Scholarlink

This paper presents a novel approach to personalized K–12 education through the design and deployment of a Personalised Learning Assistant (PLA) powered by a fine-tuned Large Language Model (LLaMA 3.1). The PLA adaptively evaluates student understanding via a multi-level diagnostic test aligned with Bloom’s Taxonomy. Using probabilistic inference and difficulty-aware weighted scoring, the system generates a learner-specific conceptual profile while minimizing the influence of random guessing. We introduce a fixed-attempt policy to ensure fairness in progression and score normalization across students with varying test trajectories. The PLA was deployed through a functional web prototype, with MCQs generated using Gemini 1.0 and a fine-tuned LLM hosted on Kaggle and a Vercel-based frontend for seamless interaction. Our evaluation demonstrates the framework’s robustness in differentiating learner capabilities and its potential to inform targeted remediation. This work bridges generative AI, adaptive assessment, and cognitive pedagogy to enable scalable and equitable learning in classroom and self-paced environments.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

From Uniform to Unique: Adaptive K–12 Assessment Using Large Language Models

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Lokesh Goenka,
  • Ajay Mukund S,
  • P. Sunil Kumar

摘要

This paper presents a novel approach to personalized K–12 education through the design and deployment of a Personalised Learning Assistant (PLA) powered by a fine-tuned Large Language Model (LLaMA 3.1). The PLA adaptively evaluates student understanding via a multi-level diagnostic test aligned with Bloom’s Taxonomy. Using probabilistic inference and difficulty-aware weighted scoring, the system generates a learner-specific conceptual profile while minimizing the influence of random guessing. We introduce a fixed-attempt policy to ensure fairness in progression and score normalization across students with varying test trajectories. The PLA was deployed through a functional web prototype, with MCQs generated using Gemini 1.0 and a fine-tuned LLM hosted on Kaggle and a Vercel-based frontend for seamless interaction. Our evaluation demonstrates the framework’s robustness in differentiating learner capabilities and its potential to inform targeted remediation. This work bridges generative AI, adaptive assessment, and cognitive pedagogy to enable scalable and equitable learning in classroom and self-paced environments.