Compare without Despair: Reliable Preference Evaluation with Generation Separability

TL;DR

This paper introduces a measure called separability to improve the consistency and reliability of preference evaluations of language models, especially in challenging scenarios with similar or variable outputs.

Contribution

The paper proposes a novel meta-evaluation measure, separability, to assess and enhance the reliability of preference judgments in language model evaluation.

Findings

High separability instances yield more consistent ratings.

Separability distribution provides insights into benchmark quality.

Incorporating separability improves LLM ranking accuracy.

Abstract

Human evaluation of generated language through pairwise preference judgments is pervasive. However, under common scenarios, such as when generations from a model pair are very similar, or when stochastic decoding results in large variations in generations, it results in inconsistent preference ratings. We address these challenges by introducing a meta-evaluation measure, separability, which estimates how suitable a test instance is for pairwise preference evaluation. For a candidate test instance, separability samples multiple generations from a pair of models, and measures how distinguishable the two sets of generations are. Our experiments show that instances with high separability values yield more consistent preference ratings from both human- and auto-raters. Further, the distribution of separability allows insights into which test benchmarks are more valuable for comparing models. Finally, we incorporate separability into ELO ratings, accounting for how suitable each test instance might be for reliably ranking LLMs. Overall, separability has implications for consistent, efficient and robust preference evaluation of LLMs with both human- and auto-raters.