Zhao, Cao, Lin & Koedinger (2026) — CMU, UNC, U Hong Kong. Accepted to LAK 2026. Online crowdsourcing experiment.
Core Finding
AI multimodal feedback matches educator feedback for learning while significantly outperforming it on student perceptions.
The authors built a real-time AI-facilitated multimodal feedback system integrating structured text, relevant slide references, and streaming AI audio narration. In a crowdsourced experiment, they compared it against fixed "business-as-usual" educator feedback.
The System
The multimodal feedback system combines three channels:
1. Structured textual explanations — targeted, specific feedback on student answers 2. Dynamic slide references — retrieval of the most relevant lecture slide for context 3. Streaming AI audio narration — spoken explanation to complement text
This multimodal approach aims to provide timely, targeted feedback at scale — something that remains a major challenge in education.
Results
Learning Effectiveness
- Equivalent learning gains between AI multimodal feedback and educator feedback
- No significant difference in post-test performance
Student Perceptions (AI significantly better)
| Dimension | AI > Educator? |
|---|---|
| Clarity | ✓ |
| Specificity | ✓ |
| Conciseness | ✓ |
| Motivation | ✓ |
| Satisfaction | ✓ |
| Reduced cognitive load | ✓ |
| Correctness | — |
| Trust | — |
| Acceptance | — |
AI feedback matched educator feedback on correctness, trust, and acceptance — but outperformed on every experiential dimension.
Behavioral Engagement Patterns
- Multiple-choice questions: Educator feedback encouraged more total submissions (students kept trying)
- Open-ended questions: AI feedback lowered revision barriers — targeted suggestions promoted iterative improvement
Significance
This is a strong result for AI feedback systems:
- Equivalence on learning is the bar most systems fail to clear — AI matched human educators
- Superior student experience across 6 dimensions suggests AI can surpass humans on consistency, specificity, and clarity
- Scalability: The system can provide real-time, context-aware support without instructor availability constraints
- Question-type effects: The engagement pattern differences suggest adaptive strategies — AI may be better for open-ended work while human-like interaction helps for multiple-choice
Connection to the Wiki
This paper provides direct empirical evidence for several wiki threads:
- sequenced-ai-feedback-learning: Cao et al. found sequenced feedback harmed learning — Zhao et al. show that direct AI feedback (not sequenced) can match educators. Together, these papers clarify how to structure AI feedback effectively.
- multimodal-ai-tutoring: Direct companion — multimodal errors and dialogue corrections in tutoring, while Zhao et al. focus on multimodal feedback delivery
- multimodal-learning-genai: The MMLD-AI unifying model for multimodal teaching/learning — Zhao et al.'s system is a concrete implementation
- formative-assessment: AI-generated multimodal feedback as a scalable formative assessment tool
- ai-peer-feedback-systems: Multi-LLM collaborative feedback — Zhao et al. add the multimodal dimension
- ai-tutor-effectiveness-review: Empirical evidence for AI feedback effectiveness, contributing to the mixed evidence base
- automated-grading: Feedback quality is downstream of grading accuracy — multimodal delivery may improve feedback without changing the grade
- engagement-assessment-video: Multimodal engagement tracking connects to multimodal feedback delivery
Methodological Notes
- Online crowdsourcing experiment — participants recruited via a platform, not classroom students
- Compared against fixed educator feedback (not live, not adaptive) — the AI system's adaptivity may partly explain its perceptual advantage
- Single-session design — long-term effects unknown
Open Questions
- Would results hold in real classroom settings with live educator feedback rather than fixed, pre-written feedback?
- Does the advantage persist over multiple sessions, or is there a novelty effect?
- How does each modality (text vs. slides vs. audio) contribute to the overall effect?
- Can multimodal AI feedback reduce the negative behavioral pathway identified in sequenced feedback studies (fewer resubmissions)?
Related Pages
- affective-text-wearable-student-health — Emotional register signals in student language predict learning-relevant outcomes
- feedback-loop — AI feedback loops and their impact on learning
- collaborative-ai-tutoring — Proactive AI tutoring with real-time support
- ai-learning-transfer — Do gains from AI feedback persist when the system is removed?
- human-in-the-loop-ai — The role of human judgment alongside AI feedback delivery
- llm-student-modeling-memory — Student modeling for personalized multimodal feedback
Citation
APA: Zhao, C. Q., Cao, J., Lin, J., & Koedinger, K. R. (2026). LLM-based Multimodal Feedback Produces Equivalent Learning and Better Student Perceptions than Educator Feedback. arXiv:2601.15280. Accepted to LAK 2026.