CHiL(L)Grader: Calibrated Human-in-the-Loop Short-Answer Grading

1. Problem Statement

The paper addresses the deployment of Large Language Models (LLMs) for Automated Short-Answer Grading (ASAG) in high-stakes educational settings. While LLMs show promise, two fundamental challenges prevent their fully autonomous use:

1. Systematic Overconfidence: Instruction-tuned LLMs often assign high confidence scores to incorrect predictions. This miscalibration makes it impossible for educators to distinguish between trustworthy and unreliable outputs, leading to potential grading errors.
2. Distribution Shift: Model performance degrades significantly when encountering new question types, evolving grading rubrics, or different student response patterns (distribution shift) that differ from the initial training data.

Existing solutions either lack uncertainty estimation (making them unsafe for deferral) or rely on static fine-tuning that fails to adapt to changing conditions without catastrophic forgetting.

2. Methodology: The CHiL(L)Grader Framework

The authors propose CHiL(L)Grader, a unified framework that integrates calibrated confidence estimation, selective prediction, and continual learning into a Human-in-the-Loop (HiL) workflow. The system operates in iterative cycles:

A. Core Components

1. Instruction-Tuned Base Model:

   • A pre-trained LLM (specifically Qwen-2.5-7B-Instruct) is fine-tuned using LoRA (Low-Rank Adaptation) on a dataset of question-answer pairs with human grades.
   • The model is trained to output a structured JSON string containing the predicted grade and the maximum possible grade.

2. Post-Hoc Calibration (Temperature Scaling):

   • To address overconfidence, the system applies Temperature Scaling on a held-out calibration set.
   • It minimizes the Expected Calibration Error (ECE) by finding an optimal temperature scalar ($T^*$) that rescales the model's logit outputs.
   • This transforms raw, overconfident probabilities into well-calibrated confidence scores that accurately reflect the model's true reliability.

3. Confidence-Based Selective Prediction:

   • A threshold ($\tau$) is applied to the calibrated confidence score.
   • High Confidence ($\hat{c} \ge \tau$): The prediction is automatically accepted as the final grade.
   • Low Confidence ($\hat{c} < \tau$): The case is deferred to a human grader for review. This ensures human effort is allocated only where the model is uncertain.

4. Continual Learning Loop with Replay:

   • Human corrections for deferred cases form a new supervision set ($H$).
   • To prevent catastrophic forgetting (where learning new data degrades performance on old data), the system employs a Replay Buffer.
   • For every new correction, the system retrieves $k$ similar historical examples from the original training data.
   • The model is fine-tuned on the union of new corrections and replayed samples using LoRA adapters, allowing it to adapt to new rubrics while retaining general grading capabilities.
   • After each update, the temperature $T^*$ is recalibrated on the current accepted samples.

B. Workflow Visualization

The process follows a loop:

1. Predict: Model generates grade + confidence.
2. Route: High confidence $\to$ Accept; Low confidence $\to$ Human Review.
3. Correct: Human provides ground truth for rejected cases.
4. Adapt: Fine-tune model on (Human Corrections + Replay Buffer) and recalibrate confidence.
5. Repeat: The model improves its coverage of high-confidence, expert-level predictions over time.

3. Key Contributions

1. Novel Framework: First ASAG system to integrate confidence calibration, selective prediction, and principled human deferral into a single, adaptive HiL loop.
2. Reliability via Calibration: Demonstrates that temperature scaling significantly reduces ECE (by 53–65% on some datasets), enabling the model to reliably identify when it should defer to humans.
3. Adaptation to Distribution Shift: Proves that HiL-based continual learning allows the model to generalize to unseen questions and evolving rubrics, maintaining performance where static models fail.
4. Expert-Level Automation: Achieves expert-level grading quality (Quadratic Weighted Kappa $\ge 0.80$) on 35–65% of student responses automatically, while routing the remaining uncertain cases to humans.
5. Reproducibility: All code, configurations, and evaluation scripts are publicly available.

4. Experimental Results

The framework was evaluated on three diverse datasets: DAMI (Data Mining, multi-scale), SciEntsBank (K-6 Science, unseen questions), and EngSAF (Engineering, coarse scale).

• Calibration Performance:
  • Baseline models had high ECE (e.g., 0.270 on DAMI).
  • After temperature scaling, ECE dropped to 0.094 (DAMI) and 0.046 (EngSAF), entering the "well-calibrated" zone ($<0.1$).
• Grading Quality (QWK):
  • DAMI: A single HiL cycle improved the QWK of accepted samples from 0.458 to 0.882 (surpassing the expert threshold of 0.80) while maintaining 35.1% automated coverage.
  • SciEntsBank: Showed progressive improvement across cycles, recovering from distribution shifts that caused temporary performance drops.
  • Routing Effectiveness: There was a significant QWK gap of +0.347 between accepted (high confidence) and rejected (low confidence) predictions, confirming the confidence gate effectively separates reliable from unreliable cases.
• Ablation Studies:
  • Removing the Replay Buffer caused catastrophic forgetting, dropping QWK to near-random levels (0.025).
  • Smaller models (3B parameters) failed to retain prior knowledge under fine-tuning.
  • The optimal confidence threshold ($\tau$) balances automation coverage and quality; lower thresholds increase coverage but may slightly reduce average QWK.

5. Significance and Conclusion

CHiL(L)Grader represents a paradigm shift from "fully autonomous" grading to "reliable partial automation."

• Practicality: It reduces instructor workload by automating the majority of high-confidence cases while ensuring safety through human oversight for uncertain cases.
• Robustness: By continuously learning from teacher feedback and using replay buffers, the system adapts to curriculum changes without losing previous knowledge.
• Trust: The integration of calibrated uncertainty quantification provides the necessary trust mechanism for deploying AI in high-stakes educational assessment.

The paper concludes that uncertainty quantification is not just a theoretical metric but a critical component for the safe and effective deployment of AI in education, enabling systems that are both efficient and trustworthy.