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

Federico Marcuzzi; Iryna Gurevych; Roy Schwartz; Xuefei Ning

REVIEW 3 major objections 2 minor 12 references

Comparative settings in bias benchmarks trigger far more social discrimination in language models than isolated assessments.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.3

2026-06-25 23:53 UTC pith:S5UME6L3

load-bearing objection The unified framework for pitting isolated vs. comparative bias tests is a practical step forward, but the standardization process risks confounding the main contrast and the abstract gives too little on methods to judge the claims. the 3 major comments →

arxiv 2606.24596 v1 pith:S5UME6L3 submitted 2026-06-23 cs.CL

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

Federico Marcuzzi , Xuefei Ning , Roy Schwartz , Iryna Gurevych This is my paper

classification cs.CL

keywords social bias evaluationlarge language modelscomparative settingsbenchmark standardizationchain-of-thought reasoningdiscrimination activationmethodological practicesmodel scale effects

verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces a unified controllable framework that standardizes existing benchmarks to directly contrast isolated demographic questions with forced-choice comparative ones. It shows that comparative setups sharply increase the activation of latent social biases, driven mainly by underspecified contexts, while isolated assessments keep prejudice low. Chain-of-thought reasoning worsens the effect in comparative cases, and the bias remains deterministic even when neutral options are supplied or models claim random answers. The comparative prejudice grows with model size across families. This matters because fragmented evaluation methods currently produce contradictory claims about model fairness.

Core claim

While isolated assessments limit prejudice activation, comparative settings act as aggressive catalysts for latent discrimination, a shift primarily driven by underspecified contexts. 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. This comparative prejudice is a generalized phenomenon that scales positively with model size.

What carries the argument

A unified and controllable framework that standardizes heterogeneous benchmarks to contrast isolated demographic assessments with forced-choice comparative settings.

Load-bearing premise

Standardizing heterogeneous benchmarks into a single controllable framework successfully isolates the effects of comparative versus isolated settings without introducing new structural artifacts or selection biases.

What would settle it

Re-running the standardized benchmarks on the same models and finding no reliable increase in bias scores when switching from isolated to comparative prompts would falsify the central claim.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

Researchers must use comparative settings to audit hidden biases that isolated tests miss.
Practitioners cannot safely rely on comparative deployments for ambiguous real-world tasks.
Chain-of-thought prompting should be avoided in comparative bias audits because it increases prejudice expression.
Bias evaluations must control for structural framing or risk contradictory conclusions across studies.
The observed prejudice scales with model size, so larger models require stricter comparative testing.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Models may internally perform similar comparisons even on single-group queries in real deployments, suggesting the need to test for implicit comparative reasoning.
The framework could be applied to non-social bias domains such as factual consistency or safety to check if comparative framing has analogous effects.
If underspecified contexts drive the effect, adding explicit neutral context might reduce the comparative prejudice without changing the evaluation structure.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Desk Editor's Note

The unified framework for pitting isolated vs. comparative bias tests is a practical step forward, but the standardization process risks confounding the main contrast and the abstract gives too little on methods to judge the claims.

read the letter

The paper's core contribution is a single controllable setup that converts existing bias benchmarks into matched isolated and forced-choice versions so the effect of comparison can be measured directly. That addresses the fragmentation the authors describe and lets them test interactions with CoT and neutral options in one place.

The new piece is the direct head-to-head on the same items and the finding that comparative framing pulls out more bias, especially in underspecified cases, with the effect growing in larger models. The guideline that comparative tests are better for auditing but risky for deployment follows from those results.

The work is useful for anyone who runs or designs bias evaluations; it gives a concrete way to make results more comparable across papers. The authors also test persistence under neutral fallbacks and random-claim conditions, which is a reasonable check.

The main soft spot is the standardization step itself. Converting open-ended items to paired choices can change prompt length, option order, or implicit framing, and nothing in the abstract shows those changes were neutral with respect to bias activation. If the observed gap partly tracks those artifacts rather than comparison per se, the central claim weakens. The abstract also states results across model families without methods, sample sizes, or error bars, so the strength of the evidence is hard to assess from what is given.

This is the sort of methodological paper that belongs in a reading group focused on evaluation practices. It deserves a serious referee because the framing question matters for real deployments and the framework idea is worth testing, even if the current version needs tighter controls and more transparent implementation details before the conclusions can be taken as settled.

Referee Report

3 major / 2 minor

Summary. The paper introduces a unified controllable framework that standardizes heterogeneous social-bias benchmarks to contrast isolated demographic assessments against forced-choice comparative settings. It reports a large paradigm gap in which comparative settings catalyze latent discrimination (primarily via underspecified contexts), CoT reasoning exacerbates bias under comparison, the effect persists even with neutral fallback options or random-answer claims, and the prejudice scales positively with model size. The authors conclude that comparative settings are necessary for robust auditing but unsafe for ambiguous real-world deployment.

Significance. If the central empirical contrast holds after controls for standardization artifacts, the work would supply a concrete methodological diagnosis for contradictory findings in the bias-evaluation literature and a practical guideline distinguishing auditing from deployment. The scaling result and the persistence under neutral options would be particularly useful for model developers and benchmark designers.

major comments (3)

[Framework] Framework section: the claim that standardization successfully isolates the comparative vs. isolated contrast rests on the untested assumption that converting open-ended items to paired choices preserves semantics, prompt length, and implicit framing. No ablation or control experiment is described that measures bias activation on the transformed items before the comparative manipulation is applied.
[Experiments / Results] Results across model families: the abstract states findings hold across families yet supplies no table or appendix listing the exact models, number of runs, statistical tests, or error bars. Without these, the reported paradigm gap and scaling trend cannot be assessed for robustness.
[Analysis] Analysis of underspecified contexts: the assertion that the shift is 'primarily driven by underspecified contexts' requires a quantitative decomposition (e.g., a table contrasting bias scores on high- vs. low-specification prompts within the same comparative setting). No such breakdown is referenced.

minor comments (2)

[Abstract] The abstract uses 'massive' and 'aggressive catalysts' without accompanying effect-size numbers; replace with quantitative descriptors once the results tables are added.
[Experimental setup] Clarify whether the neutral-fallback and random-answer conditions were run on the same prompt templates or required additional prompt engineering; this affects reproducibility.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their thoughtful review and constructive suggestions. We believe the suggested additions will enhance the clarity and robustness of our findings. Below we respond to each major comment.

read point-by-point responses

Referee: [Framework] Framework section: the claim that standardization successfully isolates the comparative vs. isolated contrast rests on the untested assumption that converting open-ended items to paired choices preserves semantics, prompt length, and implicit framing. No ablation or control experiment is described that measures bias activation on the transformed items before the comparative manipulation is applied.

Authors: We agree that an explicit control experiment would provide stronger evidence that the standardization process preserves the relevant properties. In the revised manuscript, we will include an ablation study that measures bias activation on the transformed paired-choice items (before the comparative manipulation is applied) and compare it to the original open-ended versions to verify semantic preservation. revision: yes
Referee: [Experiments / Results] Results across model families: the abstract states findings hold across families yet supplies no table or appendix listing the exact models, number of runs, statistical tests, or error bars. Without these, the reported paradigm gap and scaling trend cannot be assessed for robustness.

Authors: We will add a detailed table in the appendix listing all models evaluated, the number of runs per experiment, the statistical tests applied, and error bars for key metrics to allow assessment of robustness. revision: yes
Referee: [Analysis] Analysis of underspecified contexts: the assertion that the shift is 'primarily driven by underspecified contexts' requires a quantitative decomposition (e.g., a table contrasting bias scores on high- vs. low-specification prompts within the same comparative setting). No such breakdown is referenced.

Authors: We will include a quantitative decomposition in the revised manuscript, specifically a table or figure that contrasts bias scores on high- versus low-specification prompts within the comparative setting to substantiate the claim regarding the primary driver of the paradigm shift. revision: yes

Circularity Check

0 steps flagged

No circularity; empirical standardization framework is self-contained

full rationale

The paper introduces a controllable standardization framework to contrast isolated vs. comparative bias evaluations and reports experimental results across model families. No equations, derivations, fitted parameters, or self-citation chains appear in the provided text. The central claims rest on observed paradigm gaps in the experiments rather than reducing by construction to the framework definition or prior self-citations. The standardization procedure is presented as an independent methodological tool whose validity is tested via the reported outcomes, not presupposed.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review supplies no information on free parameters, axioms, or invented entities.

pith-pipeline@v0.9.1-grok · 5753 in / 1024 out tokens · 16442 ms · 2026-06-25T23:53:20.083200+00:00 · methodology

0 comments

read the original 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.

Figures

Figures reproduced from arXiv: 2606.24596 by Federico Marcuzzi, Iryna Gurevych, Roy Schwartz, Xuefei Ning.

**Figure 1.** Figure 1: (a) Example of the iso and cmp paradigms. (b) Example of the effect of CoT reasoning: the model will do its best to answer the user’s prompt. (c) Example of a model claims to answer randomly, but does not. critical but often implicit evaluation factor: question framing, i.e., whether bias is measured by evaluating demographic groups independently or by directly comparing competing groups or stereotypica… view at source ↗

discussion (0)

Reference graph

Works this paper leans on

12 extracted references · 3 canonical work pages · 1 internal anchor

[1]

InThe Eleventh International Confer- enceonLearningRepresentations,ICLR2023,Kigali, Rwanda, May 1-5, 2023

Quantifying memorization across neural lan- guage models. InThe Eleventh International Confer- enceonLearningRepresentations,ICLR2023,Kigali, Rwanda, May 1-5, 2023. OpenReview.net. JavierCoronado-Blázquez.2025. Deterministicorprob- abilistic? the psychology of llms as random number generators.CoRR, abs/2502.19965. Jwala Dhamala, Tony Sun, Varun Kumar, Sat...

work page arXiv 2023
[2]

In9th International Conference on Learn- ing Representations, ICLR 2021, Virtual Event, Aus- tria, May 3-7, 2021

Measuring massive multitask language under- standing. In9th International Conference on Learn- ing Representations, ICLR 2021, Virtual Event, Aus- tria, May 3-7, 2021. OpenReview.net. Mahammed Kamruzzaman and Gene Louis Kim. 2025. Prompting techniques for reducing social bias in llms through system 1 and system 2 cognitive processes. InProceedings of the ...

2021
[3]

OpenAI GPT-5 System Card

Association for Computational Linguistics. Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yu- taka Matsuo, and Yusuke Iwasawa. 2022. Large lan- guage models are zero-shot reasoners. InAdvances in Neural Information Processing Systems 35: An- nual Conference on Neural Information Processing Systems2022, NeurIPS2022, NewOrleans, LA,USA, November 28 - Decem...

work page internal anchor Pith review Pith/arXiv arXiv 2022
[4]

unrelated

Association for Computational Linguistics. Yarden Tal, Inbal Magar, and Roy Schwartz. 2022. Fewer errors, but more stereotypes? the effect of model size on gender bias. InProceedings of the 4th Workshop on Gender Bias in Natural Language Processing(GeBNLP),pages112–120,Seattle,Wash- ington. Association for Computational Linguistics. Alex Tamkin, Amanda As...

work page arXiv 2022
[5]

You are an expert in question-answering tasks
[6]

instructions:[1] Please provide your final answer as one of the following: \boxed{X} or \boxed{Y}

You are an expert in answering multiple-choice questions. instructions:[1] Please provide your final answer as one of the following: \boxed{X} or \boxed{Y}
[7]

Your final answer should be one of these: \boxed{X} or \boxed{Y}
[8]

cot footers:[1] Please reason step by step and then provide your final answer inside a \boxed{}

Provide your final answer using the following format exactly: \boxed{X} or \boxed{Y}. cot footers:[1] Please reason step by step and then provide your final answer inside a \boxed{}
[9]

Provide a step-by-step reasoning before giving your final answer in a \boxed{}
[10]

no-cot footers:[1] Provide your final answer immediately at the end of this prompt, inside \boxed{}

Think through the question carefully step by step, then give your final answer in a \boxed{}. no-cot footers:[1] Provide your final answer immediately at the end of this prompt, inside \boxed{}. My answer would be: \boxed{
[11]

My answer is: \boxed{

State your final answer directly at the end of this prompt within a \boxed{}. My answer is: \boxed{
[12]

A”/“B” or “Yes

Without any additional reasoning, provide your final answer inside \boxed{} at the end of this prompt. The final answer is: \boxed{ Prompt 14: Exhaustive list of prompt component variations (headers, instructions, and footers) used for the prompt variation analysis. Placeholders\boxed{X}and \boxed{Y}are dynamically replaced with the task-specific options ...

[1] [1]

InThe Eleventh International Confer- enceonLearningRepresentations,ICLR2023,Kigali, Rwanda, May 1-5, 2023

Quantifying memorization across neural lan- guage models. InThe Eleventh International Confer- enceonLearningRepresentations,ICLR2023,Kigali, Rwanda, May 1-5, 2023. OpenReview.net. JavierCoronado-Blázquez.2025. Deterministicorprob- abilistic? the psychology of llms as random number generators.CoRR, abs/2502.19965. Jwala Dhamala, Tony Sun, Varun Kumar, Sat...

work page arXiv 2023

[2] [2]

In9th International Conference on Learn- ing Representations, ICLR 2021, Virtual Event, Aus- tria, May 3-7, 2021

Measuring massive multitask language under- standing. In9th International Conference on Learn- ing Representations, ICLR 2021, Virtual Event, Aus- tria, May 3-7, 2021. OpenReview.net. Mahammed Kamruzzaman and Gene Louis Kim. 2025. Prompting techniques for reducing social bias in llms through system 1 and system 2 cognitive processes. InProceedings of the ...

2021

[3] [3]

OpenAI GPT-5 System Card

Association for Computational Linguistics. Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yu- taka Matsuo, and Yusuke Iwasawa. 2022. Large lan- guage models are zero-shot reasoners. InAdvances in Neural Information Processing Systems 35: An- nual Conference on Neural Information Processing Systems2022, NeurIPS2022, NewOrleans, LA,USA, November 28 - Decem...

work page internal anchor Pith review Pith/arXiv arXiv 2022

[4] [4]

unrelated

Association for Computational Linguistics. Yarden Tal, Inbal Magar, and Roy Schwartz. 2022. Fewer errors, but more stereotypes? the effect of model size on gender bias. InProceedings of the 4th Workshop on Gender Bias in Natural Language Processing(GeBNLP),pages112–120,Seattle,Wash- ington. Association for Computational Linguistics. Alex Tamkin, Amanda As...

work page arXiv 2022

[5] [5]

You are an expert in question-answering tasks

[6] [6]

instructions:[1] Please provide your final answer as one of the following: \boxed{X} or \boxed{Y}

You are an expert in answering multiple-choice questions. instructions:[1] Please provide your final answer as one of the following: \boxed{X} or \boxed{Y}

[7] [7]

Your final answer should be one of these: \boxed{X} or \boxed{Y}

[8] [8]

cot footers:[1] Please reason step by step and then provide your final answer inside a \boxed{}

Provide your final answer using the following format exactly: \boxed{X} or \boxed{Y}. cot footers:[1] Please reason step by step and then provide your final answer inside a \boxed{}

[9] [9]

Provide a step-by-step reasoning before giving your final answer in a \boxed{}

[10] [10]

no-cot footers:[1] Provide your final answer immediately at the end of this prompt, inside \boxed{}

Think through the question carefully step by step, then give your final answer in a \boxed{}. no-cot footers:[1] Provide your final answer immediately at the end of this prompt, inside \boxed{}. My answer would be: \boxed{

[11] [11]

My answer is: \boxed{

State your final answer directly at the end of this prompt within a \boxed{}. My answer is: \boxed{

[12] [12]

A”/“B” or “Yes

Without any additional reasoning, provide your final answer inside \boxed{} at the end of this prompt. The final answer is: \boxed{ Prompt 14: Exhaustive list of prompt component variations (headers, instructions, and footers) used for the prompt variation analysis. Placeholders\boxed{X}and \boxed{Y}are dynamically replaced with the task-specific options ...