RARE: Retrieval-Augmented Reasoning Enhancement for Large Language Models.

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Machine Learning

Background:

• Large language models (LLMs) struggle with factual integrity and complex reasoning.
• Existing mutual reasoning frameworks require enhancement for knowledge-intensive domains.

Purpose of the Study:

• To introduce Retrieval-Augmented Reasoning Enhancement (RARE) for improving LLM reasoning accuracy and factuality.
• To enhance LLMs for complex tasks like medical and commonsense reasoning.

Main Methods:

• RARE extends the mutual reasoning framework (rStar) with two novel Monte Carlo Tree Search actions (A6 and A7).
• Action A6 augments reasoning with retrieved data based on the problem statement.
• Action A7 uses retrieval for sub-questions and re-answers, alongside a Retrieval-Augmented Factuality Scorer.

Main Results:

• RARE enables open-source LLMs (LLaMA 3.1) to achieve performance competitive with closed-source models (GPT-4, GPT-4o).
• The proposed factuality scorer prioritizes reasoning paths with high factual integrity.

Conclusions:

• RARE offers a scalable solution for enhancing LLM performance in critical domains.
• The framework significantly improves logical coherence and factual accuracy in LLM reasoning.