The study shows that tabular foundation models (TFMs) can be dramatically simplified without losing performance. By replacing up to 96% of transformer blocks with a label‑free affine translator, researchers found that half‑depth models still outperform full‑depth TFMs and tuned GBDTs across 171 classification datasets. The simplification reduces parameters by up to 1.5 B, speeds GPU inference up to 6×, and cuts memory usage by 6× with minimal impact on AUROC.
arXiv:2608. 10837v1 Announce Type: cross Abstract: The strong performance of foundation models for tabular tasks comes at substantial inference costs.
By Mykhailo Koshil, Matthias Feurer, Katharina Eggensperger
arXiv:2607. 27546v1 Announce Type: new Abstract: Tabular Foundation Models, such as TabPFN, have received a large amount of recent attention due to their performance on in-context tabular machine learning tasks, which often exceeds classical baselines.
By Shuting Luo, Monika Mikhail Kanaan, Cameron Gordon, Anna Leontjeva, Simon Lucey
arXiv:2606. 04485v1 Announce Type: new Abstract: Tabular foundation models (TFMs) increasingly rival tree ensembles, but their performance is often compute-inefficient: with standard affine scalar tokenization, each feature injects value variation through an essentially one-dimensional channel, and feature IDs/positional signals cannot increase within-feature value degrees of freedom, yielding weak early-layer value sensitivity and redundant hidden states.
By Yuanrui Wang, Xingxuan Zhang, Han Yu, Mingchao Ming, Gang Ren, Hao Yuan, Li Mao, Yunjia Zhang, Chun Yuan, Peng Cui
Tydra is a hybrid Transformer‑State Space Model that interleaves attention and SSM layers for tabular in‑context learning. It achieves a 30% reduction in inference time compared to the Transformer‑only TabPFN while preserving most of its predictive performance. On 30 OpenML datasets, Tydra also outperforms a Hydra model that is roughly ten times larger, demonstrating that hybrid architectures can balance accuracy and efficiency for tabular foundation models.
By Mieszko Komisarczyk, Saurabh Mathur, Maurice Kraus, Sriraam Natarajan, Kristian Kersting
The paper investigates how test‑time compute techniques—adaptation, aggregation, and context construction—can improve predictions of pretrained tabular foundation models (TFMs). It introduces DiagScale, a lightweight adaptation method that trains a minuscule fraction of parameters yet matches full fine‑tuning gains, and shows that selective aggregation of predictions from diverse preprocessing variants can reduce error, while broader context expansion offers inconsistent benefits. The study concludes that adaptation and selective aggregation consistently yield benchmark‑level improvements, but their combined use demands more computation, guiding strategy selection based on available resources.
By Kanghui Ning, Marin Bilo\v{s}, James T. Wilson, Yilang Zhang, Kashif Rasul, Dongjin Song, Anderson Schneider, Yuriy Nevmyvaka
The paper introduces a reproducible benchmark for evaluating attention mechanisms in tabular foundation models, focusing on the distinct row and column attention patterns that differ from language model attention. It compares several backends—Torch SDPA, FlashAttention variants, vLLM, and SageAttention—across realistic tabular shapes on A100, H100, and B200 GPUs, revealing that optimal backend choice varies by attention type, hardware, and model specifics. The study finds FlashAttention generally performs best, but CuDNN can outperform it for column attention on longer sequences, while SageAttention excels for large row sequences beyond 16k rows.
By Maximilian Schambach, Clemens Biehl, Sam Thelin
arXiv:2610.19814v1 Announce Type: new
Abstract: Transformer-based tabular foundation models (TFMs) achieve state-of-the-art performance on small to medium-sized tabular prediction tasks while process...
By Wenwen Gong, Haotian Gu, Zhen Yang, Jiaying Ma, Hanqing Zhang, Zhongshen Wang, Jingjun Wang, Shouyi Xu, Wei Wang
TabICLv2 is a new state‑of‑the‑art tabular foundation model that outperforms existing methods on regression and classification tasks. It relies on a synthetic data generation engine for diverse pretraining, architectural innovations such as a scalable softmax attention, and optimized training protocols that replace AdamW with the Muon optimizer. On the TabArena and TALENT benchmarks, TabICLv2 surpasses the current best model, RealTabPFN‑2.5, without any tuning, while also being faster and capable of handling million‑scale datasets with limited GPU memory.
By Jingang Qu, David Holzm\"uller, Ga\"el Varoquaux, Marine Le Morvan
OrionMoE-v1.0 is a compact tabular foundation model that uses a dual‑axis Transformer and a sparse mixture‑of‑experts architecture to perform in‑context learning on labeled support tables. It selectively activates experts along the feature axis, guided by both individual feature representations and overall dataset context, enabling adaptive computation across diverse tabular problems. Pretrained on synthetic classification tasks, it achieves competitive performance on multiple benchmarks, especially on small‑context, high‑dimensional, and class‑imbalanced datasets.
By Mohamed Bouadi, Nassim Bouarour, Aditya Tanna, Vinay Kumar Sankarapu
The paper introduces Neural Spectral Capacity (NSC), a closed‑form metric derived from the singular‑value spectrum of weight matrices that can be computed solely from a network’s architectural specification. Unlike traditional measures such as #Params and #FLOPs, NSC captures architectural structure (depth, width, head, FFN allocations) and can be evaluated without instantiating the model, data, or gradients. Using a dynamic‑programming solver (NSC‑DP), the authors demonstrate that NSC can efficiently identify architectures that outperform existing training‑free proxies across Transformer and CNN families, and achieve state‑of‑the‑art results in tasks such as WikiText‑103 and commonsense reasoning with LLaMA‑7B.
whyItMatters":"NSC provides a fast, architecture‑only proxy that outperforms conventional metrics and training‑free proxies, enabling more effective design and pruning of large models without costly training or data."
By Chenyu Zhu, Ruoyu Zhao, Zhichao Lu
Support-Compiled Feature Folding (SCFF) is a training‑free inference framework that addresses the feature‑side scaling dilemma in tabular foundation models by routing support‑ranked features through bounded leaves of the native encoder, checking residual evidence, and merging encoded messages for a single contextual prediction. This approach transforms quadratic pairwise mixing into linear‑in‑width work with a bounded local working set, achieving dataset‑macro accuracy and NLL improvements across six backbones on an 18‑dataset wide‑table slice. SCFF delivers significant GPU‑memory savings (median 2.09×–2.36×) and, when constrained by a peak‑memory ceiling, further boosts accuracy by up to 4.06 points over the widest single‑leaf baseline.
whyItMatters":"SCFF demonstrates that memory‑efficient inference can simultaneously improve accuracy and reduce resource usage in tabular foundation models, offering a practical solution for deploying these models at scale."
By Tian Zhou, Beverly Jin, Xue Wang, Linxiao Yang, Wenwei Wang, Bingqing Peng, Mengni Ye, Jinjie Gu, Liang Sun