Skip to main content Skip to navigation

Explanations from Large Language Models Make Small Reasoners Better

Project Overview

The document examines the application of generative AI, specifically large language models (LLMs), in enhancing educational tools by improving the performance of smaller language models (SLMs) through the integration of generated explanations. It focuses on multi-task learning frameworks that utilize these explanations to advance reasoning skills, particularly in question-answering scenarios within educational settings. Key findings reveal that incorporating explanations generated by LLMs can lead to significant performance improvements in SLMs, sometimes even exceeding the capabilities of larger models in specific tasks. This underscores the critical role of explainable AI in education, as the generation of high-quality explanations not only boosts model effectiveness but also provides valuable insights for learners. Overall, the document highlights the transformative potential of generative AI in educational contexts, emphasizing its ability to refine learning tools and enhance students' understanding through improved communication and reasoning.

Key Applications

Multi-task Learning with Explanations from LLM

Context: Educational settings, targeting students and researchers in AI and NLP.

Implementation: Utilizing explanations generated by LLMs in a multi-task learning framework to train SLMs on reasoning tasks.

Outcomes: SLMs show significant performance improvements on reasoning tasks compared to single-task fine-tuning baselines.

Challenges: Requires extensive tuning of hyperparameters and may struggle with noisy explanations.

Implementation Barriers

Technical Barrier

Integration of explanation generation into training frameworks requires careful tuning and optimization.

Proposed Solutions: Use multi-task learning strategies to balance between prediction accuracy and explanation quality.

Quality Barrier

Generated explanations may sometimes be logically inconsistent or grammatically incorrect.

Proposed Solutions: Implement filtering mechanisms or combine multiple explanation generation strategies to enhance quality.

Project Team

Shiyang Li

Researcher

Jianshu Chen

Researcher

Yelong Shen

Researcher

Zhiyu Chen

Researcher

Xinlu Zhang

Researcher

Zekun Li

Researcher

Hong Wang

Researcher

Jing Qian

Researcher

Baolin Peng

Researcher

Yi Mao

Researcher

Wenhu Chen

Researcher

Xifeng Yan

Researcher

Contact Information

For information about the paper, please contact the authors.

Authors: Shiyang Li, Jianshu Chen, Yelong Shen, Zhiyu Chen, Xinlu Zhang, Zekun Li, Hong Wang, Jing Qian, Baolin Peng, Yi Mao, Wenhu Chen, Xifeng Yan

Source Publication: View Original PaperLink opens in a new window

Project Contact: Dr. Jianhua Yang

LLM Model Version: gpt-4o-mini-2024-07-18

Analysis Provider: Openai

Let us know you agree to cookies