Refining Answer Distributions for Improved Large Language Model Reasoning

Soumyasundar Pal, Didier Chételat, Yingxue Zhang, Mark Coates

Keywords: Reasoning, Large Language Models

TL;DR: We present Refined Answer Distributions, a novel and principled algorithmic framework to enhance the reasoning capabilities of LLMs.

Abstract: Large Language Models (LLMs) have exhibited an impressive capability to perform reasoning tasks, especially if they are encouraged to generate a sequence of intermediate steps. Reasoning performance can be improved by suitably combining multiple LLM responses, generated either in parallel in a single query, or via sequential interactions with LLMs throughout the reasoning process. Existing strategies for combination, such as self-consistency and progressive-hint-prompting, make inefficient usage of the LLM responses. We present Refined Answer Distributions, a novel and principled algorithmic framework to enhance the reasoning capabilities of LLMs. Our approach can be viewed as an iterative sampling strategy for forming a Monte Carlo approximation of an underlying distribution of answers, with the goal of identifying the mode --- the most likely answer. Empirical evaluation on several reasoning benchmarks demonstrates the superiority of the proposed approach.

Submission Number: 130