PaperlayerView exact revision on arXiv

Source-linked AI summary

Human-AI Collaborative Essay Scoring: A Dual-Process Framework with LLMs

Changrong Xiao, Wenxing Ma, Qingping Song, Sean Xin Xu, Kunpeng Zhang, Yufang Wang, Qi Fu

arXiv:2401.06431v2cs.CLcs.AI

TL;DR

Second-language learners need timely feedback, but high student-teacher ratios and AES deployment challenges complicate that support. The paper evaluates LLMs and proposes an open-source dual-process AES system; its reported results include expert-comparable novice grading with AI support and improved grader efficiency.

  • Problem

    Second-language learners need timely feedback, but high student-teacher ratios, ambiguous rubrics, and potential model errors complicate reliable AES use.

    Source: §1 · p. 1

  • Method

    The authors evaluate proprietary and open-source LLMs and build an open-source dual-process AES system with fast score prediction and slower explanation generation.

    Source: §1 · p. 2

  • Results

    The open-source framework matches proprietary-model quality with greater efficiency, while co-grading brought novice graders to expert-comparable accuracy and improved expert graders’ efficiency and consistency.

    Source: §7 · p. 8

  • Takeaways & Limitations

    The findings point to human-AI co-grading as a promising workflow, particularly for cases where model confidence is low.

    Source: §4.4 · p. 5, §7 · p. 8

    CSEE covers only two final exams for Chinese high-school English learners, and the human-AI collaboration experiment is a pilot study.

    Source: Limitations · p. 9

Abstract

from arXiv · show

Receiving timely and personalized feedback is essential for second-language learners, especially when human instructors are unavailable. This study explores the effectiveness of Large Language Models (LLMs), including both proprietary and open-source models, for Automated Essay Scoring (AES). Through extensive experiments with public and private datasets, we find that while LLMs do not surpass conventional state-of-the-art (SOTA) grading models in performance, they exhibit notable consistency, generalizability, and explainability. We propose an open-source LLM-based AES system, inspired by the dual-process theory. Our system offers accurate grading and high-quality feedback, at least comparable to that of fine-tuned proprietary LLMs, in addition to its ability to alleviate misgrading. Furthermore, we conduct human-AI co-grading experiments with both novice and expert graders. We find that our system not only automates the grading process but also enhances the performance and efficiency of human graders, particularly for essays where the model has lower confidence. These results highlight the potential of LLMs to facilitate effective human-AI collaboration in the educational context, potentially transforming learning experiences through AI-generated feedback.

1 Introduction

The introduction motivates AES as a source of timely feedback and grading support, while emphasizing the need for reliable scores and effective human-AI collaboration. It presents an open-source dual-process framework, a new essay dataset, and co-grading findings.

1 Introduction

  • High student-teacher ratios make timely feedback difficult, while ambiguous rubrics and potential model errors make unsupervised AES impractical.The introduction argues that practical systems need accurate scoring, natural-language explanations, and support for human-AI collaboration.Source: §1 · p. 1
  • The proposed dual-process AES framework is presented as accurate, efficient, and explainable, including for complex tasks with tailored grading criteria.The study explores proprietary and open-source LLMs and reports that its system matches fine-tuned LLaMA3 in grading accuracy and feedback quality.Source: §1 · p. 2
  • The authors introduce a dataset of 13,372 Chinese high-school essays scored across multiple dimensions by expert educators.The dataset is positioned as an expanded resource for AI in education.Source: §1 · p. 2
  • LLM-generated scores and explanations improved novice graders’ proficiency to an expert-comparable level and supported grader efficiency and robustness.The introduction also reports that confidence scores and explanations are integrated into this support.Source: §1 · p. 2

2 Related Work

Related work covers conventional AES approaches, human-AI decision support, and dual-process LLM architectures. The authors identify a gap in applying collaborative co-grading to AES.

2.1 Automated Essay Scoring (AES)

  • Traditional AES progressed from feature-based regression and classification through neural models to BERT-centered state-of-the-art methods (Wang et al., 2022).Earlier approaches also included CNNs, LSTMs, and pretrained language models.Source: §2.1 · p. 2
  • Prior LLM AES studies used rubric-based or few-shot prompting, but reported scores did not surpass conventional baselines or human annotators (Han et al., 2023; Stahl et al., 2024).The cited work includes IELTS scoring and CEFR-level prediction for second-language essays.Source: §2.1 · p. 2

2.2 AI-Assisted Decision Making

2.3 Dual-Process Theory

  • Dual-process theory distinguishes rapid, intuitive Fast Thinking from deliberate Slow Thinking, and LLM architectures inspired by it have been used in complex tasks (Lin et al., 2024; Tian et al., 2023).Prior adaptations sought to address issues such as social bias and hallucination.Source: §2.3 · p. 3

3 Data

The study uses the public ASAP dataset and introduces CSEE, a Chinese high-school English essay dataset scored by experienced teachers.

3 Data

  • ASAP contains 12,978 human-scored essays across eight prompts, with prompt-specific rubrics and score ranges.It is described as a commonly used public resource for AES.Source: §3 · p. 3
  • CSEE comprises 13,372 essays from two final-exam prompts, collected with 29 Chinese high schools and scored by experienced teachers.Scores cover Content, Language, Structure, and an Overall Score from 0 to 20 under Chinese National College Entrance Examination guidelines.Source: §3 · p. 3

4 Methods

The methods compare conventional AES baselines with prompted and fine-tuned LLMs, then introduce a dual-process framework that separates rapid scoring from detailed explanation generation.

4.1 Traditional Essay Scoring Baselines

  • The study evaluates traditional AES baselines, LLM approaches, and the proposed Fast and Slow Thinking framework.The BERT baseline jointly trains an unfrozen BERT model and a fully connected prediction layer; R2BERT and Tran-BERT-MS-ML-R serve as ASAP state-of-the-art benchmarks.Source: §4 · p. 3, §4.1 · p. 3

4.2 Prompting LLMs

  • Prompting experiments vary rubric context and zero-shot versus few-shot settings, with few-shot prompts adding example essays and scores.The model is also tested with and without rubrics in zero-shot settings.Source: §4.2 · pp. 3–4
  • All prompting configurations use Chain-of-Thought instructions to analyze and explain the materials before assigning a final score (Wei et al., 2022).The cited strategy is presented as supporting inference and reasoning tasks.Source: §4.2 · p. 4

4.3 Fine-tuning LLMs

  • Because ASAP and CSEE provide scores without expert explanations, the authors augment them with GPT-4-generated explanations guided by expert examples and a structured template.The resulting instructional-format data supports fine-tuning models to produce both scores and feedback.Source: §4.3 · p. 4
  • The authors fine-tune GPT-3.5-turbo and LLaMA3-8B; they describe the open-source model as a more practical option given privacy, cost, and fine-tuning flexibility concerns.The authors cite the proprietary nature of GPT-3.5 as a factor in that choice.Source: §4.3 · p. 4

4.4 Our Proposed Method

  • The proposed system assigns quick score prediction to a Fast Module and detailed rubric-grounded reasoning, explanations, and scoring to a Slow Module.The Slow Module is implemented with fine-tuned LLaMA3-8B and is described as interchangeable with another qualified LLM.Source: §4.4 · p. 4
  • A lightweight fully connected Fast Module uses input embeddings to predict scores while reducing the time and computational cost of detailed text generation.It leverages latent knowledge acquired during fine-tuning of the Slow Module.Source: §4.4 · p. 4
  • The Fast Module standardizes output-score probabilities as confidence scores; low confidence triggers Slow Module self-reflection or referral to an external human or AI judge.During inference, the Slow Module is activated when confidence is low or additional requirements call for it.Source: §4.4 · p. 5

5 Experimental Results

Fine-tuned LLMs improved over zero- and few-shot scoring but remained below traditional SOTA methods, while open-source LLaMA3-8B was comparable to fine-tuned proprietary models. Further analyses report scoring consistency, generalization patterns, and a confidence-triggered framework that improves overall performance and efficiency.

5.1 Performance of LLM-based Methods

  • GPT-4’s zero-shot QWK was 0.0423 on Set 1 and 0.0809 on Set 7, and few-shot scoring still lagged significantly behind SOTA methods.Few-shot performance improved particularly on Sets 4–6; the authors suggest broad score ranges and complex rubrics may contribute to zero-shot underperformance.Source: §5.1 · p. 5
  • Fine-tuned LLMs generally exceeded 0.7 QWK but did not surpass traditional SOTA methods; fine-tuned LLaMA3-8B was comparable to proprietary models.CSEE results also indicate reliable scores across content, language, and structure dimensions.Source: §5.1 · p. 5, Table 1 · p. 6, Table 2 · p. 6
  • In a comparison of 20 paired explanation sets, the system’s explanations received 35% wins, 40% ties, and 25% losses against GPT-4.The authors describe their explanations as high-quality and comparable to GPT-4’s.Source: §5.1 · p. 5

5.2 Further Analyses

  • Over 80% of ratings remained unchanged across three trials; fine-tuned GPT-3.5 generally outperformed BERT across subsets, with specified exceptions.GPT-3.5 underperformed when trained on Set 4 and tested on Sets 1 and 7; BERT trained on Sets 7 or 8 generalized close to random guessing.Source: §5.2 · p. 6
  • Switching to the Slow Module below a 0.2 confidence score improved integrated-system performance, while the confidence-triggered framework traded increased accuracy for efficiency.Slow Module training took around 2 hours per epoch and inference about 10 GPU hours; Fast Module training took under 0.5 hours and scoring 0.2 hours.Source: §5.2 · pp. 6–7, Figure 3 · p. 7

6 Human-AI Co-Grading Experiment

The study tested whether novice and expert graders could collaborate with the AES system to improve grading, and whether confidence and explanations helped. LLM feedback improved novice graders’ performance to a level statistically comparable to experts; experts also graded faster and more consistently, while low-confidence cases remained challenging.

6.1 Experiment Design

  • The two-stage within-group study used 50 CSEE essays on one topic, with 10 novice and 5 expert evaluators grading before and after receiving model scores, confidence, and explanations.Evaluators could revise their scores after seeing the model outputs; questionnaires used a 5-point Likert scale.Source: §6.1 · p. 7

6.2 Results

  • Novices’ mean QWK rose from 0.5256 to 0.6609 with LLM feedback (p < 0.01), with no statistical difference from experts’ 0.7117 (p = 0.27).The reported trend also held for content, language, and structure scores.Source: §6.2 · p. 7, Table 3 · p. 7, Figure 4 · p. 8
  • Expert-assisted grading raised mean QWK from 0.7117 to 0.7734, above the AES system’s 0.7302, though the improvement was not statistically significant (p = 0.13).Experts also reported needing less time, and their ratings showed lower standard deviation.Source: §6.2 · pp. 7–8, Table 4 · p. 8
  • For low-confidence cases (below 0.2), expert and novice QWK were 0.6809 and 0.5680, respectively, exceeding the Slow Module’s 0.5478.The authors describe a workflow in which the model handles high-confidence routine cases and humans review low-confidence cases.Source: §6.2 · p. 8

7 Conclusion

The paper presents an open-source Fast and Slow Thinking AES framework and reports that it matches proprietary-model quality with greater efficiency. In co-grading experiments, novice graders reached accuracy comparable to experienced graders, while experts gained efficiency and consistency.

7 Conclusion

  • The open-source Fast and Slow Thinking AES framework matches proprietary-model quality while offering greater efficiency.The paper also reports that task-specific supervised fine-tuning brought LLM performance close to conventional SOTA methods.Source: §7 · p. 8
  • In co-grading experiments, novices supported by the framework reached accuracy comparable to experienced graders, while experts gained efficiency and consistency.The paper reports that collaboration particularly enhanced performance on low-confidence cases.Source: §7 · p. 8
  • The paper concludes that human-AI co-grading has potential for AI-assisted, human-centered decision-making in education.It identifies future human-AI collaboration and AI-assisted learning as directions for further work.Source: §7 · p. 8

Limitations

The authors identify limits to the dataset’s scope and to the generalizability of the human-AI collaboration findings.

Limitations

  • The CSEE essays come from only two final exams for high school English learners in China, limiting evidence across topics and student demographics.The authors raise concerns about system robustness on a broader range of topics and student populations.Source: Limitations · p. 9
  • The human-AI collaboration experiment is a pilot study, so its findings are general rather than a thorough account of team performance.Further experiments are needed to study when humans correct errors after AI feedback and when unreliable AI feedback may mislead them.Source: Limitations · p. 9
  • The mechanisms behind complementary human-AI team performance remain insufficiently explored.The authors specifically call for identifying the circumstances in which people catch errors or are misled by unreliable AI feedback.Source: Limitations · p. 9

Ethical Considerations

The authors report IRB approval for data collection and the co-grading experiment, anonymized student information, and research-only consent.

Ethical Considerations

  • IRB approval covered data collection and the human-AI co-grading experiment, and student personal information was anonymized.The dataset contains essays and corresponding scores without additional potentially sensitive information.Source: Ethical Considerations · p. 9
  • Teachers and students explicitly consented to use the data exclusively for research, but annotation details were undisclosed.The authors did not know how many teachers scored the essays or the inter-annotator agreement level.Source: Ethical Considerations · p. 9

A Datasets

The section describes the ASAP and CSEE datasets and the BERT classifier baseline used in the study. CSEE contains 13,372 teacher-scored essays; the baseline uses a fine-tuned BERT model.

A Datasets

  • CSEE contains 13,372 essays scored by experienced English teachers using China’s National College Entrance Examination standards; Table 5 presents its descriptive statistics.Source: A Datasets · p. 11, Table 5 · p. 11
  • ASAP comprises eight subsets, each with distinct prompts and scoring rubrics; its descriptive statistics are presented in Table 6.Source: A Datasets · p. 11

B Details of BERT Classifier Baseline

C Details of LLM-based Methods

The section details LLM prompting, few-shot example selection, fine-tuning, and human-AI co-grading procedures. Retrieval-based example selection outperformed random sampling, and Structure was the most difficult grading dimension.

C.1 LLM Prompts

  • The prompt materials cover GPT-4 zero-shot and few-shot settings and GPT-3.5 and LLaMA3-8B fine-tuning and inference; Table 9 shows example prompts and output.Source: §C.1 · pp. 11–12, Table 9 · p. 15

C.2 Few-Shot GPT-4

  • GPT-4 few-shot prompts use three examples, and retrieval by cosine similarity consistently outperformed random essay sampling.The retrieval approach used text-embedding-ada-002 and excluded the target essay.Source: §C.2 · p. 12

C.3 Fine-tuning LLaMA3

  • LLaMA3-8B fine-tuning used 4-bit quantization with PEFT and LoRA rank 16, ran for two epochs, and used one RTX 4090 24GB GPU for about three weeks.The setup used an 8-bit AdamW optimizer and an initial learning rate of 2 × 10−4.Source: §C.3 · p. 12

D Human-AI Co-Grading Details

  • The co-grading study involved 10 novice college students and five experienced high-school teachers, all trained on standard scoring rubrics before evaluation.Source: D Human-AI Co-Grading Details · p. 12
  • Evaluators received writing prompts and essays, while experimental groups also received supplementary scores, confidence values, and explanations; Table 7 compares QWK generalization for fine-tuned GPT-3.5 and BERT on ASAP.The instructions and materials were translated into Chinese, and Table 8 provides the private-dataset scoring rubrics.Source: D Human-AI Co-Grading Details · p. 12, Table 7 · p. 13, Table 8 · p. 14
  • Structure had the lowest QWK among the three dimensions and was identified by experts as the most ambiguous and difficult; Content and Language followed the Overall-score trend.Source: D Human-AI Co-Grading Details · p. 12, Figure 5 · p. 13

Related papers

Loading 2401.06431v2…