GroupRAG: Cognitive-Inspired Group-Aware Retrieval and Reasoning
A novel framework named GroupRAG, drawing from cognitive science principles, seeks to enhance the efficacy of language models by tackling issues related to knowledge and reasoning. In contrast to conventional Retrieval-Augmented Generation (RAG) and Chain-of-Thought (CoT) techniques that depend on sequential reasoning paths, GroupRAG discerns hidden structural clusters within a problem and conducts retrieval and reasoning from various conceptual angles. This method facilitates a detailed interplay between the retrieval and reasoning mechanisms. The framework underwent testing using the MedQA medical dataset. The research paper can be found on arXiv with the identifier 2603.26807.
Key facts
- GroupRAG is a cognitively inspired group-aware retrieval and reasoning framework.
- It is based on knowledge-driven keypoint grouping.
- It identifies latent structural groups within a problem.
- It performs retrieval and reasoning from multiple conceptual starting points.
- It enables fine-grained interaction between retrieval and reasoning processes.
- Experiments were conducted on the MedQA medical dataset.
- The paper is available on arXiv with identifier 2603.26807.
- The approach addresses limitations of RAG and CoT in real-world settings.
Entities
Institutions
- arXiv