Self-Evaluation Improves Selective Generation in Large Language Models

TL;DR

This paper introduces a self-evaluation method for large language models that reformulates generation tasks into token-level predictions, improving the models' ability to assess and selectively generate content more reliably.

Contribution

It proposes a novel token-level self-evaluation approach that leverages LLMs' calibration, enhancing content quality assessment and selective generation capabilities.

Findings

Self-evaluation scores improve accuracy in content assessment.

Self-evaluation correlates better with overall content quality.

Method outperforms likelihood-based metrics in selective generation.

Abstract

Safe deployment of large language models (LLMs) may benefit from a reliable method for assessing their generated content to determine when to abstain or to selectively generate. While likelihood-based metrics such as perplexity are widely employed, recent research has demonstrated the limitations of using sequence-level probability estimates given by LLMs as reliable indicators of generation quality. Conversely, LLMs have demonstrated strong calibration at the token level, particularly when it comes to choosing correct answers in multiple-choice questions or evaluating true/false statements. In this work, we reformulate open-ended generation tasks into token-level prediction tasks, and leverage LLMs' superior calibration at the token level. We instruct an LLM to self-evaluate its answers, employing either a multi-way comparison or a point-wise evaluation approach, with the option to include a ``None of the above'' option to express the model's uncertainty explicitly. We benchmark a range of scoring methods based on self-evaluation and evaluate their performance in selective generation using TruthfulQA and TL;DR. Through experiments with PaLM-2 and GPT-3, we demonstrate that self-evaluation based scores not only improve accuracy, but also correlate better with the overall quality of generated content.