Chunked TabPFN: Exact Training-Free In-Context Learning for Long-Context Tabular Data

Renat Sergazinov; Shao-An Yin

REVIEW 3 major objections 6 minor 1 cited by

Chunked TabPFN: Exact Training-Free In-Context Learning for Long-Context Tabular Data

T0 review · 3 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Chunked attention lets TabPFN handle 100,000+ rows exactly

desk verdict The chunked softmax trick works and is practically useful, but the 'exact TabPFN v2' claim misses TabPFN v2's grouped feature attention, so the equivalence and the long-context results are oversold as-is. read the letter →

arxiv 2509.00326 v2 pith:BEWMTYWD submitted 2025-08-30 cs.LG

Renat Sergazinov , Shao-An Yin This is my paper

classification cs.LG

keywords tabularfoundationmodelsin-contextlearninglongcontextchunkedattentiontiledTabPFNexactscaleddot-product

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

TabPFN v2, a transformer that predicts on tabular data without per-dataset training, has a 10,000-example context limit because attention memory grows quadratically. This paper claims that computing the same attention in chunks, tiling the query and key/value tensors and merging partial softmax sums, removes that limit without retraining or preprocessing, so datasets with over 100,000 examples can be handled in a single forward pass. On 15 long-context tasks from the TabArena benchmark, the chunked model matches the original below 10,000 examples and keeps improving with additional context well past the pre-training cap, while fitting orders of magnitude faster than tuned tree-based baselines. If correct, this makes TabPFN a much more practical out-of-the-box model for large tabular data and challenges the assumption that such models cannot scale beyond their training context.

What carries the argument

Two-level chunked (tiled) attention with an online log-sum-exp merge. The query tensor is split into tiles; for each tile the key/value tensors are streamed in blocks; a row-wise running maximum plus two accumulators (sum of exponentials, sum of exponentials times values) are updated per block, and the final output is the ratio of the two accumulators. This reproduces monolithic scaled dot-product attention exactly up to floating-point associativity, with peak memory linear in tile size and unchanged FLOPs.

What would settle it

Run the chunked attention and a monolithic attention implementation on identical random Q/K/V tensors and compare outputs to the floating-point associativity tolerance; then re-run the 15-dataset scaling protocol with several random subsample seeds at each context length. If the two implementations differ beyond tolerance, or the beyond-10K gains in AUC/RMSE do not reproduce across seeds, the exactness or scaling claims are overstatements.

Watch

Extended reading notes

Core claim

The paper's central claim is that attention can be made exact over arbitrarily long contexts by a two-level tiling scheme with no change to model weights. For each query tile, the key/value stream is processed in tiles; a row-wise running maximum and two accumulators (sum of exponentials, and sum of exponentials times values) are updated across tiles, and the final answer is the ratio of the two accumulators. This reproduces softmax(QK^T/sqrt(d_k))V exactly, up to floating-point associativity, while peak activation memory scales linearly with tile size instead of quadratically with sequence length and FLOPs stay the same. Using the official pre-trained TabPFN v2 weights without any fine-tuni

Load-bearing premise

The empirical claim that more context keeps helping TabPFN beyond 10K rows rests on a single fixed-seed random subsample per dataset at each context length; if one draw is unrepresentative, the reported scaling trends could be sampling artifacts rather than real in-context learning.

Editorial extensions

If this is right

TabPFN v2 becomes applicable to datasets with more than 100,000 rows on a standard single GPU, without clustering, KNN retrieval, or learned compression.
Because the chunked computation is exact, any part of TabPFN that uses attention (self-attention over training rows or cross-attention to test rows) can be swapped to tiled form without retraining or changing predictions beyond floating-point noise.
The observed scaling curves imply that the 10K context cap of the pre-trained model is a hardware limit rather than a learned attention limit; adding more context can help.
On long-context data, the zero-shot chunked model is competitive with tuned neural and tree-ensemble baselines in accuracy while taking far less time to fit, establishing a fast training-free reference point.
The tiling is backend-agnostic, so it works with FlashAttention or native scaled dot-product attention without altering model outputs.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The exactness argument is not specific to TabPFN: any attention-based model whose Q, K, V tensors can be tiled could use the same two-level scheme, suggesting a general recipe for long-context inference in prior-data fitted networks.
The paper's scaling curves use a single fixed-seed random subsample per context length; re-running with multiple subsamples would show which datasets' gains are robust versus artifacts of one draw.
A per-layer exactness guarantee does not automatically ensure end-to-end bitwise equality across many layers, because rescaling factors propagate; a practical check is to compare full-model predictions between monolithic and chunked attention on the same minibatch.
The observed dataset-by-dataset behavior (steady gains, plateaus, and one decline) hints that a cheap context-selection policy could concentrate the benefit of scaling and reduce inference cost.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The core of this paper is a clean, correct application of blockwise softmax tiling to TabPFN v2, and the memory numbers are real: on a 32GB V100 they push contexts past 100K without retraining. The appendix proof of the log-sum-exp accumulator is right for full softmax, and the claim that accuracy keeps improving past the 10K pre-training limit on a subset of TabArena tasks is worth checking out. The authors deserve credit for a simple, practical system and for shipping code.

The problem is the 'exact TabPFN v2' claim. The paper itself says 'TabPFN groups features' and then waves it away with 'we omit this detail for concision.' TabPFN v2's feature attention is not vanilla full softmax; it uses grouped/block-sparse attention, which means a block-diagonal mask is part of the model. The chunking algorithm in Appendix A has no mask parameter and its proof covers only the unmasked case. So either the released code silently applies the mask per tile and the paper doesn't say so, or the numbers are produced by a model with full feature attention, which is not TabPFN v2. Either way the exactness claim is at least incomplete. That is a load-bearing issue, not a footnote problem.

The empirical section has a second, smaller issue. The scaling curves use random subsampling with a fixed seed at each context length and report single-seed averages without variance. The conclusion that performance continues to improve beyond 10K is a trend in those curves; with one draw per point and no confidence intervals it's hard to tell which gains are real. And while the authors are careful not to impute their own TabPFN numbers, they lean on TabArena's previously imputed baseline entries without auditing them, and they never compare against TabPFN v2's own long-context ensemble (the kNN or clustering variants). That comparison is the natural ablation for their 'exact long context' claim.

If the mask gap is fixed—say, by describing how per-tile masks are applied, or by explicitly framing the method as 'TabPFN v2 with full feature attention' and assessing the difference—the paper would be a solid systems contribution. As written, the central equivalence is overstated. I'd send it to peer review, but the authors should be told to resolve the mask issue and tighten the statistics before it's accepted. The idea is useful; the current presentation overpromises.

Referee Report

3 major / 6 minor

Summary. The paper proposes chunked/tiled attention for TabPFN v2, claiming exact equivalence to monolithic softmax attention via a running-max log-sum-exp accumulation. The method is implemented with pure PyTorch tensor operations, supports FlashAttention and native SDPA, and is used to run TabPFN v2 zero-shot on TabArena datasets with contexts beyond 10K samples. The authors report that chunked TabPFN v2 matches the standard model on contexts below 10K, continues to improve on the long-context slice up to 100K samples, and achieves competitive TabArena leaderboard results while being much faster to fit than tuned tree-based baselines. The core mathematical identity in Appendix A is standard and correct for unmasked full softmax, but the paper does not address how TabPFN's feature grouping is preserved under chunking, and the empirical scaling claims rest on a single fixed-seed subsample per context length.

Significance. If the exactness claim holds, this is a practically valuable contribution: a framework-native, training-free extension of TabPFN v2 to contexts beyond its pre-training limit, with released code and no parameters fitted to the benchmark. The running-max log-sum-exp formulation is a clean and reproducible way to tile attention. The paper also correctly avoids the circularity of fitting to the benchmark and compares against official TabArena baselines without imputation. However, the central exactness claim is not fully established for the actual TabPFN v2 architecture because the paper omits the feature-group mask, and the long-context empirical claims are supported by single-seed subsampling without confidence intervals. The contribution is therefore significant but conditional on resolving these issues.

major comments (3)

[Section 2, footnote 1; Appendix A, Algorithm 1] The exactness proof applies to unmasked softmax over the full key set. The paper states in Section 2 that 'TabPFN groups features; thus L_q depends on p' and omits this detail. If feature grouping implies block-diagonal attention masks, Algorithm 1 computes the wrong attention: for a query in one feature group, keys from other groups should be excluded, but the recurrence in Eq. (5) and the pseudocode have no mask parameter. Simply setting masked logits to -inf does not work in the running-max recurrence when an entire KV tile contains no valid keys for a given query, because max = -inf leads to e^{-inf - (-inf)} = 1 instead of 0. The claim that 'causal masks (if any) restrict valid r' per tile; the recurrence is unchanged' does not cover block-diagonal group masks. Please specify the exact mask/grouping structure of TabPFN v2's feature attention, extend the recurrence and pseudocode to
[Figure 1; Section 3.1] The claim of 'consistent performance gains from additional in-context examples, even beyond the original pre-training limits' is based on a single random subsample of training examples at each context length ('randomly subsampled with a fixed random seed'). Figure 1 shows aggregate AUC/RMSE averages across 15 datasets without confidence intervals. Since the context composition changes with each subsample, the observed monotone trends could be artifacts of one particular draw. Please report multiple subsampling seeds (or bootstrap confidence intervals) and per-dataset error bars. This is load-bearing for the second contribution, which challenges the assumption that TabPFN cannot scale beyond its pre-training context window.
[Appendix C] The main text states, 'We provide per-dataset scaling curves and a tile-size ablation in Appendix C, together with failure cases and an analysis of when more data helps.' The appendix contains per-dataset plots, but I did not find a tile-size ablation, a failure-case analysis, or a systematic analysis of when more data helps beyond the qualitative 'early plateau / continuous improvement / model failure' categorization. Please either include the missing material or remove the claim from the main text.

minor comments (6)

[Section 2, footnote 1] The footnote 'We omit this detail for concision' is not a harmless omission: the feature-group mask is essential to the exactness claim. This detail should be moved into the main body or Appendix A and handled explicitly.
[Appendix A, Implementation hints] The sentence 'Causal masks (if any) restrict valid r' per tile; the recurrence is unchanged' is misleading for block-diagonal masks, as explained in the major comment. Please clarify the distinction between causal masks and feature-group masks.
[Table 2 and Table 4] Table 2 has a column header 'Avg. AUC error,↑' which appears to be a typo for 'Avg. AUC (↑)'. The same inconsistency appears in Table 4.
[Appendix C text] There are typos: 'satasifaction' should be 'satisfaction', 'hypoethsize' should be 'hypothesize', and 'plateua' should be 'plateau'.
[Figures 2 and 3] The method name appears inconsistently as 'CUSTOMTABPFNV2', 'CustomTabPFNv2', and 'CUSTOMTABPFN2'. Please use one consistent label.
[Table 4 caption] The caption says '(context length 10,000)' but the table reports averages over the long-context datasets, which have much larger contexts. Please clarify what the table actually reports.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: chunked attention is a self-contained exact softmax identity with external evaluation.

full rationale

The paper's central claim is that blockwise online-softmax accumulation reproduces monolithic attention exactly. This is derived in Appendix A from the standard log-sum-exp identity: the update equations (5) and the correctness sketch show how a running max and accumulators over KV tiles recover the full softmax numerator and denominator. This derivation is self-contained, makes no reference to the benchmark results, and does not fit any parameter to the data. The method uses the official TabPFN v2 weights without fine-tuning, and the experimental section compares against external TabArena baselines, so no fitted input is being renamed as a prediction. References to TabPFN [8,9] are legitimate prior work and are not self-citations by the present authors (Sergazinov and Yin do not overlap with those reference lists). The skeptics' concern about omitted feature-group masks is a correctness/faithfulness question about whether the implementation exactly matches TabPFN v2's attention, not a circularity: it does not make the derivation rely on its own conclusion. Likewise, the fixed-seed random subsampling is a statistical methodology concern, not a circularity. The paper's empirical scaling results are external observations, not consequences of the chunking identity. Hence no circular step is present and the score is 0.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central method introduces no fitted parameters that affect predictions; the tile sizes are implementation choices. The load-bearing assumptions are standard math plus domain assumptions about TabPFN's architecture and the representatives of the subsampling protocol. No new entities are postulated.

free parameters (3)

Query tile length ℓ = not specified
Implementation choice controlling peak memory and loop overhead. It does not affect the exact output, only runtime and memory.
Key/value tile length r = not specified
Implementation choice for KV streaming. Does not change the mathematical result, only computational behavior.
Batch tile m = not specified
Optional batch tiling for large B. Does not affect predictive outputs, only peak memory.

assumptions (4)

standard math Online log-sum-exp with a running maximum yields the exact softmax denominator and numerator over the full key set.
Used as the correctness backbone in Appendix A, Equation (5) and Algorithm 1. The paper provides a proof sketch; it is a known identity.
domain assumption TabPFN v2's predictive distribution is permutation-invariant over the sample dimension and is computed by standard attention, so tiling along samples does not alter the model output.
The paper states this in Section 2 but does not independently verify every implementation detail (e.g., feature grouping and masks) that could affect tile-level exactness.
domain assumption The official TabPFN v2 weights, used without fine-tuning, are the same weights reflected in the baseline comparisons.
The training-free claim depends on this premise; the paper states it in Section 3.1 but provides no weight hash or artifact audit.
domain assumption Random subsampling with a fixed random seed produces representative contexts at every length.
The scaling conclusions in Figure 1 and Appendix C rest on this premise; a non-representative subsample could distort the observed trends.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Chunked TabPFN: Exact Training-Free In-Context Learning for Long-Context Tabular Data." pith.science (2026). https://pith.science/paper/BEWMTYWD

@misc{pith2026250900326,
  author       = {Pith},
  title        = {Pith review of: Chunked TabPFN: Exact Training-Free In-Context Learning for Long-Context Tabular Data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BEWMTYWD}},
  note         = {Machine review of arXiv:2509.00326}
}

read the original abstract

TabPFN v2 achieves better results than tree-based models on several tabular benchmarks, which is notable since tree-based models are usually the strongest choice for tabular data. However, it cannot handle more than 10K context tokens because transformers have quadratic computation and memory costs. Unlike existing approaches that rely on context compression, such as selecting representative samples via K-nearest neighbors (KNN), we introduce a tiled-block strategy to compute attention within the TabPFN framework. This design is compatible with standard GPU setups and, to the best of our knowledge, is the first to enable TabPFN to process long contexts without any pre-processing. We demonstrate the effectiveness of our approach on the standard TabArena benchmark, with code available at https://github.com/mrsergazinov/chunk_tabpfn.

Figures

Figures reproduced from arXiv: 2509.00326 by the authors.

**Figure 1.** Comparison of our proposed chunked scheme against the standard TabPFN v2 baseline on 15 long-context datasets. For each context length, the training examples are randomly subsampled with a fixed random seed to ensure feasibility for the standard TabPFN. All experiments were conducted on a 32GB NVIDIA V100 GPU [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗

**Figure 2.** Leaderboard position and imputation. Elo trajectories across tuning budgets. Unlike prior reports, our chunked TabPFN v2 (CUSTOMTABPFNV2 ) results are reported directly and are not imputed; historical imputed entries are indicated with a stripe pattern [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

**Figure 3.** Critical difference diagram for tuned and ensembled methods on the full benchmark. Horizontal bars connect methods that are not statistically significantly different. Our approach (CUSTOMTABPFN2 ) achieves statistically comparable results to tree-based ensembled methods. Complexity. FLOPs match standard attention (Θ(BHLqLkdk)). Peak activation memory depends linearly on ℓ and r: O [PITH_FULL_IMAGE:figures/full_fi… view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: Scaling vs. context length (part I/III). Error (RMSE (↓ ) / AUC ↑/Accuracy(↑)), wall-clock time (s), and peak GPU memory (MB) for TabPFNv2 (baseline) vs. TabPFNv2 (chunked). Datasets 1–5 of the 15 long-context tasks. 12 [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]

**Figure 5.** Figure 5: Scaling vs. context length (part II/III). Error (RMSE (↓ ) / AUC ↑/Accuracy(↑)), wall-clock time (s), and peak GPU memory (MB) for TabPFNv2 (baseline) vs. TabPFNv2 (chunked). Datasets 6–10 of the 15 long-context tasks. 13 [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]

**Figure 6.** Figure 6: Scaling vs. context length (part II/III). Error (RMSE (↓ ) / AUC ↑/Accuracy(↑)), wall-clock time (s), and peak GPU memory (MB) for TabPFNv2 (baseline) vs. TabPFNv2 (chunked). Datasets 11–15 of the 15 long-context tasks. 14 [PITH_FULL_IMAGE:figures/full_fig_p014_6.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

CRUMB: Efficient Prior Fitted Network Inference via Distributionally Matched Context Batching
cs.LG 2026-06 unverdicted novelty 6.0 of 10

CRUMB speeds up PFN inference on large tabular datasets by clustering queries and selecting MMD-matched context subsets, outperforming prior selection methods on the 51-dataset TabArena benchmark across three architec...

Reference graph

Works this paper leans on

15 extracted references · 11 canonical work pages · cited by 1 Pith paper

[1]

Tabnet: Attentive interpretable tabular learning

Sercan Arik and Tomas Pfister. Tabnet: Attentive interpretable tabular learning. InProceedings of the AAAI conference on artificial intelligence, volume 35, pp. 6679–6687, 2021

work page 2021
[2]

Flashattention: Fast and memory-efficient exact attention with IO-awareness

Tri Dao, Daniel Y Fu, Stefano Ermon, Atri Rudra, and Christopher Re. Flashattention: Fast and memory-efficient exact attention with IO-awareness. In Alice H. Oh, Alekh Agarwal, Danielle Belgrave, and Kyunghyun Cho (eds.),Advances in Neural Information Processing Systems, 2022. URLhttps: //openreview.net/forum?id=H4DqfPSibmx

work page 2022
[3]

Tabarena: A living benchmark for machine learning on tabular data, 2025

Nick Erickson, Lennart Purucker, Andrej Tschalzev, David Holzm¨ uller, Prateek Mutalik Desai, David Salinas, and Frank Hutter. Tabarena: A living benchmark for machine learning on tabular data, 2025. URLhttps://arxiv.org/abs/2506.16791

arXiv 2025
[4]

Tunetables: Context optimization for scalable prior-data fitted networks.Advances in Neural Information Processing Systems, 37:83430–83464, 2024

Benjamin Feuer, Robin T Schirrmeister, Valeriia Cherepanova, Chinmay Hegde, Frank Hutter, Micah Goldblum, Niv Cohen, and Colin White. Tunetables: Context optimization for scalable prior-data fitted networks.Advances in Neural Information Processing Systems, 37:83430–83464, 2024

work page 2024
[5]

Revisiting deep learning models for tabular data.Advances in neural information processing systems, 34:18932–18943, 2021

Yury Gorishniy, Ivan Rubachev, Valentin Khrulkov, and Artem Babenko. Revisiting deep learning models for tabular data.Advances in neural information processing systems, 34:18932–18943, 2021

work page 2021
[6]

Tabr: Tabular deep learning meets nearest neighbors in 2023.arXiv preprint arXiv:2307.14338, 2023

Yury Gorishniy, Ivan Rubachev, Nikolay Kartashev, Daniil Shlenskii, Akim Kotelnikov, and Artem Babenko. Tabr: Tabular deep learning meets nearest neighbors in 2023.arXiv preprint arXiv:2307.14338, 2023

arXiv 2023
[7]

Tabm: Advancing tabular deep learning with parameter-efficient ensembling.ICLR, 2025

Yury Gorishniy, Akim Kotelnikov, and Artem Babenko. Tabm: Advancing tabular deep learning with parameter-efficient ensembling.ICLR, 2025

work page 2025
[8]

Tabpfn: A transformer that solves small tabular classification problems in a second.International Conference on Representation Learning, 2023

Noah Hollmann, Samuel M¨ uller, Katharina Eggensperger, and Frank Hutter. Tabpfn: A transformer that solves small tabular classification problems in a second.International Conference on Representation Learning, 2023

work page 2023

Show all 15 references

[9]

Accurate predictions on small data with a tabular foundation model.Nature, 637(8045):319–326, 2025

Noah Hollmann, Samuel M¨ uller, Lennart Purucker, Arjun Krishnakumar, Max K¨ orfer, Shi Bin Hoo, Robin Tibor Schirrmeister, and Frank Hutter. Accurate predictions on small data with a tabular foundation model.Nature, 637(8045):319–326, 2025

2025
[10]

Ring attention with blockwise transformers for near- infinite context

Hao Liu, Matei Zaharia, and Pieter Abbeel. Ring attention with blockwise transformers for near- infinite context. InThe Twelfth International Conference on Learning Representations, 2024. URL https://openreview.net/forum?id=WsRHpHH4s0

2024
[11]

Tabdpt: Scaling tabular foundation models.arXiv preprint arXiv:2410.18164, 2024

Junwei Ma, Valentin Thomas, Rasa Hosseinzadeh, Hamidreza Kamkari, Alex Labach, Jesse C Cresswell, Keyvan Golestan, Guangwei Yu, Maksims Volkovs, and Anthony L Caterini. Tabdpt: Scaling tabular foundation models.arXiv preprint arXiv:2410.18164, 2024

2024
[12]

Neural oblivious decision ensembles for deep learning on tabular data.arXiv preprint arXiv:1909.06312, 2019

Sergei Popov, Stanislav Morozov, and Artem Babenko. Neural oblivious decision ensembles for deep learning on tabular data.arXiv preprint arXiv:1909.06312, 2019

1909 arXiv
[13]

Retrieval & fine-tuning for in-context tabular models.Advances in Neural Information Processing Systems, 37:108439–108467, 2024

Valentin Thomas, Junwei Ma, Rasa Hosseinzadeh, Keyvan Golestan, Guangwei Yu, Maks Volkovs, and Anthony L Caterini. Retrieval & fine-tuning for in-context tabular models.Advances in Neural Information Processing Systems, 37:108439–108467, 2024

2024
[14]

Mixture of in-context prompters for tabular pfns.ICLR, 2025

Derek Xu, Olcay Cirit, Reza Asadi, Yizhou Sun, and Wei Wang. Mixture of in-context prompters for tabular pfns.ICLR, 2025

2025
[15]

Modern neighborhood components analysis: A deep tabular baseline two decades later.ICLR, 2025

Han-Jia Ye, Huai-Hong Yin, and De-Chuan Zhan. Modern neighborhood components analysis: A deep tabular baseline two decades later.ICLR, 2025. 7 AppendixA.Implementation Details Goal.Enable inference with long contexts without altering model parameters or outputs. We keep attent...

2025

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [1]

Tabnet: Attentive interpretable tabular learning

Sercan Arik and Tomas Pfister. Tabnet: Attentive interpretable tabular learning. InProceedings of the AAAI conference on artificial intelligence, volume 35, pp. 6679–6687, 2021

work page 2021

[2] [2]

Flashattention: Fast and memory-efficient exact attention with IO-awareness

Tri Dao, Daniel Y Fu, Stefano Ermon, Atri Rudra, and Christopher Re. Flashattention: Fast and memory-efficient exact attention with IO-awareness. In Alice H. Oh, Alekh Agarwal, Danielle Belgrave, and Kyunghyun Cho (eds.),Advances in Neural Information Processing Systems, 2022. URLhttps: //openreview.net/forum?id=H4DqfPSibmx

work page 2022

[3] [3]

Tabarena: A living benchmark for machine learning on tabular data, 2025

Nick Erickson, Lennart Purucker, Andrej Tschalzev, David Holzm¨ uller, Prateek Mutalik Desai, David Salinas, and Frank Hutter. Tabarena: A living benchmark for machine learning on tabular data, 2025. URLhttps://arxiv.org/abs/2506.16791

arXiv 2025

[4] [4]

Tunetables: Context optimization for scalable prior-data fitted networks.Advances in Neural Information Processing Systems, 37:83430–83464, 2024

Benjamin Feuer, Robin T Schirrmeister, Valeriia Cherepanova, Chinmay Hegde, Frank Hutter, Micah Goldblum, Niv Cohen, and Colin White. Tunetables: Context optimization for scalable prior-data fitted networks.Advances in Neural Information Processing Systems, 37:83430–83464, 2024

work page 2024

[5] [5]

Revisiting deep learning models for tabular data.Advances in neural information processing systems, 34:18932–18943, 2021

Yury Gorishniy, Ivan Rubachev, Valentin Khrulkov, and Artem Babenko. Revisiting deep learning models for tabular data.Advances in neural information processing systems, 34:18932–18943, 2021

work page 2021

[6] [6]

Tabr: Tabular deep learning meets nearest neighbors in 2023.arXiv preprint arXiv:2307.14338, 2023

Yury Gorishniy, Ivan Rubachev, Nikolay Kartashev, Daniil Shlenskii, Akim Kotelnikov, and Artem Babenko. Tabr: Tabular deep learning meets nearest neighbors in 2023.arXiv preprint arXiv:2307.14338, 2023

arXiv 2023

[7] [7]

Tabm: Advancing tabular deep learning with parameter-efficient ensembling.ICLR, 2025

Yury Gorishniy, Akim Kotelnikov, and Artem Babenko. Tabm: Advancing tabular deep learning with parameter-efficient ensembling.ICLR, 2025

work page 2025

[8] [8]

Tabpfn: A transformer that solves small tabular classification problems in a second.International Conference on Representation Learning, 2023

Noah Hollmann, Samuel M¨ uller, Katharina Eggensperger, and Frank Hutter. Tabpfn: A transformer that solves small tabular classification problems in a second.International Conference on Representation Learning, 2023

work page 2023