SA-RAG: Structured and adaptive retrieval-augmented generation for multi-hop question answering.

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Machine Learning

Background:

• Retrieval-Augmented Generation (RAG) integrates external knowledge into Large Language Models (LLMs) but struggles with Multi-Hop Question Answering (MHQA).
• Existing RAG methods face challenges with noisy contexts, reduced accuracy at scale, and poor generalization due to overfitting in self-supervised fine-tuning.
• Distribution shifts and longer reasoning chains further degrade performance in current MHQA systems.

Purpose of the Study:

• To propose SA-RAG, a Structured and Adaptive RAG framework, to overcome limitations in current MHQA.
• To enhance multi-hop reasoning accuracy, consistency, and cross-task generalization in LLMs.
• To provide a reliable framework for building robust MHQA systems.

Main Methods:

• Developed a dynamic knowledge graph-based evidence modeling mechanism for organizing and evolving retrieved information.
• Implemented a reinforcement learning-driven adaptive policy to dynamically adjust retrieval and reasoning strategies.
• Utilized multi-dimensional feedback to balance accuracy, coverage, and efficiency in the RAG process.

Main Results:

• SA-RAG demonstrated significant improvements in accuracy compared to existing methods on multiple benchmarks.
• The framework showed enhanced reasoning consistency and adaptability, particularly in complex MHQA tasks.
• Experimental results validated the effectiveness of the knowledge graph and adaptive policy mechanisms.

Conclusions:

• SA-RAG offers an effective solution for improving LLM performance in Multi-Hop Question Answering.
• The structured and adaptive approach addresses key challenges in evidence organization and strategy selection.
• This framework contributes to building more reliable and generalizable AI question-answering systems.