Tabular foundation-model researcher at BIH/Charité working on TabPFN-scale architectures and efficient FP8 inference.
2
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
7
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-15) 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—
Not yet scored: publication coverage in the underlying record is incomplete, so total output volume and authorship positions cannot be measured reliably. This axis would aggregate works counts, publication venues, and author roles across all indexed sources.
Impact—
Not yet scored: citation coverage in the snapshot is insufficient — only an h-index of 7 and i10-index of 7 are on record — so citation influence cannot be assessed. This axis would track citations, adoption, and downstream reuse.
Focus0/100
The recorded focus score is 0, the lowest value on this radar, yet the underlying field distribution is narrow rather than scattered: it splits evenly (0.5/0.5) between two adjacent machine-learning lines — efficient inference and tabular deep learning, and tabular deep learning with foundation models. Topic counts reinforce this, led by machine learning and data classification (11), domain adaptation and few-shot learning (7), and neural networks and applications (6).
Collaboration78/100
Collaboration scores 78. The record lists 49 distinct coauthors, and recurring partners include Frank Hutter, Jonas M. Kübler, Benjamin Jäger, Oscar Key, Nick Erickson, Léo Grinsztajn, Klemens Flöge, Felix Birkel, Kürşat Kaya, and Adèle Frankel. Both 2026 catalog papers list him in middle authorship without a corresponding role, consistent with contribution to larger team-driven engineering efforts rather than solo-led projects.
Activity83/100
Activity scores 83, the strongest axis on this radar. Works counts by year rise from a single work in 2017 to 3 in 2022, 5 in 2024, and a peak of 6 in 2025, with 5 recorded for 2026; 16 works fall within the most recent three years. This trajectory indicates a steep ramp into sustained and continuing productivity.
Reproducibility55/100
Reproducibility scores 55, computed from two evaluated papers (reproMean 55, reproPapers 2). Evolve's review of the TabPFN-3.5 technical report finds broad, convincing benchmark evidence but closed-source headline variants, while the attention-quantization paper is described as rigorous yet Blackwell-only, with three seeds and no code stated — both landing at medium readiness.
Research summary
Noah Hollmann is a machine-learning researcher at the Berlin Institute of Health at Charité - Universitätsmedizin Berlin whose OpenAlex-resolved record centers on tabular deep learning and tabular foundation models. His reviewed output concentrates on the TabPFN line: the 2026 TabPFN-3.5 technical report (middle authorship) claims state-of-the-art results across seven tabular benchmarks through new cell encodings and a doubled-width architecture, and a companion paper on attention quantization reports FP8 attention kernels delivering up to 1.7x speedup for TabPFN- and TabICL-class models with negligible accuracy loss. Evolve rates both papers at medium reproducibility readiness, citing closed-source headline variants and, respectively, Blackwell-only hardware, three seeds, and no stated code. Bibliometrically, the record shows 23 works since 2017 with clear acceleration — 16 works in the most recent three years, peaking at 6 in 2025 — alongside 49 distinct coauthors, an h-index of 7, and an i10-index of 7. Topic clusters are led by machine learning and data classification (11), domain adaptation and few-shot learning (7), and neural networks and applications (6). Output and impact axes remain unscored because citation coverage in the underlying data is insufficient, so this profile should be read as an active, highly collaborative contributor in an early maturation phase of a narrowly focused tabular-ML agenda.
Collaboration pattern
Hollmann works in sizeable, stable teams: 49 distinct coauthors appear on his record, with repeat partners including Frank Hutter, Jonas M. Kübler, Benjamin Jäger, Oscar Key, Nick Erickson, Léo Grinsztajn, Klemens Flöge, Felix Birkel, Kürşat Kaya, and Adèle Frankel. On both 2026 catalog papers he sits in middle authorship without a corresponding role, suggesting he contributes specialist expertise within larger, engineering-heavy collaborations.
Research keywords
Highlights
TabPFN-3.5: new cell encodings and doubled width claim SOTA on seven tabular benchmarks
FP8 attention delivers up to 1.7x speedup for tabular foundation models
Timeline
- 2017First work appears in the OpenAlex counts-by-year series (1 work), the earliest signal in this record.
- 2022Output begins to pick up: 3 works recorded, up from 1 in 2021.
- 20245 works recorded, extending the upward ramp in annual output.
- 2025Most active year on record: 6 works, part of 16 works across the most recent three years.
- 2026Two TabPFN-line papers enter the Evolve catalog — the TabPFN-3.5 technical report and FP8 attention quantization — with 5 works 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.
- Bibliometric indicators come from an OpenAlex-resolved identity (identity: openalex, ORCID 0000-0001-8556-518X); coverage may be incomplete.
- Output and impact dimensions are unscored because citation coverage in the snapshot is insufficient; the quoted h-index (7) and i10-index (7) are the only impact signals available.
- Reproducibility evidence rests on only 2 evaluated papers, both 2026 arXiv preprints, so the score may shift as code releases and peer review proceed.
- Reviewed-paper evidence covers two 2026 preprints only; earlier activity (2017-2025) is represented solely by bibliometric counts.