Mitra-v2 Technical Report

Yefan Tao; Xiyuan Zhang; Xinyi Liu; Boran Han; Danielle Maddix; Haoyang Fang; Zhen Han; Jiading Gai

cs.LGSep 3, 2026

Mitra-v2 Technical Report

Authors: Yefan Tao, Xiyuan Zhang, Xinyi Liu, Boran Han, Danielle Maddix, Haoyang Fang, Zhen Han, Jiading Gai,

Organizations: Bernie · Amazon

Abstract

We introduce Mitra-v2, a tabular foundation model that delivers state-of-the-art performance on real-world classification and regression problems, from credit-risk scoring and clinical prediction to equipment-failure detection and house-price estimation. Mitra-v2 is trained only on synthetic data, with a pretraining distribution that is much larger and more diverse than Mitra-v1's. Built on a small 2D Transformer backbone, Mitra-v2 supports longer contexts and larger feature spaces. Improved optimization lets it learn from this larger task distribution. We evaluate Mitra-v2 on the TabArena and TALENT benchmarks, comprising more than 300 real-world datasets under two evaluation protocols. On the full TabArena benchmark, Mitra-v2 delivers state-of-the-art performance at the level of the industry-scale TabFM and EXAONE Tabular models, while surpassing TabPFN-3 by a wide margin in both classification and regression. Mitra-v2 matches the 1.6B-parameter TabFM with only 5% of its size (77M parameters), delivering frontier performance at a fraction of the cost. On TALENT, Mitra-v2 remains among the leading models, clearly outperforming TabPFN-3 and TabICLv2. It also ranks first on classification tasks with more than ten classes, even though it was pretrained only on tasks with at most ten classes. These results make Mitra-v2 one of the strongest and most broadly applicable open tabular foundation models released to date. We release the model weights, the inference and fine-tuning code, and our evaluation results under the Apache-2.0 license.

Figures & tables

Figure 2: Overview of Mitra-v2. Top: the mixture of synthetic pretraining priors: classical SCM, gradient-boosting, and forest/tree generators alongside the new Hybrid SCM prior, which composes neural, tree, Gaussian-process, and autoregressive mechanisms inside a single causal graph. Bottom: the scaled-up pretraining envelope, the element-wise 2D-attention backbone (attention across rows and across columns) carried over from Mitra-v1, and the hybrid optimizer.

Component	Mitra-v1	Mitra-v2
Parameters	cls 75.7M / reg 75.7M	cls 75.7M / reg 76.7M
Maximum support rows in pretraining	512	5,120
Query rows in pretraining	128	1,280
Maximum features in pretraining	16	50
Prior mixture	SCM + tree priors	SCM + Hybrid SCM + tree priors
Training steps	22,000	cls $\approx$ 14,000 / reg $\approx$ 14,000 ${+}$ 3,000 FT

Table 1: Mitra-v1 to Mitra-v2 changes. Only components that differ are listed; both versions share the same 12-layer 2D attention backbone (hidden width 512, four attention heads) and the at-most-ten-class pretraining head.

Generator	Weight
Base TabPFN/SCM	0.35
Hybrid SCM (new)	0.35
Decision tree	0.06
ExtraTrees	0.06
Gradient boosting	0.06
Random forest	0.06

Table 2: Configured outer prior mixture of Mitra-v2 (classification). Hybrid SCM is the generator newly introduced in Mitra-v2.

Panel of Figure 3 — Figure 3: Two deployment-time wrappers that let the frozen checkpoint handle tables beyond the pretraining caps, both reusing the same checkpoint with no additional pretraining. (a) Wide tables: train-only feature reduction shrinks the column set to a fixed budget before fine-tuning and bagging. (b) Many classes: a balanced hierarchical decomposition routes a query through a shallow tree of at-most-10-class Mitra nodes.

Method	Elo	95% int.	Avg. rank	Gap to best (%)
AutoGluon 1.6 (noncommercial, 4h)	1,792.9	$+101/-58$	8.38	7.4
Mitra-v2 (default)	1,774.6	$+95/-70$	9.02	7.1
TabFM (default)	1,773.5	$+99/-95$	9.06	5.3
EXAONE Tabular (default)	1,749.4	$+71/-55$	9.96	8.3
AutoGluon 1.6 (extreme, 4h)	1,742.0	$+86/-55$	10.25	8.1
AutoGluon 1.5 (4h)	1,652.3	$+67/-59$	14.35	9.1

Table 3: Combined TabArena board (51 datasets, 816 units): the top 24 of the 43-method pool, ranked by Elo; lower average rank is better. The parenthetical marks the configuration setting (default; tuned; or tuned and ensembled, abbreviated “tuned + ens.”); the AutoGluon rows are four-hour AutoML portfolios rather than single foundation models, and the TabFM+ ensemble is excluded from main-text boards per the display policy of Section 3.1 . Mitra-v2 is the default fine-tuned and bagged system. All 24 rows shown have 0% imputation; the complete 43-method pool, including TabFM+ and the partially imputed tail, is in Appendix B ( Table 10 ). Gap to best is the mean relative gap to the best entry of the pool on each task (TabArena’s improvability, in %): unlike Elo and average rank, which count wins and losses, it retains the margins, and 0 would mean placing first on every task; lower is better.

Method	Elo	95% int.	Avg. rank	Gap to best (%)
AutoGluon 1.6 (noncommercial, 4h)	1,760.8	$+84/-54$	9.52	9.4
Mitra-v2 (default)	1,756.3	$+120/-78$	9.69	9.0
TabFM (default)	1,755.1	$+125/-111$	9.74	6.3
EXAONE Tabular (default)	1,750.8	$+80/-59$	9.91	10.0
AutoGluon 1.6 (extreme, 4h)	1,711.1	$+61/-44$	11.59	10.1
AutoGluon 1.5 (4h)	1,654.2	$+78/-67$	14.39	10.6

Table 4: TabArena classification board (38 datasets, 594 units); top of a 41-method pool (83 configuration entries). Gap to best as defined in Table 3 . All rows shown have 0% imputation; the complete pool is in Table 11 .

Method	Elo	95% int.	Avg. rank	Gap to best (%)
AutoGluon 1.6 (noncommercial, 4h)	2,078.1	$+208/-155$	4.99	1.7
AutoGluon 1.6 (extreme, 4h)	2,015.5	$+176/-115$	6.28	2.2
TabFM (default)	1,986.8	$+170/-96$	6.96	2.4
Mitra-v2 (default)	1,985.6	$+223/-134$	6.99	1.5
EXAONE Tabular (default)	1,879.5	$+137/-95$	10.07	3.5
TabPFN-3 (default)	1,798.2	$+212/-125$	13.04	3.2

Table 5: TabArena regression board (13 datasets, 222 units); top of a 40-method pool (82 configuration entries). Gap to best as defined in Table 3 . All rows shown have 0% imputation; the complete pool is in Table 12 .

Configuration	Support cap	Train-only FS	Elo	$\Delta$ Elo
Regression (full 222-unit protocol)
Cap at the fine-tuning value	8,192	on	1,848.4	—
Cap raised at prediction time	32,768	on	1,935.9	$+$ 87.5
Classification (38-dataset single-split diagnostic)
Base configuration	8,192	off	1,678	—
$+$ feature selection	8,192	on	1,720	$+$ 42

Table 6: What the two deployment-time levers are worth on the frozen Mitra-v2 checkpoint. Rows within a block differ only in the two environment-variable levers; weights, harness, and comparison pool are held fixed, and

\Delta

Elo is the gain over the row above. The regression block is measured on the full 222-unit protocol; feature selection is always enabled there, so only the support cap is toggled. The classification block is a 38-dataset single-split diagnostic used for attribution, not the full-protocol board; its final row lands within bootstrap noise of the release board of Section 3.3 .

Figure 4: In-context sample efficiency of the released checkpoints. Per-dataset test error (log-loss or

1{-}\mathrm{AUC}

for classification, RMSE for regression) as a fraction of its cap-512 value, as the prediction-time support cap grows: every curve starts at

1.0

, and a value of

0.34

means the error has fallen to about a third of its cap-512 level (a

66\%

reduction). Single TabArena split, fine-tuning support fixed at 8192, all other settings held fixed. Legend entries give the training-set size and the total error reduction. Hollow markers denote caps larger than the rows available to a bag child (

7/8

of the training set); from the first hollow marker onward the curve is exactly flat, an internal control confirming that only the cap varies.

Figure 5: Time cost of the same sweep: end-to-end wall-clock of the full evaluation unit relative to its cap-512 value; legend entries give the absolute cap-512 time on a single H100. Each unit’s cost has two parts that the curve separates visually: fine-tuning all bagged children, which never depends on the prediction-time cap and forms the flat floor of every curve, and bagged prediction, which grows with the cap and accounts for all of the rise. Hollow markers as in Figure 4 ; small excursions below

1.0

are run-to-run fit jitter (about

\pm 4\%

) on datasets where the cap never binds.

Figure 6: Accuracy versus recorded cost on the frozen boards: (a) classification, (b) regression. Each point is one

0\%

-imputed leaderboard entry; the

x

-axis is the median over datasets of the mean fit-plus-prediction time per evaluation unit, as recorded in the reference artifacts (log scale). For Mitra-v2 the harness does not time prediction separately, so its recorded time is already the end-to-end unit cost. Comparator foundation-model defaults are blue circles, AutoGluon AutoML portfolios are dark diamonds, and Mitra-v2 is the orange star; the shaded orange band is the 95% bootstrap interval around Mitra-v2’s rating, and TabFM+ is excluded as in the main-text boards. The top of each panel is a statistical cluster inside that band, reached at per-unit costs spanning more than an order of magnitude; no entry that spends more than Mitra-v2 reaches the band, and every member of the band sits well inside the benchmark’s one-hour per-fit budget (dashed vertical line). No Pareto frontier is drawn: recorded times mix hardware and for TabFM include fits that exceed the one-hour budget ( Section 3.1 ), so the panels compare operating points only.

Method	Elo	95% int.	Rank	Elo	95% int.	Rank	Imp.
	Train-only context			Train $+$ val context
TabFM	1,503.2	$+62/-56$	4.03	1,527.7	$+67/-58$	3.82	13 / 12
Mitra-v2	1,479.7	$+60/-54$	4.39	1,490.2	$+66/-60$	4.35	0 / 0
EXAONE Tabular	1,451.8	$+52/-46$	4.83	1,519.0	$+57/-52$	3.94	1 / 0
TabPFN-3	1,414.8	$+54/-50$	5.48	1,404.8	$+52/-48$	5.78	1 / 1
TabICLv2	1,399.6	$+54/-49$	5.76	1,393.9	$+53/-49$	5.99	1 / 1

Table 7: TALENT joint board, main274 preset (274 datasets: 173 classification with at most ten classes, 12 with more than ten, 89 regression), 15 seeds per method on fixed splits, no time limit, all methods at author-recommended defaults; top 12 of the pool by Elo (RandomForest anchored at 1,000), scored under the two matched context policies of the text. On the train-only board every in-context model consumes the train split alone (TabFM and EXAONE re-run accordingly); on the train

+

val board TabFM, EXAONE, and Mitra-v2 fold the validation rows into context, while TabPFN-3 and TabICLv2 keep their (stronger) train-only configuration in both. Classical baselines are identical inputs to both pools; their ratings shift by under 2 Elo between the fits. “Imp.” counts imputed dataset cells out of 274 on the train-only / train

+

val board.

Train-only context
	Cls $\leq$ 10 classes	Cls $>$ 10 classes	Regression
Method	(173 datasets)	(12 datasets)	(89 datasets)
TabFM	1,518.8 (3.93)	—	1,480.2 (4.01)
Mitra-v2	1,444.2 (5.16)	2,070.1 (2.08)	1,499.3 (3.75)
EXAONE Tabular	1,432.6 (5.38)	1,770.1 (5.25)	1,466.7 (4.20)
TabPFN-3	1,404.1 (5.94)	1,850.3 (4.17)	1,395.3 (5.33)

Table 8: TALENT main274 per-slice Elo (average rank in parentheses) for the foundation-model group, under both matched context policies. Slices are rated independently and are never mixed: accuracy scores the two classification slices, RMSE the regression slice. TabFM is absent from the more-than-ten-class slice by construction (its API caps the class count at ten). TabPFN-3 and TabICLv2 use train-only context in both blocks, their stronger configuration; classical pool ratings shift by under 3 Elo between the two fits on the two large slices (under 12 on the 12-dataset many-class slice).

Method	Elo	95% int.	Avg. rank	Win rate	Imp.
TabFM	1,516.6	$+64/-58$	3.86	0.905	12
EXAONE Tabular	1,510.4	$+55/-51$	3.94	0.902	0
Mitra-v2	1,488.6	$+69/-59$	4.26	0.891	0
TabPFN-3 ∗	1,369.8	$+50/-45$	6.34	0.822	26
TabICLv2 ∗	1,361.8	$+49/-45$	6.50	0.817	26
RealMLP	1,167.6	$+41/-39$	11.43	0.652	0

Table 9: TALENT joint board, full300 preset (all 300 datasets, no exclusions), scored under the train

+

val context policy (the train-only re-runs of TabFM and EXAONE cover the main274 datasets only, so a train-only full300 board would impute their 26 excluded cells; the text mentions it only to note its imputation count); same protocol, pool, and columns as Table 7 , top 12 by Elo, with “Imp.” counted out of 300. Rows marked ∗ are not cleanly comparable on this preset (see note below).

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 7: Per-task mean metric error difference (comparator $-$ Mitra-v2) against TabPFN-3, TabFM, and EXAONE Tabular on the frozen one-hour classification board; orange favors Mitra-v2, and each column header gives the number of datasets where Mitra-v2 has lower error. Rows are sorted by training-set size (largest on top; row labels give training rows). Magnitudes mix ROC-AUC error and log-loss (descriptive only).

Method	Regime	Elo	95% int.	Avg. rank	Win rate	Gap (%)	Imp.
TabFM+	Ensemble	1,817.4	$+97/-78$	7.58	0.922	5.0	–
AutoGluon 1.6 (noncommercial, 4h)	AutoML	1,792.9	$+101/-58$	8.38	0.912	7.4	–
Mitra-v2	Default	1,774.6	$+95/-70$	9.02	0.905	7.1	–
TabFM	Default	1,773.5	$+99/-95$	9.06	0.904	5.3	–
EXAONE Tabular	Default	1,749.4	$+71/-55$	9.96	0.893	8.3	–
AutoGluon 1.6 (extreme, 4h)	AutoML	1,742.0	$+86/-55$	10.25	0.890	8.1	–

Appendix

Table 10: Complete overall TabArena board (51 datasets, 816 units): all 43 methods, ranked by Elo.

Method	Regime	Elo	95% int.	Avg. rank	Win rate	Gap (%)	Imp.
TabFM+	Ensemble	1,790.7	$+117/-83$	8.42	0.909	6.2	–
AutoGluon 1.6 (noncommercial, 4h)	AutoML	1,760.8	$+84/-54$	9.52	0.896	9.4	–
Mitra-v2	Default	1,756.3	$+120/-78$	9.69	0.894	9.0	–
TabFM	Default	1,755.1	$+125/-111$	9.74	0.893	6.3	–
EXAONE Tabular	Default	1,750.8	$+80/-59$	9.91	0.891	10.0	–
AutoGluon 1.6 (extreme, 4h)	AutoML	1,711.1	$+61/-44$	11.59	0.871	10.1	–

Appendix

Table 11: Complete classification TabArena board (38 datasets, 594 units): all 41 methods, ranked by Elo.

Method	Regime	Elo	95% int.	Avg. rank	Win rate	Gap (%)	Imp.
AutoGluon 1.6 (noncommercial, 4h)	AutoML	2,078.1	$+208/-155$	4.99	0.951	1.7	–
TabFM+	Ensemble	2,078.0	$+205/-146$	5.00	0.951	1.3	–
AutoGluon 1.6 (extreme, 4h)	AutoML	2,015.5	$+176/-115$	6.28	0.935	2.2	–
TabFM	Default	1,986.8	$+170/-96$	6.96	0.926	2.4	–
Mitra-v2	Default	1,985.6	$+223/-134$	6.99	0.926	1.5	–
EXAONE Tabular	Default	1,879.5	$+137/-95$	10.07	0.888	3.5	–

Appendix

Table 12: Complete regression TabArena board (13 datasets, 222 units): all 40 methods, ranked by Elo.

Figure 8: UMAP projections (after PCA whitening) of datasets sampled from the Hybrid SCM data prior with $1$ – $16$ feature dimensions, the range shared with Mitra-v1. Each panel is one sampled dataset, with points colored by class label $y$ . The prior spans a broad range of geometric structures, including compact clusters, rings, filaments, and curved manifolds.

Figure 9: UMAP projections (after PCA whitening) of datasets sampled from the Hybrid SCM data prior with $17$ – $50$ feature dimensions, a higher-dimensional setting that Mitra-v1 does not generate. Each panel is one sampled dataset, with points colored by class label $y$ . Compared with the $1$ – $16$ range (Figure 8 ), these datasets are dominated by denser, blob-like clusters.

Tabular Foundation Models

Tabular foundation models learn reusable representations or in-context prediction skills for structured data such as spreadsheets and databases.

See all 188 papers

TabICLv2: A better, faster, scalable, and open tabular foundation model
Feb 11, 2026Jingang Qu, David Holzmüller, Gaël Varoquaux +1Synthetic Data Pretraining Tabular Foundation Models
SODA Team, INRIA Saclay, Palaiseau, France · Probabl, France
Tabular foundation models for non-tabular tasks
Aug 23, 2026Goran Nakerst, John Brennan, Wouter Beugeling +1Handwritten Digit Recognition Tabular Foundation Models
Institut für Theoretische Physik, Technische Universität Dresden, 01062 Dresden, Germany · Department of Physics, Maynooth University, Ireland · Physikalisches Institut (EP3) and Institute for Topological Insulators, Universität Würzburg, Am Hubland, 97074 Würzburg, Germany
EXAONE Tabular 1.0 : Technical Report
Aug 26, 2026Moonjung Eo, Min-Kook Suh, Hye-Seung Cho +4Tabular Foundation Models Tabular ICL
LG AI Research

Feb 11, 2026J/K move · Enter open · S save

TabICLv2: A better, faster, scalable, and open tabular foundation model

Jingang Qu, David Holzmüller, Gaël Varoquaux, Marine Le Morvan

SODA Team, INRIA Saclay, Palaiseau, France · Probabl, France

Synthetic Data Pretraining Tabular Foundation Models Synthetic Tabular Data Generation Tabular ICL Tabular Prediction

Open paper AskPDF arXiv

Tabular foundation models, such as TabPFNv2 and TabICL, have recently dethroned gradient-boosted trees at the top of predictive benchmarks, demonstrating the value of in-context learning for tabular data. We introduce TabICLv2, a new state-of-the-art foundation model for regression and classification built on three pillars: (1) a novel synthetic data generation engine designed for high pretraining diversity; (2) various architectural innovations, including a new scalable softmax in attention improving generalization to larger datasets without prohibitive long-sequence pretraining; and (3) optimized pretraining protocols, notably replacing AdamW with the Muon optimizer. On the TabArena and TALENT benchmarks, TabICLv2 without any tuning surpasses the performance of the current state of the art, RealTabPFN-2.5 (hyperparameter-tuned, ensembled, and fine-tuned on real data). With only moderate pretraining compute, TabICLv2 generalizes effectively to million-scale datasets under 50 GB GPU memory while being markedly faster than RealTabPFN-2.5. We provide extensive ablation studies to quantify these contributions and foster open research by releasing code for inference, pretraining, and synthetic data generation at https://github.com/soda-inria/tabicl.

Abstract

Tabular Foundation Models

Explore similar work

TabICLv2: A better, faster, scalable, and open tabular foundation model

Tabular foundation models for non-tabular tasks

EXAONE Tabular 1.0 : Technical Report