EduAlly: AI Learning & Feedback Platform

• EduAlly is an AI-assisted educational platform designed for writing-intensive tasks, implementing a cyclical feedback and revision loop with teacher oversight.
• It operationalizes the AI-Educational Development Loop by combining adaptive AI feedback with reflective, iterative student revisions guided by instructional input.
• Empirical studies show statistically significant performance improvements through iterative revisions, highlighting alignment among AI, teacher, and student assessments.

Searching arXiv for the cited papers to ground the article in current arXiv records. arXiv search: EduAlly / AI-EDL / educational AI alignment papers. EduAlly is an AI-assisted educational platform for writing-intensive, feedback-sensitive tasks, implemented as the practical instantiation of the AI-Educational Development Loop (AI-EDL), a theory-driven framework that combines human-in-the-loop AI with reflective, iterative learning (Yu et al., 1 Aug 2025). Its central design premise is that AI should provide formative, rapid, and scalable feedback without displacing instructor authority: students submit an initial response, receive adaptive AI feedback and a provisional grade, revise their work, and then receive final evaluation through teacher oversight. In this formulation, EduAlly is not a fully automated grader, but a pedagogically supervised environment intended to support transparency, self-regulated learning, iterative revision, and growth-oriented assessment (Yu et al., 1 Aug 2025).

1. Conceptual definition and pedagogical basis

EduAlly is defined in the literature as an AI-assisted platform built to support learning tasks in which feedback quality, revision, and instructor judgment are central. The platform is explicitly associated with the AI-Educational Development Loop, a cyclical model in which learners move through knowledge gap → learning → trial → assessment → reflection → loop continuation → goal achievement (Yu et al., 1 Aug 2025). This developmental sequence treats assessment as a formative intervention rather than a terminal act of scoring.

The framework is grounded in several classical theories. Socratic questioning/dialogue is used to frame learning as questioning, reflection, and critical examination. Aristotelian virtue ethics appears in the emphasis on habituation toward intellectual excellence and purpose-driven growth. Hull’s drive-reduction theory is used to characterize knowledge gaps as sources of cognitive tension that motivate action. Skinner’s operant conditioning underlies the role of feedback and reinforcement in shaping later performance. Zimmerman’s self-regulated learning/metacognition informs the requirement that learners interpret feedback, reflect, and revise (Yu et al., 1 Aug 2025).

This theoretical grounding distinguishes EduAlly from systems that use AI primarily for automation. A plausible implication is that EduAlly belongs to a class of educational AI systems designed to mediate learning processes rather than merely optimize throughput. That interpretation is consistent with later work on trustworthy educational AI, which argues that alignment in education requires not only embedding human values into AI systems but also equipping teachers, students, and institutions with the skills to interpret, critique, and guide these technologies (Shen, 25 Dec 2025).

2. AI-EDL as a cyclical learning model

The core contribution associated with EduAlly is the AI-Educational Development Loop itself. AI-EDL is described as a cyclical model of learning and revision in which the learner is repeatedly brought back into contact with the consequences of prior performance through feedback, reflection, and resubmission (Yu et al., 1 Aug 2025). The framework is explicitly developmental: evaluation is a feedback moment that triggers reflective modification.

The paper formalizes the instructional package prepared by the teacher as

$I = \{M, Q, K, R\}$

where $M$ denotes instructional materials, $Q$ questions, $K$ the answer key, and $R$ the rubric (Yu et al., 1 Aug 2025). This package is submitted to the AI system, which initializes the backend and later generates adaptive feedback $F$ and a provisional grade $GA$ from student answers.

The paper identifies three actors: Ț for Teacher, Ș for Student, and A for AI Agent, alongside D for the backend database, F for adaptive feedback, and G / GA for grades (Yu et al., 1 Aug 2025). It also provides an AI-planning-style formalization of the loop:

• Submit(Ț, I, Ą)
• Init(Ą, I, D)
• Render(Ą, M, Ș)
• Digest(Ș, M)
• Render(Ą, Q, Ș)
• Answer(Ș, Q, As)
• (F, GA) = g(As, K, R)
• Render(Ą, {F, GA}, Ș)
• Reflect(Ș, {M, Q, As, F, GA})
• Revise(Ș, As, A's)
• Forward(Ą, A's, Ț)
• G = gT(A's, K, R)
• Render(Ą, G, Ș) (Yu et al., 1 Aug 2025)

This formalization makes explicit that the AI system is responsible for rendering materials, evaluating initial responses, and generating formative feedback, whereas the teacher remains the final grading authority. The architecture therefore encodes a human-in-the-loop division of labor rather than end-to-end automated assessment.

3. System architecture, workflow, and design principles

EduAlly operationalizes AI-EDL as an adaptive learning workflow organized into sequential phases. The reported workflow comprises: input preparation; data analysis and backend setup; material rendering; student engagement; question rendering; answer submission; feedback generation; feedback rendering; feedback review and reflection; second attempt; attempt forwarding; teacher grading; grading submission; and final feedback rendering (Yu et al., 1 Aug 2025). In practice, the teacher prepares the instructional package, the AI agent generates formative feedback and a provisional grade after the initial submission, the student revises, and the teacher grades the revised response.

Several design principles are repeatedly emphasized. Transparency is treated as essential: students receive visible and interpretable feedback, and instructors can monitor the process (Yu et al., 1 Aug 2025). Self-regulated learning is built into the requirement that students review feedback, identify deficiencies, and submit a second attempt. Instructor oversight is not peripheral but constitutive: teachers define the rubric and answer key, review AI-mediated outputs, and retain final judgment. Iterative revision is the mechanism through which the developmental loop operates, and growth orientation is expressed in the use of constructive feedback rather than judgment alone (Yu et al., 1 Aug 2025).

This workflow places EduAlly in a broader family of pedagogically aligned AI systems. EduChat, for example, is described as an LLM-based chatbot for intelligent education that supports open question answering, essay assessment, Socratic teaching, and emotional support, with domain-specific pre-training on educational corpora and fine-tuning for educational functions (Dan et al., 2023). EduAlign similarly proposes a reinforcement-learning framework that optimizes LLM tutors along Helpfulness, Personalization, and Creativity (HPC), using an 8k educational interaction dataset, a reward model called HPC-RM, and GRPO-based fine-tuning on 2k prompts (Song et al., 27 Jul 2025). EduAlly differs in emphasis: rather than centering reward-model alignment or general conversational tutoring, it centers a structured revision loop with preserved teacher authority.

4. Empirical evaluation

The reported evaluation of EduAlly is a mixed-methods pilot study conducted at a mid-size comprehensive public university in the northeastern United States in special education courses taught by the second author (Yu et al., 1 Aug 2025). Participants were undergraduate and graduate students, and the study investigated three issues: alignment between AI-generated grades, teacher evaluations, and student self-assessments; performance change from Attempt I to Attempt II after AI feedback; and student perceptions of the feedback process (Yu et al., 1 Aug 2025).

The tasks were writing-intensive, feedback-sensitive responses to content-specific questions, graded with a three-level rubric:

• Satisfactory = 2
• Improvement Needed = 1
• Not Assessable = 0 (Yu et al., 1 Aug 2025)

The descriptive statistics reported in the paper are as follows. For Attempt I, the AI grade had $n = 918$, mean 1.40, SD 0.73, range 0 to 2, while the Teacher grade had $n = 962$, mean 1.56, SD 0.73, range 0 to 2. For Attempt II, the Teacher grade had $n = 963$, mean 1.77, SD 0.57, range 0 to 2. For holistic evaluations, Student self-evaluation had $M$0, mean 1.99, SD 0.12, range 1 to 2, and Teacher final grade had $M$1, mean 1.89, SD 0.37, range 0 to 2 (Yu et al., 1 Aug 2025).

The inferential results indicate statistically significant improvement after revision. The Wilcoxon signed-rank test for Attempt I: Teacher Grade - AI Grade yielded $M$2, $M$3, indicating that teachers gave significantly higher grades than the AI on Attempt I. The test for Teacher Grade: Attempt II - Attempt I yielded $M$4, $M$5, indicating significantly higher scores on resubmission. The comparison Teacher Grade Attempt II - AI Grade Attempt I yielded $M$6, $M$7, and the comparison Attempt II: Teacher Final Grade - Student Self-Eval yielded $M$8, $M$9, indicating no statistically significant difference between final teacher grades and student self-evaluations (Yu et al., 1 Aug 2025).

The paper also reports Spearman correlations between student level and assessment outcomes: AI Grade - Attempt I $Q$0, $Q$1, $Q$2; Teacher Grade - Attempt I $Q$3, $Q$4, $Q$5; Teacher Grade - Attempt II $Q$6, $Q$7, $Q$8; Student Self-Evaluation $Q$9, $K$0; and Teacher Final Grade $K$1, $K$2 (Yu et al., 1 Aug 2025). The reported interpretation is that graduate students tended to receive slightly higher AI and teacher grades on content-specific attempts, while self-evaluations and final holistic grades were not significantly associated with student level.

Alignment statistics further characterize the relation among AI, teacher, and student judgments. Where both AI and teacher grades were available for Attempt I, Same grades: 654 / 782 = 83.63% and Different grades: 128 / 782 = 16.37%. Comparing teacher grades for Attempt I vs Attempt II, the paper reports Attempt I > Attempt II: 0.31%, Attempt I < Attempt II: 15.18%, and Attempt I = Attempt II: 84.51%. Comparing AI grades for Attempt I vs teacher grades for Attempt II, it reports AI > Teacher Attempt II: 0.26%, AI < Teacher Attempt II: 26.18%, and AI = Teacher Attempt II: 73.56% (Yu et al., 1 Aug 2025). These figures are presented as evidence that AI feedback aligned reasonably well with instructor judgment while supporting later improvement.

5. Student experience, interpretability, and trustworthy use

The qualitative analysis is based on open coding of student self-reflections and reports generally positive but mixed perceptions of the system (Yu et al., 1 Aug 2025). Positive themes included balance and constructiveness, clarity and specificity, encouragement to elaborate, immediacy and efficiency, helpfulness and learning support, and growth and opportunity to revise. Students reportedly valued feedback that identified strengths and areas for growth, explained what needed improvement and why, and enabled immediate reflection through rapid response (Yu et al., 1 Aug 2025).

Mixed or negative themes included emotional complexity, feedback not always context-sensitive, technical issues, and lack of feedback archiving (Yu et al., 1 Aug 2025). Some students expressed appreciation, curiosity, frustration, skepticism, or anxiety; some questioned grading accuracy; some felt the system requested elaboration even when answers were already conceptually correct; and some reported answers not saving, page reloads causing loss of work, confusion about submission or saving, and reduced trust after glitches. Students also wanted a way to save or revisit feedback later (Yu et al., 1 Aug 2025).

These findings align with broader work on trustworthy educational AI. The chapter on bidirectional human-AI alignment in education identifies equity, privacy, autonomy, transparency, and trust as major concern areas, and argues that trustworthy learning environments arise not only from embedding human values into AI systems but also from equipping teachers, students, and institutions with the skills to interpret, critique, and guide these technologies (Shen, 25 Dec 2025). EduAlly’s explicit emphasis on transparency, reflection, and teacher-finalized grading is consistent with that orientation. This suggests that EduAlly can be understood as an instance of educational AI designed to preserve agency and oversight while scaling feedback.

6. Relation to adjacent research, infrastructure, and scope

EduAlly occupies a specific position within the contemporary educational AI landscape. Relative to dataset-centered work, EdNet provides a large-scale hierarchical dataset of diverse student activities collected by Santa, with 131,441,538 interactions from 784,309 students over more than two years, four hierarchical abstraction levels, and action types ranging from question solving to lecture consumption and item purchasing (Choi et al., 2019). Such datasets support tasks including knowledge tracing, learning path recommendation, score prediction, dropout prediction, and student simulators. A plausible implication is that systems such as EduAlly could benefit from similarly rich interaction modeling if extended beyond writing-intensive revision tasks.

Relative to platform and interoperability work, the $K$3Ed API proposes a shared, platform-independent API for educational microservices initially specified for feedback, assessment, and educational chatbots, with /evaluate and /chat implemented in v0.1 and /generate, /recommend, and /analyze planned (Sölch et al., 20 Feb 2026). Because EduAlly combines AI feedback generation, assessment-related workflows, and instructor-facing oversight, it fits naturally within the class of systems that could be decomposed into interoperable educational microservices.

Relative to pedagogical interface design, SpaceRaceEdu represents an educational multiplayer videogame prototype for self-study and self-assessment that balances competition and cooperation, supports teacher-authored question banks, and rewards correct answers without penalizing wrong ones (Gómez et al., 2024). Its pedagogical framing differs substantially from EduAlly’s writing-focused revision loop, but both systems share a concern with low-stakes formative support, self-assessment, and teacher configurability.

The limitations of EduAlly are explicitly acknowledged. The pilot was conducted at a single institution, involved a small sample size, had uneven self-evaluation data, was domain specific to writing-intensive tasks, and was affected by technical issues and missing data (Yu et al., 1 Aug 2025). Future directions include extending AI-EDL to new disciplines, expanding instructor engagement, refining AI feedback mechanisms, improving technical reliability, creating an accessible feedback archive, and tailoring feedback more effectively to student emotional and cognitive responses (Yu et al., 1 Aug 2025).

Taken together, the current literature presents EduAlly as a theory-driven, human-in-the-loop AI platform whose distinctive contribution lies in operationalizing a developmental feedback loop rather than merely attaching AI generation to existing assessment workflows. Its significance lies less in autonomous decision-making than in its attempt to formalize a supervised cycle of feedback, reflection, revision, and final instructor judgment (Yu et al., 1 Aug 2025).