants: 63 high school graduates from a southeastern Chinese province, all recent National College Entrance Examination takers ranked within the top 20.4% provincially (20 male, 43 female); none had prior formal instruction on AI agents or LLMs.
• Materials: two conceptual multiple-choice items — which of two unequal candles in a closed glass box extinguishes first, and how air-conditioner airflow direction affects temperature uniformity in a sealed room. A pilot with 113 comparable students found only 32.86% accuracy on the candle item, confirming it targets a genuine reasoning gap.
• Design: assignment by the parity of ID-number last digit (32 teacher-centered, 31 student-centered); a 10-minute independent pretest, roughly 20 minutes of dialogue with the assigned agent, then post-test and questionnaires. Four students who did not provide personal information in the post-test were excluded, leaving 59 analyzed.
• Instruments: the two questions as a knowledge test (5 points per correct answer), plus scales for cognitive load (three dimensions, Cronbach's α = 0.857), flow experience (enjoyment, engagement, control), and empathy perception. Non-normal distributions (Kolmogorov–Smirnov p < 0.01) led to Mann–Whitney U and Wilcoxon signed-rank tests for the knowledge test.

Results

• Learning performance. Pretest scores did not differ (student-centered 5.83, SD 3.24; teacher-centered 5.86, SD 3.01; U = 435, Z = 0.000, p = 1.000). Both conditions improved significantly — student-centered Z = −4.07, p < 0.001, r = 0.76; teacher-centered Z = −3.21, p < 0.01, r = 0.59 — but the student-centered condition finished higher at post-test (9.67, SD 1.27 vs. 7.93, SD 2.84; U = 298, Z = 2.89, p < 0.01, r = 0.38, a moderate advantage).
• Cognitive load. Intrinsic load did not differ (t = −0.57, p = 0.58, d = −0.15). The student-centered condition reported lower extraneous load (8.33 vs. 10.76; t = −2.84, p = 0.01, d = −0.74) and higher germane load (14.27 vs. 12.62; t = 2.16, p = 0.04, d = 0.56) — less wasted effort, more effortful schema-building.
• Flow experience. Total flow was higher in the student-centered condition (43.73 vs. 37.97; t = 3.54, p = 0.003, d = 0.92), driven by enjoyment (15.17 vs. 12.66; d = 1.00), engagement (13.13 vs. 11.76; d = 0.60), and control (15.43 vs. 13.55; d = 0.75).
• Empathy perception. Students rated the student-centered agent as substantially more empathic (21.27, SD 2.56 vs. 18.24, SD 3.70; U = 170, Z = 4.05, p < 0.01, r = 0.53).

Why the role design mattered

The authors attribute the student-centered advantage to interaction structure rather than content coverage. That agent sequenced its help — first diagnosing the cause of the misconception, then explaining the essential concept, then offering an analogous example — consistent with stepwise guidance proposals such as review–guidance–inspiration–correction–summary. It also closed each round by asking whether the student had understood the point, which the authors read as building a sense of participation (the user taking an active role rather than receiving information passively), keeping students immersed in the dialogue and supporting comprehension, Motivation, and Student Engagement. The physics education literature is cited as already arguing that LLM implementation in physics instruction should prioritize a student-centered approach.

What this means for practice

• Instructors. Specify the agent's role, skills and constraints as carefully as its content accuracy: with model, platform and temperature held constant, only the prompt design differed, and the student-centered condition finished higher at post-test (9.67 vs. 7.93, r = 0.38). The accuracy-optimized teacher-centered agent — bounded to a textbook knowledge source and built for concise, accurate answers — still finished lower on every cognitive and affective measure.
• Instructors. Script the sequence rather than only the explanation: diagnose the cause of the misconception first, name the relevant concept, then transfer to an analogous real-world case, and close each round by asking whether the student understood.
• Instructors. Hold prompt design accountable for affect: students rated the student-centered agent as more empathic (21.27 vs. 18.24) purely on the strength of scripted moves — perspective-taking openings, misconception diagnosis, comprehension checks.
• Learners. Treat dialogue with the agent as work rather than delivery — the student-centered condition reported lower extraneous load (8.33 vs. 10.76) and higher germane load (14.27 vs. 12.62), that is, less wasted effort and more schema-building.
• Researchers. Use the extraneous/germane split as an inexpensive instrument for comparing tutoring designs, since a single total load score would not have separated the two agents.

Limitations

• The sample was 63 high school graduates from a single southeastern Chinese province, all ranked within the top 20.4% of the National College Entrance Examination provincially; four were excluded for incomplete post-test information, leaving 59 analyzed.
• The intervention was a 10-minute pretest plus roughly 20 minutes of dialogue, and learning performance was measured by two conceptual multiple-choice physics questions (5 points each).
• Assignment used ID-number parity rather than random allocation, and the knowledge-test distributions were non-normal (Kolmogorov–Smirnov p < 0.01), requiring Mann–Whitney U and Wilcoxon signed-rank tests.
• Cognitive load, flow and empathy perception were self-reported on questionnaires, so the affective advantages describe students' perceptions of the agent rather than observed pedagogical behavior.

Connected Concepts

• Physics Education
• Pedagogical Agent
• Prompt Engineering
• Generative AI
• Large Language Models (LLMs)
• Intelligent Tutoring
• Affective Computing
• Student-AI Interaction
• Misconceptions about AI
• Motivation
• Student Engagement
• Teaching
• Hallucination Risk
• Personalized Learning
• Human AI Collaboration

Connected Articles

• Pragmatic users and skeptical nonusers: A qualitative typology of ChatGPT adoption in physics education — Typology of ChatGPT use in physics learning
• A Bottom-Up Taxonomy of Student Discourse with a Socratic AI Physics Tutor — Taxonomy of Socratic AI physics tutors
• Analyzing Undergraduate Problem-Solving in Physics Through Interaction With an AI Chatbot — Socratic physics chatbot
• It's Not the Tool, It's the Task: A Framework for Cognitively Activated AI Augmentation in Physics Instruction — AIRIS: cognitively activated AI augmentation in physics
• Trust-utility gap in introductory physics education: Students' adoption, domain-specific skepticism, and preferences for AI integration — The trust–utility gap in physics AI tools
• Probing AI-Generated Physics Solutions and Preparing Students to Critique Them — Probing AI-generated physics solutions
• A multi-agent AI classroom based on dual-process reasoning hazards: a pilot with prospective physics teachers — Multi-agent classroom simulation with physics teachers
• Adversarial Stress Testing of Role-Playing Language Agents using Multi-Agent Evaluation — Adversarial stress testing of role-playing agents
• Catching The Correct Answer Trap: Characterising AI Tutor Blind Spots When Analysing Student Reasoning — When hints help and when they encourage over-reliance
• How motivation and roles influence metacognitive engagement in student-GenAI interaction — Motivation and roles in metacognitive GenAI engagement

Citation

Wang, Y., Chen, X., Xiong, Y., Xu, S., Li, Q., & Zhou, S. (2026). Comparing teacher-centered and student-centered agents based on prompt engineering: Effects on learning performance, cognitive load, flow experience, and empathy perception in physics learning. Physical Review Physics Education Research, 22(2), 020131.