mleyvaz/geo-bias-llm

Geo Bias LLM — Institutional Prestige as Geographic Bias in Large Language Models

Experimental data for the paper "Institutional Prestige as Geographic Bias in Large Language Models: Evidence from Three Factorial Experiments with Bootstrap Confidence Intervals" (Leyva-Vázquez, 2026).

• Paper page: https://huggingface.co/papers/2608.18107
• Code: https://github.com/mleyvaz/geo-bias-llm
• Author: Maikel Leyva-Vázquez — Universidad Bolivariana del Ecuador / Universidad de Guayaquil

What the experiments measure

Four LLMs (Claude Haiku 4.5, GPT-4o-mini, Gemini 2.0 Flash, Llama 3.1 8B) score otherwise identical applicant profiles on a 0–10 scale across five decision domains (scholarship, hiring, credit, health, public policy). The profiles vary only in signals that should be irrelevant to most of those decisions:

• Name origin — Anglo, Latino, Arabic
• Institutional affiliation — from Tier 1 (MIT, Columbia) to Tier 5 (Universidad de Guayaquil)
• Country development status and journal prestige (Study 3)

The headline result is that name origin moves scores very little, while institutional affiliation moves them consistently — including in domains where institutional prestige carries no information, such as credit scoring and clinical health experience.

Scores are also reported as neutrosophic bias triplets (T, I, F), where T is the degree of favourable judgement, I the indeterminacy, and F the degree of unfavourable judgement.

Configurations

Config	Rows	Granularity	Contents
study3_raw (default)	1,440	one row per API call	Institution (MIT/UGye) × Journal (Nature/NCML), per model, domain and repetition
study1_pilot	80	model × profile × domain	2×2: Name (Anglo/Latino) × Institution (Tier 1/Tier 5)
study1_full	240	model × profile × domain	3×4: Name (Anglo/Latino/Arabic) × Tier (T1 MIT / T2 UChile / T3 UNAL / T5 UGye)
study2	240	model × profile × domain	3×4 recast as 2×2 prestige × country development
bootstrap_ci	12	contrast	Observed effects with bootstrap 95% confidence intervals

Only study3_raw is call-level. The other configurations are the aggregated statistics reported in the paper (means, standard deviations, gaps and neutrosophic triplets); the per-call responses behind them were not retained.

from datasets import load_dataset

calls = load_dataset("mleyvaz/geo-bias-llm", "study3_raw", split="train")
pilot = load_dataset("mleyvaz/geo-bias-llm", "study1_pilot", split="train")

Fields

study3_raw — model, cell, inst_label, journal_label, inst_prestige, journal_prestige, name_origin, domain, rep, score.

study1_* and study2 — model, profile, domain, name_group, inst_tier, mean_score, std, gap, NBI_T, NBI_I, NBI_F.

bootstrap_ci — contrast, subgroup, obs, ci95_lo, ci95_hi, n_a, n_b. Contrasts prefixed s1_ belong to Study 1 and s2_ to Study 2.

The original, unflattened JSON files from the repository are preserved verbatim with the raw_ prefix.

Provenance and limitations

All model responses were collected through OpenRouter in May 2026. Model versions are those served at that date and are not pinned by hash, so exact regeneration is not guaranteed — the released scores are the record of what was observed. The stimuli are synthetic profiles written for the experiment; no personal data of real individuals is involved.

Scores are single-rater LLM outputs on a coarse 0–10 scale, and the five domains are represented by a limited set of items, so the estimates are informative about the direction and rough size of the effects rather than their precise magnitude. Bootstrap confidence intervals for the main contrasts are provided in bootstrap_ci for that reason.

Citation

@article{leyvavazquez2026geobias,
  title   = {Institutional Prestige as Geographic Bias in Large Language Models:
             Evidence from Three Factorial Experiments with Bootstrap Confidence Intervals},
  author  = {Leyva-V{\'a}zquez, Maikel},
  year    = {2026},
  eprint  = {2608.18107},
  archivePrefix = {arXiv}
}

Released under the MIT license, matching the code repository.