Direct-Inverse Prompting: Analyzing LLMs' Discriminative Capacity in Self-Improving Generation

TL;DR

This paper investigates how large language models can use discriminative prompts to reduce output uncertainty and improve their self-generated responses, introducing three prompt strategies and analyzing their effectiveness.

Contribution

It is the first systematic analysis of LLMs' discriminative capacity to address generative uncertainty using direct, inverse, and hybrid prompts.

Findings

Hybrid prompts show the most promise in reducing uncertainty.

Discriminative prompts improve LLM consistency across responses.

Insights guide when to use each discriminative prompt strategy.

Abstract

Mainstream LLM research has primarily focused on enhancing their generative capabilities. However, even the most advanced LLMs experience uncertainty in their outputs, often producing varied results on different runs or when faced with minor changes in input, despite no substantial change in content. Given multiple responses from the same LLM to the same input, we advocate leveraging the LLMs' discriminative capability to reduce this generative uncertainty, aiding in identifying the correct answers. Specifically, we propose and analyze three discriminative prompts: direct, inverse, and hybrid, to explore the potential of both closed-source and open-source LLMs in self-improving their generative performance on two benchmark datasets. Our insights reveal which discriminative prompt is most promising and when to use it. To our knowledge, this is the first work to systematically analyze LLMs' discriminative capacity to address generative uncertainty.