Self-prompted Chain-of-Thought on Large Language Models for Open-domain Multi-hop Reasoning

• DOI: 10.48550/arxiv.2310.13552
• Publication Date: 2023-01-01
• Authors: Jin‐Yuan Wang; Junlong Li; Hai Zhao
• Keywords: FOS: Computer and information sciences; Computer Science - Computation and Language; Artificial Intelligence (cs.AI); Computer Science - Artificial Intelligence; Computation and Language (cs.CL)

In open-domain question-answering (ODQA), most existing questions require single-hop reasoning on commonsense. To further extend this task, we officially introduce multi-hop (ODMR) by answering with explicit steps in setting. Recently, large language models (LLMs) have found significant utility facilitating ODQA without external corpus. Furthermore, chain-of-thought (CoT) prompting boosts the capability of LLMs to a greater extent manual or automated paradigms. However, methods lack quality assurance, while approaches suffer from limited scalability and poor diversity, hindering capabilities LLMs. paper, propose Self-prompted Chain-of-Thought (SP-CoT), an framework mass-produce high CoTs LLMs, for SP-CoT introduces generation pipeline ODMR datasets, adaptive sampler in-context CoT selection self-prompted inference via learning. Extensive experiments four benchmarks show that our proposed not only significantly surpasses previous SOTA large-scale (175B) but also nearly doubles zero-shot performance small-scale (13B) Further analysis reveals remarkable elicit direct concise intermediate recalling $\sim$50\% answers MuSiQue-Ans dataset.