"WisdomBot: Tuning Large Language Models with Artificial Intelligence Knowledge"

This paper enhances LLMs for education by incorporating educational theories and retrieval augmentation.

Introduces WisdomBot, an educational LLM

It addresses the limitations of general LLMs in education, such as limited knowledge, personalized learning needs, and the need for concise explanations.

-----

Paper - https://arxiv.org/abs/2501.12877

Original Problem 😞:

→ General LLMs struggle in education due to limited specialized knowledge, lack of personalization, and difficulty explaining complex concepts concisely.

→ They often give inaccurate responses due to constrained comprehension and outdated knowledge.

-----

Solution in this Paper 🤔:

→ WisdomBot combines LLMs with educational theories.

→ It leverages self-instructed knowledge concepts and instructions based on Bloom’s Taxonomy for training.

→ During inference, it uses local knowledge base and search engine retrieval to enhance responses.

-----

Key Insights from this Paper 💡:

→ Aligning LLMs with educational theories like Bloom's Taxonomy improves their performance in educational tasks.

→ Self-instruction can effectively generate training data for specialized knowledge and cognitive processes.

→ Retrieval augmentation enhances factual accuracy and provides broader context for responses.

-----

Results 👍:

→ WisdomBot outperforms baselines on various educational tasks, including professional question answering, test problem generation, and intelligent tutoring, with a minimum winning rate of 63%.

→ On C-Eval benchmark, WisdomBot shows superior performance, especially in information and computer science subjects, demonstrating the relevance of training data to those fields.