Technical Summary: DeanLLM – A Framework for Automated Quality Review of AI-generated Feedback

Problem Statement

Large Language Models (LLMs) are increasingly deployed as tutors to generate educational feedback. However, existing research has predominantly focused on the generation of feedback rather than its evaluation. This gap creates significant risks: LLM-generated feedback may lack pedagogical value, fail to align with learning goals, or contain hallucinations (fabricated, irrelevant, or inconsistent information) that mislead students. Current automated evaluation methods often address single dimensions (e.g., correctness or praise) or fail to integrate hallucination detection with pedagogical quality assessment. There is a critical need for a comprehensive, automated review framework that can evaluate feedback for content quality, educational effectiveness, and safety before it is delivered to learners.

Methodology

The study introduces DeanLLM, an automated review framework designed to evaluate LLM-generated tutor feedback prior to student delivery. The methodology involved a three-stage offline design:

1. Data Preparation and Synthesis:

   • The researchers utilized real computer science assignment data from a university (2021 semester) but could not release it due to privacy and copyright.
   • They generated 1,000 synthetic assignment descriptions and submissions using a one-to-one LLM-based imitation procedure (using o3-pro and o4-mini) to preserve task structure and semantic meaning while removing student identities.
   • A "privacy gate" using o4-mini ensured no real student identifiers were preserved in the synthetic data.

2. Evaluation Framework Development (16 Dimensions):

   • A comprehensive rubric was developed covering three domains:
     • Content Quality (6 dimensions, 0–2 scale): Alignment with goals, Specificity, Motivational Tone, Strength, Weakness, and Clarity/Understandability.
     • Educational Effectiveness (7 dimensions, binary): Feedback direction (Feed up, Feed back, Feed forward) and Feedback level (Task, Process, Self-regulation, Self).
     • Hallucination Risks (3 dimensions, binary): Input-conflicting, Context-conflicting, and Fact-conflicting hallucinations.

3. Experimental Stages:

   • Ground Truth Creation: Three human experts annotated 200 feedback instances (generated by 8 different LLM tutors) using the 16-dimension rubric to create a ground truth dataset.
   • Reviewer Calibration (RQ2): The study tested various configurations of LLMs (GPT-4.1, o4-mini, o3, o3-pro) as automated reviewers. Conditions included:
     • Zero-shot and Few-shot prompting.
     • Supervised Fine-Tuning (SFT) on human-labelled data with three variations: plain labels only, explanatory labels (with rationales), and mixed data.
   • Benchmarking (RQ3): The best-performing reviewer configuration was used to evaluate feedback generated by 10 commercial LLM tutors (OpenAI, Google, DeepSeek) on 2,000 instances (1,000 real, 1,000 synthetic).

4. Analysis:

   • Psychometric analyses (correlation matrices, item-total correlations, Exploratory Factor Analysis, t-SNE) were used to validate the framework's structure and compare human vs. LLM evaluation patterns.
   • Agreement metrics (Accuracy, Cohen's $\kappa$, F1) were calculated to measure reviewer alignment with human experts.
   • Composite scores (0–100) were derived for Content Quality, Educational Effectiveness (using theory-driven weights), and Safety/Factuality (using a strict zero-tolerance veto rule).

Key Contributions

1. DeanLLM Framework: The introduction of a scalable, automated review architecture that integrates pedagogical quality assessment with hallucination detection. It operates as a "quality gate" to screen, rank, or trigger regeneration of feedback candidates before delivery.
2. Comprehensive 16-Dimension Rubric: The operationalization of feedback quality across content, educational effectiveness, and safety, explicitly addressing the gap in integrating hallucination risks into automated educational feedback evaluation.
3. Calibration Insights: Empirical evidence demonstrating that supervised fine-tuning with plain-labelled human examples significantly outperforms zero-shot and few-shot prompting, as well as fine-tuning with explanatory rationales, in aligning automated reviewers with human expert judgments.
4. Model Benchmarking: A comparative analysis of 10 commercial LLMs, revealing that model choice significantly impacts the pedagogical depth and factual reliability of generated feedback.

Results

• Framework Validity: Psychometric analyses confirmed the 16 dimensions form a reliable basis for evaluation. However, a structural divergence was observed: human reviewers tended to evaluate feedback holistically (showing inter-dimensional correlations or a "halo effect"), whereas LLM reviewers evaluated dimensions mechanically and orthogonally (near-zero correlations between dimensions).
• Reviewer Performance:
  • Standard prompting (zero-shot/few-shot) showed limited agreement with human experts on Content Quality ($\kappa \approx 0.0$).
  • GPT-4.1 fine-tuned on plain-labelled data achieved the strongest alignment with human experts: Accuracy = 79.8%, $\kappa = 0.686$, and F1 = 79.4%.
  • Fine-tuning with explanatory rationales performed worse than plain-label fine-tuning, suggesting that adding open-ended explanations increased complexity without improving label accuracy for this specific task.
  • Reasoning-oriented models (e.g., o3-pro) were particularly effective at detecting hallucinations.
• Tutor Benchmarking:
  • Content Quality: Most high-capacity models scored near-ceiling (e.g., Gemini 2.5 Pro: 99.38).
  • Educational Effectiveness: Reasoning models (o3, o4-mini) outperformed others, generating more feed-forward and process-level feedback. Lightweight models (e.g., GPT-4.1 nano) scored lower.
  • Safety/Factuality: Hallucination risks were concentrated in specific models. GPT-4.1 nano had the lowest safety score (80.50), while Gemini 2.5 Pro and GPT-4.1 achieved near-perfect safety scores (99.50 and 99.00).
  • The study found no single uniform ranking; a model could excel in content quality but fail in safety or pedagogical depth.

Significance and Claims

The paper claims that DeanLLM offers a scalable pathway to improve the reliability and safety of LLM tutor feedback by treating automated review as a calibration and governance problem rather than a simple text-scoring task.

• Pedagogical Safety: The framework demonstrates that automated review can effectively screen for hallucinations and misaligned feedback, acting as a necessary safeguard before feedback reaches students.
• Calibration Necessity: The findings emphasize that "out-of-the-box" prompting is insufficient for nuanced educational evaluation; high-quality, task-aligned fine-tuning is critical for aligning automated reviewers with human pedagogical judgment.
• Model Selection: The study argues that model choice is a critical deployment variable. Reasoning models are better suited for generating pedagogically effective feedback, while lightweight models may require stricter oversight or human-in-the-loop validation due to higher hallucination risks.
• Governance Layer: DeanLLM positions automated review not as a replacement for teacher judgment, but as a complementary layer that ensures feedback meets specific safety and quality thresholds before delivery, supporting responsible AI deployment in education.

The authors remain modest, noting that the study is offline and does not measure actual student learning outcomes. They acknowledge that the synthetic data, while privacy-preserving, cannot fully replicate the complexity of real classroom dynamics, and that the framework requires ongoing validation against diverse educational contexts.