Reproduce “RISE: Reasoning Enhancement via Iterative Self-Exploration in Multi-hop Question Answering” on OpenResearch

About this paper

Bolei He, Xinran He, Mengke Chen, Xianwei Xue, Ying Zhu, Zhenhua Ling

The RISE framework enhances Large Language Models' (LLMs) complex reasoning abilities in multi-hop question answering through an iterative self-exploration process. It enables smaller LLMs, such as LLaMA-3.1-8B, to achieve performance competitive with or superior to larger models like GPT-3.5-turbo on several benchmarks by refining question decomposition, evidence integration, and self-critique.

Key insights

• Integrating a self-iterative paradigm with Retrieval-Augmented Generation effectively allows LLMs to autonomously refine their reasoning, mitigating issues like noisy retrieval and error propagation in multi-hop QA.
• A structured self-critique mechanism is crucial for improving logical consistency and ensuring the relevance of generated sub-questions and answers in complex reasoning paths.
• Joint training across question decomposition, retrieve-then-read, and self-critique tasks leads to synergistic improvements, where each component's learning complements the others.

Summary by alphaXiv.