DFKI NLP researcher working on mechanistic interpretability of multilingual LLMs: steering, probing, translation.
4
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
38
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-17) 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—
Evolve has not yet computed an output score for this profile. This axis would quantify publication volume and productivity trajectory over time from the indexed works record; the underlying metric is missing here, so no score is estimated.
Impact—
No impact score is available yet. This axis would measure citation-based influence — total citations, citation trajectory and field-normalized impact — but Evolve has not computed it for this profile, so nothing is inferred or estimated here.
Focus0/100
The radar assigns focus a score of 0, reflecting a historically broad topical footprint: 341 outputs in NLP techniques, 173 in topic modeling, 102 in speech and dialogue systems, 101 in syntax/semantics and 94 in speech recognition and synthesis. Notably, the field distribution of recent catalog work is far tighter, split evenly between LLM interpretability and mechanistic interpretability of multilingual LLMs.
Collaboration46/100
The radar scores collaboration at 46, a mid-range value consistent with a compact recent network: 44 works over three years involving 9 distinct coauthors. On all four 2026 catalog papers he is a middle author rather than corresponding author, and recurring partners such as Tanja Baeumel, Simon Ostermann and Daniil Gurgurov point to stable small-team work.
Activity100/100
Activity scores a perfect 100. The counts-by-year record shows continuous annual output from 1993 through 2026, with a historical peak of 26 works in 2010, a second high of 24 in 2020, 44 works across the most recent three years, and 19 already recorded for 2026 — a sustained late-career pace.
Reproducibility78/100
Reproducibility scores 78, computed as the mean across two repro-assessed papers. In the current catalog, two papers — compositional attribute steering and tokenization-as-supervision — carry high reproducibility readiness, while the stereotype-tracing and syntax-separation studies are rated medium, with Evolve noting sparsity confounds and missing statistical testing respectively.
Research summary
Josef van Genabith is a researcher at the German Research Centre for Artificial Intelligence (DFKI) whose OpenAlex-linked record spans more than three decades, from one indexed output in 1993 to 19 already recorded in 2026. The profile carries an h-index of 38 and an i10-index of 157, with 44 works in the most recent three-year window. The historical footprint is broad: 341 outputs tagged natural language processing techniques, 173 topic modeling, 102 speech and dialogue systems, 101 syntax, semantics and linguistic variation, and 94 speech recognition and synthesis. Recent work, by contrast, sits entirely within interpretability — the field distribution splits evenly between NLP (LLM interpretability) and mechanistic interpretability of multilingual LLMs. Four 2026 arXiv preprints in the Evolve catalog, on all of which he is a middle author, mark out the current agenda: compositional activation steering of multilingual attributes, tracing stereotype features from representation to output, output tokenization as a hidden supervision signal, and a mechanistic account of word-order determination in multilingual translation. Evolve rates reproducibility at 78 (the mean over two repro-assessed papers), with two catalog papers rated high and two medium readiness. Radar dimensions currently cover collaboration (46), activity (100), focus (0) and reproducibility (78); output and impact scores are pending.
Collaboration pattern
Collaboration scores 46 with a compact recent network: 44 works over three years involving 9 distinct coauthors. On all four 2026 catalog papers he appears as a middle author, none as corresponding author. Recurring partners — Tanja Baeumel, Simon Ostermann, Daniil Gurgurov, Mikhail Sonkin, Hyun Gu Kang, Ariun-Erdene Tumurchuluun, Yusser Al Ghussin, Pinzhen Chen and Koel Dutta Chowdhury — suggest stable, small-team interpretability work rather than large consortium authorship.
Research keywords
Highlights
Training-free compositional steering of multilingual LLM attributes
Stereotype features: decodable, influential, transferable — but not the same thing
Output tokenization as a hidden supervision signal
Word-order heads fix target syntax in multilingual machine translation
Timeline
- 1993Earliest year in the OpenAlex record, with 1 indexed output — the start of a three-decade publication trail.
- 2010Historical peak of the counts-by-year record: 26 outputs in a single year.
- 2020Second-highest annual output at 24 works, part of a sustained late-2010s surge (21 works in 2016, 24 in 2020).
- 2026Four arXiv preprints on mechanistic interpretability of multilingual LLMs enter the Evolve catalog; 19 outputs already recorded for the year.
Papers in the Evolve catalog
Frequent co-authors
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.
- All bibliometric figures come from an OpenAlex-linked snapshot; citation and output coverage may lag or differ from other databases.
- Output and impact dimension scores have not been computed by Evolve and are left unscored in this profile.
- All four catalog papers are 2026 arXiv preprints with middle-author position and no corresponding-author role; peer-review status is not reflected in the data.
- Identity is name-resolved from public bibliographic data (OpenAlex, ORCID 0000-0003-1322-7944); the listed country code (IN) appears inconsistent with the DFKI affiliation and should be treated cautiously.