Mario Sanz-Guerrero

Research profile

Research summary

Mario Sanz-Guerrero is an early-career natural language processing researcher at Johannes Gutenberg University Mainz whose work centers on how large language models should be evaluated and how reliably they behave across languages and domains. The OpenAlex-linked record (identity resolved via ORCID 0009-0001-3581-8369) shows a sharply concentrated profile: the field distribution is 100% natural language processing, and topic frequencies are led by topic modeling (18) and NLP techniques (15), with adversarial robustness (9), domain adaptation and few-shot learning (8), and software engineering research (3) rounding out the mix. Activity is heavily concentrated in the last two recorded years — 13 works in 2025 and 9 in 2026, 22 works in the trailing three-year window — while citation-based indicators remain nascent (h-index 1, i10-index 0). Two Evolve-reviewed papers anchor the qualitative picture. A first-authored 2026 position paper argues that calibration should sit beside every NLP benchmark metric — per Evolve, a persuasive adoption-gap case that names no default metric and adds no new evidence. A second 2026 study collects stakeholder perspectives on cross-lingual consistency of LLMs for medical questions, offering useful first evidence tempered by thin statistics and persona averaging. Frequent co-authors include Katharina von der Wense, Minh Duc Bui, Abteen Ebrahimi, Sagi Shaier, Peter Herbert Kann, and Manuel Mager; with only 6 distinct co-authors on record, the network is compact, and neither reviewed paper lists him as corresponding author.

Research keywords

Highlights

• Position paper makes the case for calibration as a first-class LLM evaluation criterion

  In a first-authored 2026 arXiv position paper, Sanz-Guerrero argues that calibration — spanning expected calibration error and verbalized confidence — should sit beside every NLP benchmark metric, alongside LLM-as-a-judge practice. Evolve finds the adoption-gap case persuasive but notes it names no default metric and adds no new evidence; reproducibility readiness is rated low.

• First stakeholder evidence on cross-lingual consistency of LLMs for medical questions

  A 2026 co-authored arXiv study gathers stakeholder perspectives on whether multilingual LLMs answer medical questions consistently across languages, touching cultural adaptation and persona simulation. Evolve calls it useful first evidence — with stakeholders split — while flagging thin statistics and persona averaging as limits; reproducibility readiness is medium.

• Sharply focused NLP agenda with a two-year output ramp

  The OpenAlex record assigns 100% of output to natural language processing, led by topic modeling (18) and NLP techniques (15), plus adversarial robustness (9) and domain adaptation / few-shot learning (8). Counts by year jump to 13 works in 2025 and 9 in 2026 so far — 22 works in the trailing three years — while citation indicators (h-index 1, i10-index 0) remain nascent.

Timeline

1. 202513 works recorded in OpenAlex — the first full year in the record and the start of a steep activity ramp.
2. 20269 works recorded so far in 2026; the first-authored position paper 'Calibration as a First-Class Criterion in LLM Evaluation' appears on arXiv.
3. 2026Co-authors 'Perspectives on Cross-Lingual Consistency in LLMs for Medical Questions' on arXiv, in a middle authorship position.
4. 2026Trailing three-year output reaches 22 works, while citation indicators remain nascent (h-index 1, i10-index 0).

Papers in the Evolve catalog

• 2026-09-22first author

  Calibration as a First-Class Criterion in LLM Evaluation

• 2026-09-07

  Perspectives on Cross-Lingual Consistency in LLMs for Medical Questions

Frequent co-authors