What Happened

On June 23, 2026, researchers Federico Marcuzzi, Xuefei Ning, Roy Schwartz, and Iryna Gurevych published a paper on arXiv (2606.24596) titled "To Compare, or Not to Compare: On Methodological Practices in Evaluating Social Bias." The paper introduces a unified, controllable framework designed to standardize the fragmented landscape of LLM social bias benchmarks.

The core problem the researchers identify is structural: current bias evaluation literature suffers from widespread methodological fragmentation that produces contradictory conclusions. Different benchmarks use different framing — some assess demographics in isolation (e.g., asking a model to evaluate a single candidate), while others use forced-choice comparative settings (e.g., asking a model to choose between candidates from different demographic groups). These structural differences are rarely controlled for, leading to inconsistent findings.

The researchers' framework standardizes heterogeneous benchmarks to systematically contrast these two paradigms. Their evaluation across multiple model families revealed what they call a "massive, systematic paradigm gap": isolated assessments tend to limit prejudice activation, while comparative settings act as aggressive catalysts for latent discrimination. This shift is primarily driven by underspecified contexts — when models lack sufficient information to make a clear decision, comparative framing pushes them toward demographic-based reasoning.

Three additional findings are particularly significant:

1. Chain-of-thought reasoning exacerbates bias under comparative settings. This is especially concerning given the growing adoption of CoT prompting in agentic and reasoning-heavy workflows.
2. Bias persists deterministically even with neutral fallback options. Even when models are given the option to decline or answer neutrally, and even when they claim to answer randomly, the systemic bias remains.
3. Comparative prejudice scales positively with model size. Larger models — which enterprises increasingly adopt for performance — carry deeper latent prejudices that surface under comparative conditions.

Why It Matters

This paper exposes a critical blind spot in how the AI industry measures and mitigates social bias. Most production bias audits rely on isolated assessments or single-turn evaluations that, according to this research, may dramatically underreport real-world discrimination. If your compliance team signed off on an LLM deployment based on isolated bias testing, that sign-off may be worth significantly less than assumed.

The chain-of-thought finding is particularly alarming for operators. CoT reasoning has become a default pattern in agentic workflows, tool-use pipelines, and complex reasoning tasks. If CoT actively worsens comparative bias — and this paper provides evidence that it does — then the industry's push toward more reasoning-capable systems may be inadvertently amplifying discriminatory behavior in exactly the contexts where it matters most: comparative decisions like hiring, lending, admissions, and resource allocation.

The scaling result creates a direct tension with enterprise AI strategy. Organizations are migrating to larger models for better performance, but this research suggests that larger models harbor deeper latent biases that only surface under comparative conditions. This means the performance gains from scaling may come with a hidden compliance and reputational cost that current evaluation pipelines don't capture.

Finally, the deterministic persistence of bias — even when neutral fallbacks are provided — undermines a common mitigation strategy. Many practitioners assume that giving models an explicit "I don't know" or neutral option will reduce biased outputs. This paper shows that assumption is wrong.

Who Is Affected

AI startups building products with LLM-driven comparative decisions (candidate ranking, product recommendations, content moderation, loan assessment) face the most immediate risk. Their bias audits may be producing false negatives, and their CoT-based reasoning pipelines may be amplifying the very biases they're trying to mitigate.

Enterprise AI teams deploying frontier models in regulated industries need to reassess their evaluation frameworks. If your bias testing doesn't include comparative settings with CoT reasoning enabled, you're likely missing the worst-case scenarios.

Benchmark creators and open-source evaluators must contend with the paper's call for standardized evaluation framing. The fragmentation the authors describe means that cross-model bias comparisons in the current literature may be comparing apples to oranges.

Strategic Implications

For AI Startup Founders

If your product uses LLMs for any comparative decision — ranking, choosing, filtering, or recommending between options that map to demographic attributes — your current bias testing is likely insufficient. Prioritize building evaluation pipelines that test under comparative settings with CoT enabled, as these represent the worst-case bias scenario. This is not a nice-to-have; it's a compliance and liability issue.

For Developers/Operators Building with AI APIs

Chain-of-thought reasoning is now a default pattern in agentic workflows, but this research shows it actively worsens social bias in comparative contexts. If you're using CoT prompting in production systems that make comparative judgments, implement additional bias safeguards. Consider whether CoT is necessary for your specific use case, or whether simpler prompting patterns might achieve similar results with lower bias risk.

For Non-Technical Business Owners Evaluating AI Tools

When vendors claim their AI is "bias-tested" or "fair," ask specific questions: Did the tests use comparative settings or isolated assessments? Were chain-of-thought reasoning patterns tested? What model size was evaluated? The gap between isolated and comparative bias can be massive, and isolated tests may hide discrimination that surfaces in real-world comparative deployments.

What to Watch Next

Monitor whether major AI labs (OpenAI, Anthropic, Google) respond to these findings by updating their model cards or bias evaluation methodologies. Watch for adoption of the paper's framework by independent benchmark organizations like HELM or MLCommons. If the framework gains traction, expect it to become a standard component of responsible AI audits.

Frequently Asked Questions

Q: What is the difference between isolated and comparative bias evaluation in LLMs?

A: Isolated assessment asks a model to evaluate or respond to a single demographic representation (e.g., rating one candidate's resume). Comparative evaluation asks the model to choose between or rank multiple demographic representations (e.g., selecting between two candidates). The research shows comparative settings activate significantly more latent bias than isolated assessments.

Q: Does chain-of-thought reasoning increase bias in LLMs?

A: According to this research, yes — specifically under comparative settings. Chain-of-thought reasoning was found to exacerbate social biases when models are asked to make comparative decisions between demographic groups. This is particularly relevant for agentic workflows and reasoning-heavy production systems.

Q: Do larger LLMs have more bias than smaller ones?

A: The paper finds that comparative prejudice scales positively with model size, meaning larger models show more latent bias under comparative evaluation conditions. However, this doesn't mean smaller models are unbiased — rather, the bias becomes more pronounced as models scale up.