Madelon Hulsebos

Research profile

Research summary

Madelon Hulsebos is a researcher at Centrum Wiskunde & Informatica (CWI) in the Netherlands whose work sits at a deliberate intersection: bibliometric field attribution splits evenly (50/50) between information retrieval and databases — specifically LLM-based tabular data analysis — and machine learning, specifically tabular deep learning and foundation models. The two 2026 catalog papers capture both sides: a middle-author position on the TabPFN-3.5 technical report, and the last-author slot on the Open Tabular Insight Extraction survey. TabPFN-3.5 claims state-of-the-art results on seven tabular benchmarks through new cell encodings and a doubled-width architecture; Evolve finds the evidence broad and convincing while noting the headline variants are closed-source. The OpenTI paper unifies the area under one framework and surveys 58 systems across 42 benchmarks, with headline statistics that Evolve flags as resting on appendix methods not shown in review. Activity metrics point to a steep output curve: recorded works rise from 3 (2018) to 6 (2021) and peak at 18 (2025), with 32 works in the most recent three-year window, alongside an h-index of 10 and an i10-index of 10. Collaboration is broad — 48 distinct coauthors, including recurring partners such as Maarten de Rijke and Léo Grinsztajn. Reproducibility reads as a work in progress, averaging 38 across the two reviewed papers (medium for TabPFN-3.5, low for OpenTI). Output and impact composite scores are not yet computed in this snapshot.

Research keywords

Highlights

• TabPFN-3.5: pushing tabular foundation models to claimed SOTA

  The technical report introduces new cell encodings and a doubled-width architecture, claiming state-of-the-art on seven tabular benchmarks with non-i.i.d. evaluation via the TabArena benchmark. Evolve judges the evidence broad and convincing, while flagging that the headline variants are closed-source (repro readiness: medium; Hulsebos is a middle author).

• OpenTI: a unifying framework for open tabular insight extraction

  As last author, Hulsebos contributes to a position paper that unifies open tabular insight extraction under the OpenTI framework and surveys 58 systems across 42 benchmarks, spanning table question answering, text-to-SQL, data analysis agents, and data lakes. Evolve finds the synthesis compelling but notes the headline statistics rest on appendix methods not shown in review (repro readiness: low).

• A two-sided profile: tabular ML meets data management

  Bibliometric attribution splits evenly (50/50) between machine learning (tabular deep learning and foundation models) and information retrieval and databases (LLM-based tabular data analysis). Combined with 48 distinct coauthors and a maximum activity score of 100, the record shows a researcher moving between model building and data systems rather than settling in either community alone.

Timeline

1. 2018Earliest year in this profile's activity record: 3 works logged, at the start of a curve that later accelerates sharply.
2. 20216 works recorded, marking steady mid-period output between the 2018 start and the later surge.
3. 2025Peak activity year: 18 works recorded, and 32 works now fall within the most recent three-year window.
4. 2026Two Evolve-cataloged papers appear: the TabPFN-3.5 technical report (tabular foundation models, middle author) and the Open Tabular Insight Extraction survey (LLM-based tabular data analysis, last author).

Papers in the Evolve catalog

• 2026-09-17

  TabPFN-3.5: Technical Report

• 2026-09-07

  Open Tabular Insight Extraction: Where Do We Stand, and Where Should We Go?

Frequent co-authors