To Compare, or Not to Compare: On Methodological Practices in Evaluating Social Bias

2026-06-23

AI Summary

Current evaluations of social bias in large language models yield inconsistent conclusions due to methodological fragmentation, primarily stemming from a neglect of the structural design of benchmarks. This work proposes a unified and controlled framework that standardizes heterogeneous bias benchmarks, enabling systematic comparison between isolated demographic assessments and forced-choice comparative setups through controlled experiments. By disentangling the confounding effects of structural factors—such as chain-of-thought reasoning and neutral response options—on bias measurement, the study reveals for the first time that comparative paradigms significantly amplify latent model discrimination, with this effect intensifying as model scale increases. Moreover, even when neutral options are provided or models claim random responding, they exhibit deterministic bias. These findings suggest that while comparative evaluation effectively audits implicit biases, it may not be suitable for real-world scenarios characterized by ambiguity.

Abstract

As Large Language Models are increasingly deployed in critical applications, robustly evaluating their social biases is paramount. However, the current literature suffers from widespread methodological fragmentation, which yields contradictory conclusions. This stems largely from ignoring the structural framing of benchmark-level evaluations. To resolve this, we introduce a unified and controllable framework that standardizes heterogeneous benchmarks to systematically contrast isolated demographic assessments with forced-choice comparative settings. Crucially, this allows us to disentangle the confounding effects of Chain-of-Thought reasoning, neutral fallback options, and other structural artifacts in social bias evaluations. Our evaluation across multiple model families reveals a massive, systematic paradigm gap: while isolated assessments limit prejudice activation, comparative settings act as aggressive catalysts for latent discrimination, a shift primarily driven by underspecified contexts. Alarmingly, CoT reasoning exacerbates social biases under comparative settings, and this systemic bias persists as a deterministic prejudice even when models are provided neutral fallback options or claim to answer randomly. Finally, we demonstrate that this comparative prejudice is a generalized phenomenon that scales positively with model size. Ultimately, we offer a crucial methodological guideline: while researchers must leverage comparative settings to robustly audit hidden biases, practitioners cannot safely rely on comparative deployments in ambiguous real-world tasks.

Problem

• social bias
• methodological fragmentation
• comparative evaluation
• benchmarking
• large language models

Innovation

• comparative evaluation
• social bias
• Chain-of-Thought reasoning
• benchmark standardization
• latent discrimination