TabPFN-3.5: Technical Report
Benjamin J\"ager, Nick Erickson, L\'eo Grinsztajn, Felix Birkel, Klemens Fl\"oge, Oscar Key, K\"ur\c{s}at Kaya, Jonas K\"ubler, et al.
cs.LG · stat.ML
Quick verdict
Claims SOTA on seven tabular benchmarks via new cell encodings and doubled width; evidence is broad and convincing, though headline variants are closed-source.
Summary
Prior Labs presents TabPFN-3.5, a 220M-parameter tabular foundation model trained jointly for classification and regression on up to 1M rows and 6,000 features, doubling the in-context transformer width of TabPFN-3 (512 to 1024) and adding Fourier value encodings, in-context ECDF rank features, QK-norm, and a synthetic prior emphasizing high-cardinality, wide, and grouped data. Open checkpoints include TabPFN-3.5 and the 84M-parameter TabPFN-3.5-Fast; proprietary API variants (Plus, Thinking) add native text handling and inference-time compute. The report claims rank-1 results on seven benchmarks: TabArena (Elo 1866 vs TabFM+ 1823), BeyondArena, STRABLE, MulTaBench, RelArena-a, TALENT (average rank 2.95), and ScoringBench (mean CRPS rank 2.85); Fast trades roughly 150 Elo for up to 3x speed. Strengths include transparent protocols (Appendix C), significance testing, and a reproduction check of TabPFN-3's published CRPS. Main limitations: proprietary variants decide rank-1 on four of seven benchmarks yet their implementation is undisclosed; TFM wrappers were re-engineered and re-timed while baseline timings were inherited; grouped-slice gains partly rely on exposing group identifiers, a deviation from the official BeyondArena protocol; and open weights forbid commercial use.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
4 strengths · 4 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- TabPFN-3.5 sets a new state of the art on seven benchmarks: TabArena, BeyondArena, STRABLE, MulTaBench, RelArena-a (via TabPFN-Rel), TALENT, and ScoringBench.
- The gains come from revised per-cell encodings (Fourier features and in-context ECDF), doubled model width, a single joint classification/regression checkpoint, and a synthetic prior retuned toward high-cardinality, wide, and grouped data.
- TabPFN-3.5-Fast runs up to 3x faster than TabPFN-3 while remaining about 150 Elo points ahead of it on TabArena.
- Cached inference keeps per-row latency and cache size essentially unchanged from TabPFN-3 despite roughly quadrupled parameters.
- TabPFN-3.5-Thinking pushes accuracy further and is up to 12x faster than TabPFN-3-Thinking, using only inference-time computation on the base model without LLMs or external data.
Strengths
Evaluation breadth with transparent protocol detail
Seven benchmarks spanning i.i.d. prediction (TabArena, 51 datasets / 816 tasks), non-i.i.d. splits (BeyondArena), string columns (STRABLE, 108 datasets), multimodal text+image (MulTaBench, 40 datasets), relational data (RelArena-a, 21 tasks), TALENT (274/337 datasets), and distributional quality (ScoringBench, 101 datasets), each with the full protocol documented in Appendix C, complete leaderboard tables (Tables 2-6) and pairwise win-rate matrices (Figs 10, 12).
Statistical hygiene beyond mean scores
Wilcoxon signed-rank test for per-dataset improvement over TabPFN-3 (p < 0.0001, Fig 1b), bootstrap confidence intervals on Elo (Figs 3-4, 16), a one-sided sign test on ScoringBench (85/101 datasets, p = 8e-13, median +1.3% CRPS), and a verification re-run showing TabPFN-3 reproduces its published CRPS to within 1e-3 relative (C.7).
Honest reporting of residual gaps
Concrete, mechanistic model changes
Weaknesses
Proprietary variants carry the headline
Asymmetric timing methodology
Grouped-data gains hinge on a protocol change
Cross-paper comparisons and licensing friction
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
- All 4 strengths & 4 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is TabPFN-3.5 and how is it different from TabPFN-3?
TabPFN-3.5 is Prior Labs' tabular foundation model: a 220M-parameter transformer that makes predictions on tables in-context, without per-dataset training, supporting up to 1 million rows and 6,000 features. Compared with TabPFN-3 it doubles the transformer width (512 to 1024), trains one joint checkpoint for classification and regression, adds Fourier value encodings and in-context ECDF rank features, and uses a synthetic training prior enriched for high-cardinality, wide, and grouped data. Code is Apache 2.0; weights are open but non-commercial.
Can I use TabPFN-3.5 commercially?
Not with the open weights. The TABPFN-3.5 License v1.0 permits research, evaluation, internal benchmarking and testing on your own data, but production or revenue-generating use requires a Commercial Enterprise License, obtained through the Prior Labs API, AWS SageMaker, Azure AI Foundry, SAP AI Core, or on-prem deployment. The more capable Plus and Thinking variants are available only through these commercial channels.
Does TabPFN-3.5 beat XGBoost and CatBoost?
On the benchmarks reported (TabArena, TALENT, STRABLE, ScoringBench), the TabPFN-3.5 family outscores tuned and ensembled gradient-boosted trees - for example on TabArena, TabPFN-3.5 reaches Elo 1866 versus 1394 for tuned+ensembled CatBoost - and even the faster TabPFN-3.5-Fast stays well ahead. However, on BeyondArena's grouped, temporal, and large-table slices, tuned and ensembled MLPs remain the highest performers, with TabPFN-3.5 narrowing but not closing those gaps.
How fast is TabPFN-3.5 compared to its predecessor?
The core model is up to 2x slower than TabPFN-3 on large training sets because its width doubled, but cached single-row prediction matches TabPFN-3's speed. TabPFN-3.5-Fast (84M parameters, 4 estimators) runs up to 3x faster than TabPFN-3 while remaining about 150 Elo points ahead of it on TabArena. TabPFN-3.5-Thinking is up to 12x faster than TabPFN-3-Thinking due to inference-time improvements.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures arrive as flattened numeric labels (e.g., Figs 3-7, 13-22); error-bar sizes and exact plot groupings could not be visually verified.
- Table formatting was linearized during extraction; per-row values in Tables 2-6 were reconstructed from the flattened text and may contain alignment slips.
- Claims about external benchmarks and competitor models (TabFM, EXAONE-Tabular, RelArena-a) rest solely on this report's description; no external verification was possible from the text alone.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.