In the present era, the proliferation and widespread adoption of Large Language Models (LLMs) have substantially elevated the capacity for understanding and reasoning in various textual tasks. Yet, there is a current shortage of effective methodologies for tackling Deliberative Questions, which require thoughtful consideration and may encompass evolving scenarios. To bridge this gap, this paper introduces a systematic approach that employs relevant impact analysis to build a cognitive framework, whose results guide LLMs to assist generate hypotheses and testing, known as Abductive Reasoning, for addressing deliberative questions, with the goal of fostering a deliberative thinking mindset. Our comprehensive evaluation on deliberative questions gathered from Reddit and Zhihu showed that pipeline generates superior answers in 65.5% and 73.6% of the cases on our datasets in ChatGPT and QwenMAX view, particularly excelling with an 88.3% dominance rate in addressing reasoning-type questions and and 84.4% in predictive-type questions.

错误:搜索内容不能为空,请输入英文关键词
错误:关键词超出字数限制,请精简
高级检索

LLM Assists Hypothesis Generation and Testing for Deliberative Questions

  • Fuchun Wang,
  • Xian Zhou,
  • Wenpeng Hu,
  • Zhunchen Luo,
  • Wei Luo,
  • Xiaoying Bai

摘要

In the present era, the proliferation and widespread adoption of Large Language Models (LLMs) have substantially elevated the capacity for understanding and reasoning in various textual tasks. Yet, there is a current shortage of effective methodologies for tackling Deliberative Questions, which require thoughtful consideration and may encompass evolving scenarios. To bridge this gap, this paper introduces a systematic approach that employs relevant impact analysis to build a cognitive framework, whose results guide LLMs to assist generate hypotheses and testing, known as Abductive Reasoning, for addressing deliberative questions, with the goal of fostering a deliberative thinking mindset. Our comprehensive evaluation on deliberative questions gathered from Reddit and Zhihu showed that pipeline generates superior answers in 65.5% and 73.6% of the cases on our datasets in ChatGPT and QwenMAX view, particularly excelling with an 88.3% dominance rate in addressing reasoning-type questions and and 84.4% in predictive-type questions.