The Unreliability of Explanations in Few-shot Prompting for Textual Reasoning

TL;DR

This paper investigates whether explanations improve large language model reasoning, finding limited benefits from explanations, highlighting their unreliability, but also proposing methods to assess and improve explanation quality post-hoc.

Contribution

The study systematically evaluates the impact of explanations in prompting LLMs for reasoning tasks and introduces a calibration method to assess explanation reliability.

Findings

Explanations yield small to moderate accuracy improvements.

Generated explanations may not be factually grounded or entail predictions.

Calibrators can improve performance by assessing explanation reliability.

Abstract

Does prompting a large language model (LLM) like GPT-3 with explanations improve in-context learning? We study this question on two NLP tasks that involve reasoning over text, namely question answering and natural language inference. We test the performance of four LLMs on three textual reasoning datasets using prompts that include explanations in multiple different styles. For these tasks, we find that including explanations in the prompts for OPT, GPT-3 (davinci), and InstructGPT (text-davinci-001) only yields small to moderate accuracy improvements over standard few-show learning. However, text-davinci-002 is able to benefit more substantially. We further show that explanations generated by the LLMs may not entail the models' predictions nor be factually grounded in the input, even on simple tasks with extractive explanations. However, these flawed explanations can still be useful as a way to verify LLMs' predictions post-hoc. Through analysis in our three settings, we show that explanations judged by humans to be good--logically consistent with the input and the prediction--more likely cooccur with accurate predictions. Following these observations, we train calibrators using automatically extracted scores that assess the reliability of explanations, allowing us to improve performance post-hoc across all of our datasets.