Scholarlink

Large Language Models (LLMs) are frequently used as study assistants in education, yet their reliability is often compromised by overconfidence, hallucinations, and challenges in handling ambiguity or outdated information. This study explores enhancing LLM performance through fine-tuning, Retrieval-Augmented Generation (RAG), and confidence calibration. Fine-tuning was conducted on datasets like FEVER and SQuAD v2, enabling task-specific improvements in claim verification and question answering. Confidence calibration, combining logit-based scoring with self-assessment, improved reliability and interpretability. Results show significant gains in accuracy and self-awareness compared to baseline models. Future work includes integrating these advancements into a user-friendly platform where students can access tailored assistance, and tutors can upload resources to enhance RAG pipelines. This research lays the groundwork for building robust and effective educational tools powered by LLMs.

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

Empowering University AI Assistants: Improving LLM Performance with RAG and Self-evaluation

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Polina Shnaider,
  • Anastasiia Chernysheva,
  • Anton Govorov,
  • Maksim Khlopotov

摘要

Large Language Models (LLMs) are frequently used as study assistants in education, yet their reliability is often compromised by overconfidence, hallucinations, and challenges in handling ambiguity or outdated information. This study explores enhancing LLM performance through fine-tuning, Retrieval-Augmented Generation (RAG), and confidence calibration. Fine-tuning was conducted on datasets like FEVER and SQuAD v2, enabling task-specific improvements in claim verification and question answering. Confidence calibration, combining logit-based scoring with self-assessment, improved reliability and interpretability. Results show significant gains in accuracy and self-awareness compared to baseline models. Future work includes integrating these advancements into a user-friendly platform where students can access tailored assistance, and tutors can upload resources to enhance RAG pipelines. This research lays the groundwork for building robust and effective educational tools powered by LLMs.