Feedback Generation in Education using Large Language Models: A Survey of Recent Advances

ACL ARR 2026 January Submission6400 Authors

Keywords: Educational feedback generation, automated feedback, hint generation, socratic scaffolding, Decision-making / feedback policies, Verification / verifiers, Rubric-based feedback / rubric adherence, Outcome-grounded feedback, Student modeling / student state estimation, Human-in-the-loop

Abstract: Large language models (LLMs) have made it possible to generate formative educational feedback at scale, but naïve feedback generation often fails to meet educational requirements. A rapidly growing line of work therefore reframes feedback generation as a decision-making problem: systems think pedagogically before they speak, by making intermediate choices—what to target, what action to take, how much support to provide, and what evidence should justify the message—before realizing the feedback. This survey reviews recent research on LLM-based educational feedback generation that incorporates such deliberative structure. We organize existing systems by where decision-making lives: prompting and in-context planning, training-time alignment, inference-time candidate selection, and scripted pedagogical scaffolds. We highlight open challenges in signal and judgement reliability, construct validity, generalization, personalization and scaling. We conclude with recommendations for building more auditable, controllable, and pedagogically-grounded feedback systems.

Paper Type: Long

Research Area: NLP Applications

Research Area Keywords: educational applications, feedback generation, human-in-the-loop, human-AI interaction/cooperation, human-centered evaluation, prompting, LLM agents, reinforcement learning in agents

Contribution Types: Surveys

Languages Studied: English

Submission Number: 6400