TL;DR
This paper introduces SURE, a human-in-the-loop grading system that uses LLMs to grade student work. By leveraging self-consistency and selective human review, SURE significantly improves grading accuracy and efficiency, reducing manual grading time by up to 90% while maintaining human-level performance.
What Problem Does This Paper Address?
Grading open-ended assessments like essays or coding assignments is a time-consuming and often tedious task for educators. While Large Language Models (LLMs) show promise for automating this process, they currently suffer from inconsistent accuracy, systematic biases, and limited reliability across different assignments. This makes fully automating grading risky, as highlighted by the EU's AI Act classifying AI grading as "high risk" and mandating human oversight. The authors aim to bridge this gap by developing a system that combines the efficiency of LLMs with the reliability of human review, ensuring accurate and fair grading without overwhelming educators.
Methods
The core of the proposed system, SURE (Selective Uncertainty-based Re-Evaluation), involves repeatedly prompting LLMs to grade the same student answer multiple times (20 times in this study). This generates a distribution of scores from which a predicted score (e.g., majority vote) and a certainty estimate (based on score variability) are derived. Low-certainty scores, indicating potential errors or ambiguities, are flagged for selective human review.
To enhance the LLM's output diversity and thus improve uncertainty estimation, the researchers explored several strategies:
- Parameter Variation: Adjusting temperature and top-p parameters to control randomness.
- Prompt Perturbations: Shuffling rubric order, using different grader personas (strict, lenient), and multilingual prompting.
- LLM Ensembles: Combining the outputs of multiple LLMs to reduce model-specific biases and stabilize predictions.
The system was trained on one assignment to tune certainty thresholds and then evaluated on the remaining assignments using a fixed threshold.
Datasets
The study utilized data from 46 graduate students enrolled in an introductory programming course. This dataset comprised 130 coding exercises and open-ended questions across five assignments. The data was used for both training and testing the SURE system. Human graders provided ground-truth scores for comparison.
Key Insights and Ideas
- Self-consistency is key: Repeatedly prompting LLMs and observing the consistency of their scores is a strong indicator of grading correctness. When LLMs are highly consistent, their scores are likely accurate.
- Low certainty flags errors: Conversely, high variability in LLM scores for the same answer signals low certainty and is a strong predictor of incorrect grading. This allows for targeted human intervention.
- Ensembles are better than single models: Combining multiple LLMs, especially those with diverse reasoning capabilities, significantly improves the ability to distinguish between correct and incorrect scores, leading to more reliable uncertainty estimates.
- Human oversight is still valuable, but selective: SURE demonstrates that a human-in-the-loop approach, where humans only review flagged low-certainty cases, can achieve human-level accuracy while drastically reducing manual grading workload.
- Reasoning matters more than size: The study suggests that the reasoning capabilities of LLMs are more critical for accurate grading than simply having a large number of parameters.
Key Results
- Fully automated grading with a single LLM prompt resulted in significant underscoring.
- Majority voting (averaging 20 prompts) improved accuracy but didn't eliminate bias.
- SURE, with selective human review, reduced manual grading time by 40-90% while improving grading accuracy.
- LLM ensembles consistently approached human-level accuracy, with 70-90% of student grades falling within human-grader ranges.
- In the test set, SURE with an ensemble achieved 84-91% of grades within target ranges with minimal grade point deviations.
- Time savings varied, with individual LLMs sometimes exceeding 80% reduction in manual grading time, while ensembles offered more moderate but consistent savings.
Limitations and Future Work
The authors acknowledge several limitations:
- Human grading as a benchmark: While interrater reliability was high, human grading is not infallible, and the study's findings are based on aggregated human scores, which might not be absolute ground truth. Future work could incorporate more objective measures like unit tests where applicable.
- Simulated human regrading: The random sampling of human graders for regrading might underestimate the impact of individual grader biases. Future work should consider how to best simulate real-world grading scenarios with potential grader tendencies.
- Uncertainty operationalization: The study primarily used prompt agreement for uncertainty. Future research could explore alternative distributional metrics and their impact on flagging.
- Generalizability: The study was conducted on a single programming course with specific assignments. Future work needs to validate SURE across diverse domains, educational levels, and task formats.
- Threshold sensitivity: Optimal flagging thresholds appear to be assignment- and model-sensitive, posing a practical challenge for setting thresholds in advance. Future work should investigate heuristics or adaptive thresholding.
- Bias in flagging: The potential for certain answer styles or languages to systematically increase or decrease flagging likelihood needs further investigation to ensure fairness.
Promising future directions include combining SURE with student self-grading for enhanced feedback and learning, and exploring more diverse ensembles and advanced uncertainty metrics.
Summary updated