Explaining Length Bias in LLM-Based Preference Evaluations

TL;DR

This paper investigates the length bias in LLM-based preference evaluations, decomposes the evaluation metric into desirability and information mass, and proposes AdapAlpaca to correct for length effects ensuring fair content quality assessments.

Contribution

It introduces a decomposition of preference metrics into length-independent and length-dependent components and proposes AdapAlpaca to mitigate length bias in evaluations.

Findings

Length impacts evaluation through information mass.

Decomposition clarifies bias sources.

AdapAlpaca improves fairness in preference assessments.

Abstract

The use of large language models (LLMs) as judges, particularly in preference comparisons, has become widespread, but this reveals a notable bias towards longer responses, undermining the reliability of such evaluations. To better understand such bias, we propose to decompose the preference evaluation metric, specifically the win rate, into two key components: desirability and information mass, where the former is length-independent and related to trustworthiness such as correctness, toxicity, and consistency, and the latter is length-dependent and represents the amount of information in the response. We empirically demonstrated the decomposition through controlled experiments and found that response length impacts evaluations by influencing information mass. To derive a reliable evaluation metric that assesses content quality without being confounded by response length, we propose AdapAlpaca, a simple yet effective adjustment to win rate measurement. Specifically, AdapAlpaca ensures a fair comparison of response quality by aligning the lengths of reference and test model responses under equivalent length intervals.