HBISS Recap: The Foundation Model Revolution for Structured Data; Frank Hutter, Prior Labs. Sept 22, 2026

HBISS Seminars
, , , , ,
1

Hi @AWGall September 22, 2026

Today’s Horizons in Biosciences and Informatics Seminar series (HBISS) with ~50 in attendance featured Dr. Frank Hutter on tabular foundation models: AI built for the rows and columns of spreadsheets and databases, which is where most of science, medicine, and spaceflight data actually lives. :robot::bar_chart: Recording here.

Frank is founder and co-CEO of Prior Labs, the company behind TabPFN, a PI at the ELLIS Institute Tübingen, and Director of the ELLIS unit Freiburg. Before this he pioneered automated machine learning (AutoML). Fun fact: he interned at NASA Ames in 2001 and 2002 at the start of his PhD. His Google Scholar.

Here is the link to view the recording if you were unable to join. Slides will be uploaded here soon.

What Frank covered

Why tables were the holdout. Deep learning transformed text and images, yet for two decades tabular prediction stayed with gradient-boosted trees (XGBoost, CatBoost). LLMs are poor at this job: they misread numbers, ignore the fact that swapping two rows or columns leaves a table unchanged, and hallucinate. In the last two years tabular foundation models changed that, and adoption is spreading fast, with hundreds of published applications, 104 of them biomedical.

How TabPFN works. TabPFN never trains on your data, or on anyone’s real data. Prior Labs pre-trains it once on hundreds of millions of synthetic tables, each generated from a random cause-and-effect graph. By predicting hidden values across all those tables, the network learns how columns relate to each other in general, the way vision models learned pixel relationships and language models learned word relationships. At use time you hand it your labeled rows plus the new unlabeled rows, and it predicts the missing labels in about a second, in two lines of code, with honest uncertainty attached. No training run, no hyperparameter tuning. This also answers the “but there is no space prior” question: the model learned structure, not domain facts, so it transfers to spaceflight data the same way it transferred to medicine and finance.

The scaling story. TabPFN-1 (2022) proved the idea on tiny clean datasets. TabPFN-2 landed in Nature in January 2025, Nature’s most-cited AI paper of 2025 (~1,900 citations, per Frank): a 5-second prediction beat trees tuned for 4 hours. TabPFN-3 scaled to a million rows, 160 classes, and text columns, cut prediction time to milliseconds, and ranked near the top of time-series forecasting benchmarks without ever seeing a time series. TabPFN-3.5, released last week, leads the TabArena leaderboard. On a classic 2015 Kaggle AutoML challenge with 3,500 teams, the best automated system needed 24 hours on 96 CPUs to crack the top 35; TabPFN-3.5 reached rank 1 in one minute on one GPU.

Small data is the sweet spot. This matters most for us. Spaceflight biology runs on small N, and TabPFN is strongest exactly there: it is trained to stay calibrated rather than overconfident, and in production you skip train/test splits entirely and put all labeled data in as context. A 2026 Nature Communications study built a TabPFN-based test for early pancreatic cancer from serum metabolomics that held ~0.90 accuracy with only 50 to 70 training cases.

Trust and access. Frank closed on trustworthiness: interpretability, robustness, fairness, and causal reasoning that moves past correlation. Do-PFN (NeurIPS 2025) predicts what happens after an intervention using only observational data. Every TabPFN version is open-weight; from 2.5 onward the license is non-commercial, and entirely free for research: pip install tabpfn, runs offline on a CPU or an old gaming GPU. There is also a free API for publishable data and an MCP server so your AI agent of choice can run TabPFN for you.

Key discussion highlights

A strong Q&A that ran well past the hour.

  • Nil Mu @Nil asked why TabPFN beats AutoML computationally. One forward pass replaces model selection, tuning, and ensembling; the pass finishes before a single XGBoost would even train.
  • gabyc (unsure which Gabby in the AWG) asked whether TabPFN needs more data than other tools. The opposite: the massive pretraining is Prior Labs’ job, done once; you bring your 100 rows.
  • James Casaletto @james.casaletto asked about splits and minimum sample size. For real use, all labeled data goes in; splits exist only to score methods against each other. There is no minimum-N rule, and that is by design: with 20,000 columns and 10 rows, the honest answer is high uncertainty, and TabPFN says so.
  • Jian Gong @jgong asked about missing values (handled natively since TabPFN-3 with a missingness flag plus imputation; @asaravia followed up and Frank pointed to Fig 5, p6 of the report), edge devices (runs fine on modest hardware), and what the new “Thinking” mode is (extra compute wrapped around the base model at prediction time, no LLM involved).
  • Matthew Gebre @mgebre256 asked which data TabPFN handles worst. The toughest case the team found was data whose true pattern is itself a random forest, where tree methods keep a small edge. He also raised long COVID detection (Frank: bring data!) and space weather prediction after solar particle events. :high_voltage:
  • @lauren.sanders asked about, then it was confirmed by Frank that the synthetic pretraining tables mix numeric and categorical features.
  • Felipe @pinedaalpa (ETH Zurich) shared that TabPFN powered clinical biomarker mapping in his PhD.
  • The closing thread, with Ryan and James Casaletto @james.casaletto → how do we build ground truth for spaceflight life sciences when only ~650-700 humans have flown and perhaps ~80 maybe have accessible data? Frank’s answer: the biggest gift this community could give causal AI is good benchmark datasets that pair observational data with intervention outcomes, which are rare everywhere. OSDR’s curated, API-accessible studies are strong candidates, and Ryan floated a possible spring hackathon with the Prior Labs team on OSDR spaceflight data. :confetti_ball: :confetti_ball:
  • Ryan also flagged OSDR environmental telemetry as additional prime territory for these models: RadLab and the Environmental Data App.

Next HBISS

October 22, 2026: the first X-ray ever taken in space. :x_ray: The science team behind the Fram2 mission radiograph will cover the full arc, from flight readiness to the device itself, including how to prepare non-clinician crew for point-of-care imaging in deep space. Here is the Forum-Space invite.

All links from the chat and presentation

3 Likes