Yusser Al Ghussin

German Research Centre for Artificial Intelligence · Germany

Unclaimed · AI-structured profileOpenAlex

DFKI researcher auditing pretraining corpora for cultural coverage and tracing stereotype bias in multilingual LLMs

papers here

reviewed by Evolve

first-author

corresponding

—

works · OpenAlex

h-index · OpenAlex

Bibliometric figures come from OpenAlex (2026-10-16) and cover the author's entire public record, not just the papers listed here. Reproducibility signal comes from Evolve's own reviews.

Research profile

Output—

The output axis is not yet computed: publication-volume and venue-quality scoring is unavailable for this profile. It would measure aggregate production — works per year and venue tiers. Only raw per-year counts (2023–2026) exist, so no score is given here.

Impact—

The impact axis is not yet computed and no citation-based score is available. It would measure citation reach — h-index, i10-index and per-paper citations. Raw indicators exist (h-index 1, i10-index 0) but are too early-stage, so the axis remains unscored.

Focus0/100

Focus scores 0, the lowest concentration: the field distribution is split exactly 50/50 between corpus auditing / cultural evaluation resources and mechanistic interpretability of multilingual LLMs. On a single-field axis this reads as low concentration, yet both halves sit inside NLP and share methodological threads visible in the topic profile — topic modeling and NLP techniques (22 mentions each) and explainable AI (10). The picture is a deliberate two-track agenda rather than scattered output.

Collaboration44/100

Collaboration scores 44, a moderate level: the record lists 8 distinct coauthors. Repeated collaborators are concentrated at DFKI — Cristina España-Bonet, Josef van Genabith, Simon Ostermann, Eva Gavaller, Koel Dutta Chowdhury, Ariun-Erdene Tumurchuluun and Pinzhen Chen. The reviewed papers show both modes: first author leading the corpus-auditing resource, middle author on the interpretability study — consistent with a small, stable in-house team rather than a wide external network.

Activity91/100

Activity scores 91, the strongest axis in this profile. Counts by year rise from 3 works in 2023 and 1 in 2024 to 8 in 2025 and 13 in 2026, with 22 works in the most recent three-year window — a steep acceleration across 2025–2026. Topic volume backs this up: 22 mentions each for topic modeling and NLP techniques, 10 for explainable AI.

Reproducibility55/100

Reproducibility averages 55 across the 2 reviewed papers, both rated medium readiness. The first-author FineWeb-CLaR resource is candid about validity limits — 22% human agreement on culture labels — which supports transparent reuse; the stereotype-tracing study is careful, though its cross-SAE comparison is sparsity-confounded. A mid-range score fits: valuable, documented artifacts with disclosed caveats rather than turnkey replication.

Research summary

Yusser Al Ghussin is a researcher at the German Research Centre for Artificial Intelligence (DFKI) working at the intersection of corpus auditing and mechanistic interpretability in natural language processing. The OpenAlex-derived record shows 25 works between 2023 and 2026 — 3 in 2023, 1 in 2024, 8 in 2025 and 13 in 2026 — with 22 works in the last three years, driving an activity score of 91. Research output splits evenly (50/50) between cultural corpus-auditing resources and mechanistic interpretability of multilingual LLMs, which yields a focus score of 0 on the concentration axis even though both lines share NLP methodology: topic modeling, NLP techniques and explainable AI recur across the topic profile. Two Evolve-reviewed 2026 papers anchor the profile: FineWeb-CLaR, a first-author culture-language-region annotation of 30.9B FineWeb documents for benchmark-aligned corpus auditing, candid about weak culture-label validity (22% human agreement), and a study tracing stereotype features from representation to output in multilingual LLMs. Reproducibility readiness averages 55 across the two papers (both rated medium). Citation indicators remain early-stage (h-index 1, i10-index 0), and the output and impact axis scores are not yet computed. Collaboration is moderate (score 44; 8 distinct coauthors), centered on DFKI colleagues.

Collaboration pattern

Collaboration is moderate (score 44) with 8 distinct coauthors on record. The recurring circle is DFKI-based: Cristina España-Bonet, Josef van Genabith, Simon Ostermann, Eva Gavaller, Koel Dutta Chowdhury, Ariun-Erdene Tumurchuluun and Pinzhen Chen. The reviewed papers span both roles — first author of the corpus-auditing resource, middle author of the interpretability study — suggesting work inside a stable in-house NLP group rather than a broad external network.

Research keywords

cultural NLPpretraining corpus auditingFineWebregion attributionURL geolocationtopic modelingmultilingual LLMsstereotype biasmechanistic interpretabilitylinear probingattribution patchingsparse autoencoders

Highlights

First-author annotation of 30.9B FineWeb documents for culture, language and region auditing
FineWeb-CLaR adds culture, language and region annotations to 30.9 billion FineWeb documents, aligned with benchmarks for corpus auditing. Evolve's review calls it a valuable auditing resource with exploratory analyses, built on URL geolocation and topic modeling.
Candid validity reporting on culture labels (22% human agreement)
The same paper openly reports weak validity of its culture (CT) labels — 22% human agreement — rather than overstating coverage. This honesty about label quality shapes how downstream users should treat the resource.
Tracing stereotype features from representation to output in multilingual LLMs
A 2026 mechanistic-interpretability study shows that decodability, output influence and cross-lingual transfer of stereotype features dissociate. Evolve notes it is a careful study, while flagging that the cross-SAE comparison is sparsity-confounded.

Timeline

2023First works appear in the OpenAlex record: 3 publications in 2023.
2024Output dips to a single publication in 2024, the quietest year on record.
2025Output accelerates to 8 publications, beginning the current high-output phase.
2026Most productive year: 13 publications, including both Evolve-reviewed arXiv papers — first-author FineWeb-CLaR and the multilingual stereotype-tracing study.

Papers in the Evolve catalog

2026-09-21first author
FineWeb-CLaR: Culture, Language, and Region Annotations for Benchmark-Aligned Corpus Auditing
2026-09-08
Tracing Stereotypes from Representation to Output in Multilingual LLMs

Frequent co-authors

Eva Gavaller×1 Cristina España-Bonet×1 Josef van Genabith×1 Simon Ostermann×1 Ariun-Erdene Tumurchuluun×1 Pinzhen Chen×1 Josef van Genabith×1 Koel Dutta Chowdhury×1

About this profile

This profile is AI-structured by Evolve.Pub from public bibliographic data (OpenAlex) and Evolve's own paper reviews. It is not written or endorsed by the researcher unless the profile is claimed. Radar dimensions are computed from real bibliometrics with fixed formulas — raw figures are shown alongside; missing dimensions mean insufficient data, never zero. Errors? Claim the profile to correct or hide it.

Bibliometric indicators are derived from OpenAlex only; no ORCID is linked, so the profile is name-resolved from public bibliographic data.
Output and impact dimension scores are not computed; citation indicators (h-index 1, i10-index 0) reflect an early-stage record and may lag recent 2026 arXiv output.
Reproducibility and quality assessments rest on only 2 Evolve-reviewed papers, both 2026 arXiv preprints; the 55 mean may shift as more work is reviewed.
Per-year work counts come from OpenAlex countsByYear and are not verified against venue-level records.