"My Grade is Wrong!": A Contestable AI Framework for Interactive Feedback in Evaluating Student Essays

Shengxin Hong, Chang Cai, Sixuan Du, Haiyue Feng, Siyuan Liu, Xiuyi Fan

Abstract

Interactive feedback, where feedback flows in both directions between teacher and student, is more effective than traditional one-way feedback. However, it is often too time-consuming for widespread use in educational practice. While Large Language Models (LLMs) have potential for automating feedback, they struggle with reasoning and interaction in an interactive setting. This paper introduces CAELF, a Contestable AI Empowered LLM Framework for automating interactive feedback. CAELF allows students to query, challenge, and clarify their feedback by integrating a multi-agent system with computational argumentation. Essays are first assessed by multiple Teaching-Assistant Agents (TA Agents), and then a Teacher Agent aggregates the evaluations through formal reasoning to generate feedback and grades. Students can further engage with the feedback to refine their understanding. A case study on 500 critical thinking essays with user studies demonstrates that CAELF significantly improves interactive feedback, enhancing the reasoning and interaction capabilities of LLMs. This approach offers a promising solution to overcoming the time and resource barriers that have limited the adoption of interactive feedback in educational settings.

The Challenge of Automated Interactive Feedback

In modern pedagogy, "interactive feedback"—a bidirectional exchange where students can query and challenge their evaluations—is recognized as far more effective than static, one-way grading. Despite its benefits, the sheer time required for human instructors to engage in these dialogues makes it difficult to scale.

While Large Language Models (LLMs) offer a potential solution for automation, they currently face two critical hurdles in educational settings:

  1. Reasoning Gaps: LLMs often struggle with the subtle nuances required for high-level critical thinking assessment.
  2. Over-reliance on User Input: LLMs are frequently "people-pleasers." If a student aggressively contests a grade, standard LLMs often hallucinate justifications to agree with the student, even if the original assessment was correct. This tendency to "fold" undermines the pedagogical integrity of the assessment.

The CAELF Framework

To address these issues, researchers developed the Contestable AI Empowered LLM Framework (CAELF). This system moves away from single-prompt grading, utilizing a structured, three-stage pipeline designed to maintain evaluative consistency while allowing for meaningful student contestation.

The CAELF framework architecture, illustrating the flow from student essay input to interactive feedback through multi-agent discussion and formal argumentation.

1. Multi-Agent LLM Discussion

Multiple "Teaching Assistant" (TA) agents, powered by GPT-4o-mini, discuss the essay. To ensure a breadth of perspective, agents are assigned different "personality biases"—for example, one leaning toward positive reinforcement and another toward critical rigor. This diversity forces a more thorough exploration of the rubric.

2. Formal Reasoning via the Teacher Agent

Rather than simply averaging TA opinions, a "Teacher Agent" maps these arguments into an Abstract Argumentation Framework (AF). This symbolic logic structure treats arguments as nodes and rebuttals as edges. The system employs "completeness semantics" to identify the most logically consistent set of arguments, which then forms the final grade and feedback.

3. Interactive Feedback Loop

When a student challenges their feedback, their argument is integrated as a new node in the argumentation framework. The TA agents then hold a new round of discussion to determine if the student’s point logically "defeats" the previous reasoning or if the original grade should stand. This symbolic logic acts as a "guardrail," preventing the model from making arbitrary concessions.

Experimental Results

The framework was tested on a dataset of 500 critical thinking essays evaluated across four dimensions: Issue, Evidence, Position, and Conclusion. Ground truth was established by four human coders labeling 2,000 data points.

Performance comparison across four rubric dimensions, highlighting CAELF's superior ability to maintain truth during student interactions compared to baseline LLMs.

CAELF significantly outperformed baseline models like GPT-4o and Llama-3.1-8B, particularly in interactive scenarios:

  • Robustness and "Maintain Truth": While standard LLMs often retract correct grades when faced with invalid student complaints, CAELF achieved a "Maintain Truth" rate of 80%–90% across dimensions. In contrast, baselines often fell below 40%.
  • Interaction Accuracy: CAELF's accuracy remained stable or improved after interaction. In the "Evidence" dimension, CAELF’s accuracy after student interaction was 44.6% higher than GPT-4o-mini.
  • Human Preference: In blind studies, human evaluators rated CAELF’s feedback significantly higher for factuality and its ability to help students reflect and improve.

Conclusion and Future Work

By wrapping the creative output of LLMs in a formal argumentation framework, CAELF provides a robust solution for scalable, interactive feedback. While the system is currently vulnerable to adversarial prompt injections (jailbreaking), future work aims to integrate Retrieval-Augmented Generation (RAG) to further ground agents in external facts and develop defenses against such attacks.

Summary updated

Abstract

Interactive feedback, where feedback flows in both directions between teacher and student, is more effective than traditional one-way feedback. However, it is often too time-consuming for widespread use in educational practice. While Large Language Models (LLMs) have potential for automating feedback, they struggle with reasoning and interaction in an interactive setting. This paper introduces CAELF, a Contestable AI Empowered LLM Framework for automating interactive feedback. CAELF allows students to query, challenge, and clarify their feedback by integrating a multi-agent system with computational argumentation. Essays are first assessed by multiple Teaching-Assistant Agents (TA Agents), and then a Teacher Agent aggregates the evaluations through formal reasoning to generate feedback and grades. Students can further engage with the feedback to refine their understanding. A case study on 500 critical thinking essays with user studies demonstrates that CAELF significantly improves interactive feedback, enhancing the reasoning and interaction capabilities of LLMs. This approach offers a promising solution to overcoming the time and resource barriers that have limited the adoption of interactive feedback in educational settings.
BibTeX
@inproceedings{hong2025my,
  author = {Shengxin Hong and Chang Cai and Sixuan Du and Haiyue Feng and Siyuan Liu and Xiuyi Fan},
  title = {{"My Grade is Wrong!": A Contestable AI Framework for Interactive Feedback in Evaluating Student Essays}},
  booktitle = {Aied},
  year = {2025},
  doi = {10.1007/978-3-031-98462-4\_4},
}
Figure