Beyond IID: How General Are Tabular Foundation Models, Really?

BeyondArena

• BeyondArena unifies IID, temporal, and grouped tasks across sample-size and dimensionality scales, including high-cardinality categorical and text features from diverse applied disciplines.It is integrated into TabArena’s [11] open-source ecosystem to reduce fragmentation in tabular benchmarking.Source: §1 · p. 2

Data curation

• The authors manually curate 142 datasets and introduce DataFoundry, a Python framework and metadata schema for reproducible tabular data curation.The datasets are curated through rigorous protocols across scales and feature types.Source: §1 · p. 2

Evaluation scope

• The benchmark systematically compares 3 state-of-the-art open-source TFMs with 8 strong traditional baselines across diverse tasks.Its motivation is to assess how TFMs perform in real-world predictive applications and close the gap between academic evaluations and practice.Source: §1 · p. 2

Scope and limitations

• BeyondArena covers predictive tabular classification and regression from 100 to 1 million samples, while excluding few-shot settings, image features, survival analysis, and relational learning.The evaluated feature types include high-cardinality categorical and text features.Source: §1 · p. 2

Main findings

• TFMs dominate tiny, small, and IID data but fail to compete with traditional tree-based and deep learning models on non-IID, large-scale, high-dimensional, and high-cardinality categorical datasets.The study also examines the importance of outer test splits, inner validation splits, preprocessing, and probability calibration for reliable evaluation.Source: §1 · p. 2

• The appropriate split depends on the practitioner’s application: random splits suit retrospective fraud investigation, whereas temporal splits simulate real-time fraud prevention and avoid leakage-driven overestimation [15] [75].Using a temporal split can underestimate performance when retrospective evaluation is the actual application.Source: §2 · pp. 2–3
• IID data permits randomly held-out test samples because the application imposes no particular structure on the test distribution.Source: §2 · p. 3
• Non-IID data requires temporal splits for future observations or grouped splits for unseen entities, with grouped tasks supporting label-per-group and label-per-sample objectives.Temporal splits require a time index and strictly later test samples; grouped splits keep each group entirely within training or testing.Source: §2 · p. 3
• Split choice changes the emphasis on temporal or grouped dependencies without removing either, and BeyondArena excludes time-series forecasting because its assumptions require different validation procedures.The curation process distinguishes temporal tabular regression from time-series forecasting, although tabular models can still be competitive for forecasting [78] [79] [81] [82] [83] [84].Source: §2 · p. 3

IID Tabular Benchmarks

• BeyondArena extends predominantly IID benchmarks by unifying IID and non-IID tabular data, including text, while leaving image and time-series benchmarks for future work.Prior benchmarks covered model development, deep learning, AutoML, data-centric evaluation, few-shot or tiny data, broad domains, large data, and real-world tasks, but mostly treated data as IID.Source: §3 · p. 3

Non-IID Tabular Benchmarks

• BeyondArena builds on TabReD with more temporal datasets, more baselines, and tabular foundation models, while enabling IID and temporal research in one framework.TabReD addressed a gap in temporal tabular benchmarks after decades of broader non-IID research.Source: §3 · p. 3
• BeyondArena is the first benchmark to incorporate grouped data from longitudinal, algorithm-selection, and distribution-shift fields, filling a non-IID benchmarking gap across application domains.The paper notes that dedicated benchmarks for tabular deep learning on grouped data had not previously existed.Source: §3 · p. 3

Contribution

• BeyondArena unifies benchmarking through representative datasets in one shared metadata format and rigorous evaluation of state-of-the-art tabular models across tasks and multimodal data types.These contributions are described as the summary distinction from related work.Source: §3 · p. 4

Data Foundry

• Data Foundry provides a Python curation package, metadata schema, dataset checks, train-test splits, and reproducible notebooks for every BeyondArena dataset.The notebooks cover metadata edits, preprocessing, validation, splitting, and artifact export to benchmarks and data repositories.Source: §4 · p. 4

Dataset Sources

• The dataset pool combined 304 datasets reconsidered from 14 prior benchmarks, data from seven non-IID or multimodal benchmarks, and new repository or competition searches.Searches were guided by popularity, task, and application domain while avoiding obvious duplicates.Source: §4 · p. 4

Dataset Selection Criteria

• The revised criteria admit IID, temporal, and grouped tasks; datasets with at least 100 training samples; larger datasets via subsampling; and tabular data with text or date features.Datasets must be unique, predictive, representative of real-world classification or regression, suitable for tabular modeling, and free of obvious ethical concerns.Source: §4 · p. 4

Dataset Processing

• Each selected dataset was standardized with unified formats, feature-type annotations, task metadata, curation notes, and task-specific feature engineering when necessary.Metadata includes train-test splits, target columns, and group or time indices; the process was manually verified because data quality affects benchmark conclusions.Source: §4 · pp. 4–5

Curation Outcome

• 142 of 1,128 investigated datasets met the selection criteria, with roughly 12.6% selected and broad diversity across scales, feature types, tasks, sources, and application domains.Most excluded datasets were duplicates, did not represent real applications, or were not originally predictive tasks.Source: §4 · p. 5, Figure 2 · p. 5

BeyondArena Models

• BeyondArena compares 11 models spanning linear, tree-based, neural, and tabular foundation models, with TabPFN-2.6 and TabDPT restricted to datasets with at most 100k training samples.The benchmark uses verified, open-source implementations and imputes unavailable TabPFN-2.6 and TabDPT results with default Random Forest performance.Source: §5 · pp. 5–6

Outer Splits

• The benchmark uses sample-size-dependent outer splits, repeated or group-based cross-validation, stratification for classification, and application-specific rolling temporal splits that exclude overly unrepresentative training windows.Datasets above 1.25 million samples use 1 million training and 250k test samples; temporal splits require at least 50% of the original data for training.Source: §5 · p. 6

Inner Validation and Tuning Protocol

• Models are tuned with task-matched inner validation, 25 random-search configurations, and time limits, while TFMs are evaluated only through untuned in-context learning to assess pretrained generality.TFM evaluation uses default preprocessing without dataset-specific weight updates; TabPFN-2.6 and TabICLv2 receive extended 12-hour limits, and TabICLv2 uses one ensemble member on the five largest datasets.Source: §5 · p. 6

Multimodal and non-IID Preprocessing

• A unified preprocessing pipeline encodes dates and text and handles grouped data, using 10-dimensional datetime features, 32-dimensional Qwen3 text embeddings, and group-specific index transformations.Label-per-sample datasets drop group indices, whereas label-per-group datasets replace them with 50-dimensional group encodings.Source: §5 · p. 6

Metrics

• Evaluation uses ROC AUC, log-loss, and RMSE for binary classification, multiclass classification, and regression, respectively, and aggregates performance with Elo and mean Improvability.Elo is calibrated so that 1000 corresponds to default XGBoost, while Improvability measures error relative to the best method on each outer split.Source: §5 · p. 7

6 Results

• BeyondArena evaluates 11 models across 142 datasets using task, scale, dimensionality, and special-feature sub-benchmarks, with full-benchmark Elo and Improvability comparisons.The sub-benchmarks cover IID, grouped, temporal, tiny, small, medium, large, low- and high-dimensional, text, high-cardinality, and all-dataset regimes.Source: §6 · p. 7, Figure 4 · p. 8, Figure 5 · p. 8
• TFMs perform best on tiny, small, and IID datasets, but fail to compete with traditional models on non-IID, large-scale, high-dimensional, and high-cardinality datasets.TFMs also lead on average Improvability and default performance, whereas tuned and ensembled traditional models benefit substantially from optimization.Source: §6 · p. 7, Figure 5 · p. 8, Figure 4 · p. 8
• Default TFM training costs, including cross-validation, are substantially lower than tuned competitors’ costs, but TFMs continue to have large inference costs.Nonparametric tests indicate that global rankings and pairwise comparisons across models and sub-benchmarks are generally significant.Source: §6 · p. 7, Figure 4 · p. 8
• TFMs provide peak performance on 70% of datasets, although TabICLv2 ranks first on 19%, TabPFN-2.6 on 10.5%, and TFMs are clearly outperformed on 42 datasets.TFMs significantly outperform other models on 49% of datasets and match the best non-TFM on an additional 21%; grouped, temporal, and large datasets yield peak performance only rarely.Source: §6 · p. 7, Figure 6 · p. 9

BeyondArena Ablations

• BeyondArena-Core is 5× faster than BeyondArena and has 2.3× more stable per-dataset win rates than BeyondArena-Lite, which is 9× faster.Using inappropriate IID outer splits for grouped data can drastically distort rankings, with τ = 0.49 and τ = 0.60 on two exemplary datasets.Source: BeyondArena Ablations · p. 8
• Using 5 × 5 inner cross-validation for datasets with fewer than 500 training samples consistently improves test performance, while non-IID inner splits improve average performance and peak results.IID and non-IID inner splits produce similar ranks, but non-IID splits are appropriate for temporal and grouped data.Source: BeyondArena Ablations · p. 8, Table 1 · p. 9
• BeyondArena preprocessing is generally better on grouped data, but TabPFN-2.6 improves substantially without preprocessing; Qwen3 is better for short text, whereas TF-IDF is better for long text.Rankings remain mostly stable for L-P-G and short-text datasets, while L-P-S rankings are more mixed; text-encoder differences only slightly affect rankings.Source: BeyondArena Ablations · pp. 8–9, Table 1 · p. 9
• Post-hoc calibration improves log-loss most of the time, although TabPFN-2.6 and RealMLP perform worse with calibration.The ablation supports using calibration by default for other models.Source: BeyondArena Ablations · p. 9, Table 1 · p. 9

Conclusion

• BeyondArena evaluates 11 models across 142 datasets and shows that TFMs perform well on small- to medium-scale IID data but are outperformed by traditional methods on non-IID, large-scale, and high-dimensional settings.Together with DataFoundry, it enables standardized, reproducible evaluation across diverse task types, dataset scales, feature dimensionalities, and feature types.Source: §7 · p. 9

Limitations and Societal Impact

• The study is limited by constrained hyperparameter tuning, evaluation of only in-context learning, inclusion of only the top three TabArena TFMs, and uneven dataset representation across sub-benchmarks.Validation and preprocessing best practices for temporal, grouped, and text data also remain to be explored.Source: §7 · p. 9
• BeyondArena improves rigor and reproducibility and narrows the gap between academics and practitioners, but incurs non-trivial computational cost mitigated through BeyondArena-Core.Future evaluations could extend to few-shot prediction, multimodal tabular data, relational learning, and survival analysis.Source: §7 · p. 9

Conclusion

• BeyondArena establishes a broader evaluation standard that surfaces critical limitations of current TFMs beyond the paradigms where they already perform well.The results indicate promising in-context learning performance, but current TFMs are not yet fully general for tabular data.Source: §7 · p. 9

Competing Interests

• The authors report roles in RealMLP, TabPFN-2.6, TabICLv2, TabArena, AutoGluon, skrub, and scikit-learn, alongside affiliations with Prior Labs and probabl; they declare no other competing interests.D.H. authored RealMLP; L.P. and F.H. authored TabPFN-2.6; D.H. and G.V. authored TabICLv2; several authors maintain TabArena or AutoGluon; G.V. contributes to or cofounded skrub and scikit-learn.Source: Competing Interests · p. 10

A.1 Time-series Forecasting Tasks vs. Temporal Tabular Tasks

• Temporal tabular tasks predict at future time points using future covariates and a refit horizon, unlike forecasting tasks with a fixed forecast horizon.Applying forecasting validation to temporal tabular data can produce unrepresentative evaluations and may allow reading the target at prediction time.Source: §A.1 · p. 33
• BeyondArena therefore separates forecasting from temporal tabular data during curation because their validation protocols and real-world contexts differ.Forecasting datasets are excluded from BeyondArena and left to time-series forecasting benchmarks.Source: §A.1 · p. 33, §B.2 · p. 34

B.1 Towards Automated Data Curation?

• Human judgment remained necessary for most curation metadata because automated procedures lacked structured information and agentic LLMs produced hallucinations and faulty decisions.Agentic tooling was most useful for documenting human-created metadata, including through Agent Skills via Claude Code.Source: §B.1 · p. 34

B.2 Details on Dataset Selection Criteria

• BeyondArena selects unique, sufficiently sized, predictive classification or regression datasets with random, temporal, or grouped splits, while excluding few-shot, forecasting, non-predictive, and task-specific benchmark data.Datasets with fewer than 100 training samples are excluded, whereas larger datasets may be subsampled.Source: §B.2 · pp. 34–35
• The curation criteria prioritize real-world representativeness by excluding noncompetitive modalities, artificial or trivial data, irreversible leakage, insufficiently documented datasets, and ethically ambiguous tasks.Vectorized image, text, audio, or time-series data may be included when tabular models remain competitive; datasets may also be excluded when subjects or creators request non-use for machine-learning research, as with PIDD.Source: §B.2 · p. 35

B.3 Details on Dataset Processing

• Dataset processing standardizes identifiers, missing values, names, feature types, dates, scaling, and sample order, while temporal construction verifies prediction-time availability and prevents leakage.Temporal splits are ordered with the most recent test point first, and grouped temporal structure is considered when repeated entities occur over time.Source: §B.3 · pp. 35–36

B.4 Details on Curation Outcome

• The curated collection spans diverse domains and scales, includes text and high-cardinality categorical features, and shows larger average temporal datasets alongside smaller multiclass problems.High-cardinality features comprise 14% of categorical columns; two of 51 accepted TabArena datasets were removed because one was trivial and one raised ethical concerns.Source: §B.4 · p. 36

B.5 BeyondArena Datasets Overview

• BeyondArena contains 142 datasets organized across domains, row and column scales, IID, temporal, and grouped tasks, and problem types.The overview also reports text-feature datasets spanning varied sample sizes, text-column counts, and text lengths, alongside the distribution of categorical cardinalities.Source: Figure B.1 · p. 37, Figure B.2 · p. 37, Figure B.3 · p. 37

C.1 Time Limit Impact

• Only ∼0.003% of jobs—31 out of 785,208—exceeded the four-hour limit, while ∼1.72% exceeded one hour and ∼0.01% exceeded 3.5 hours.The four-hour limit accommodates the benchmark’s larger datasets and applies to one configuration on one train split, including all inner folds.Source: §C.1 · p. 41, Figure C.1 · p. 41

C.2 Additional Details

• Text features are embedded with Qwen3-Embedding-8B into 4096-dimensional vectors, reduced to 32 dimensions using a supported Matryoshka slice, and cached.Datetime features are converted into ten numerical and spline-based periodic features, while validation folds are reduced when minority classes are insufficient.Source: §C.2 · p. 41
• Grouped preprocessing replaces label-per-group indices with 50-dimensional transductive encodings built from within-group feature aggregations.The procedure sorts by group time when available, selects high-variance aggregations, and assumes all samples from a group are available at prediction time.Source: §C.2 · p. 42
• Experiments use GCP virtual machines with A100 GPUs for MLPs and tabular foundation models, while other models run on 16-vCPU, 64-GB-RAM instances.The GPU memory allocation is 40 GB for smaller-data MLPs and 80 GB for larger-data MLPs and foundation models.Source: §C.2 · p. 42

D.1 Inference Time Overview

• Figure D.1 compares models’ improvability and inference time under default hyperparameters, tuning, and tuning with post-hoc ensembling; default TFM performance equals in-context learning performance.Source: Figure D.1 · p. 43

D.2 BeyondArena Leaderboard as a Table

• Table D.1 reports per-model performance across all BeyondArena datasets, but the supplied passage does not provide the table’s numerical comparisons.The leaderboard therefore establishes the scope of the comparison without supporting a specific model ranking here.Source: §D.2 · p. 43

D.3 TabArena-v0.1 Datasets vs. BeyondArena Datasets in the Same Scope

• TFMs dominate the 49 accepted TabArena datasets but are outperformed by RealMLP on 20 newly added datasets within the same scope.Non-foundation-model Elo remains stable, whereas all three evaluated TFMs degrade substantially on the new datasets.Source: §D.3 · p. 43, Figure D.2 · p. 45
• The new datasets are harder because several lie near TabArena’s upper sample-size limit and often have high dimensionality or high-cardinality categorical features.Thus, datasets within previous benchmark scope can still expose future TFM improvements.Source: §D.3 · p. 43
• Table D.1 reports default, tuned, and tuned-plus-ensembled performances, with imputed TabPFN-2.6 and TabDPT results above 100k samples and approximate 95% Elo confidence intervals.The table highlights the three best column values with gold, silver, and bronze colors.Source: Table D.1 · p. 44

E Statistical Analysis

• Methods are compared using normalized error scores from 0 for best to 1 for worst among 27 methods, averaged across cross-validation folds over heterogeneous ROC AUC, log-loss, and RMSE tasks.The analysis controls family-wise error rates with Holm–Šidák correction and reports effect sizes alongside corrected p-values.Source: E Statistical Analysis · p. 46

E.1 Test 1: Are the methods really different from each other?

• The Friedman test rejects the null hypothesis that the 27 methods are equivalent across the 142 datasets, supporting an informative leaderboard ranking.Because ranks are computed within each dataset, the test accounts for every method being evaluated on the same datasets.Source: §E.1 · p. 46
• The pairwise Wilcoxon signed-rank analysis tests paired per-dataset differences using the ranks of their absolute magnitudes while retaining their signs.Its two-sided hypotheses compare symmetric differences around zero against nonsymmetry, with rejection when the corrected p-value is below 0.05.Source: §E.1 · p. 46
• 260 of 351 method pairs (74.6%) show performance differences surviving Holm correction at α = 0.05.Pairwise Wilcoxon signed-rank tests therefore indicate that most individual method pairs can be reliably distinguished.Source: §E.1 · p. 46

E.2 Test 2: At what training-set size does using a TFM stop being a reasonable choice?

• TFM degradation begins at approximately 100,000 rows: tiny, small, and medium datasets are indistinguishable, but every large-scale comparison is significant with large effects.The Kruskal-Wallis test gives H = 30.6 on 3 degrees of freedom with p = 1.0 × 10^-6; pairwise large-scale effects are r = 0.68–0.79.Source: §E.2 · p. 47, Table E.1 · p. 47

E.3 Test 3: How much do tuning and ensembling help non-TFM models?

• All 16 paired tests are significant at padj < 0.001, showing that tuning and ensembling consistently improve every traditional algorithm across the 142-dataset benchmark.RealMLP has the largest tuning-to-ensembling effect (r = 0.87), while LightGBM has the largest default-to-tuned effect (r = 0.89).Source: §E.3 · p. 48, Table E.2 · p. 48

E.4 Test 4: What dataset properties predict that GBDTs will beat TFMs?

• Dataset size, high-cardinality categorical features, maximum categorical cardinality, and missing values predict a larger GBDT advantage over TFMs.The comparison uses Δ = e_TFM − e_GBDT, so positive associations indicate increasing GBDT advantage.Source: §E.4 · p. 48, Table E.3 · p. 48

E.5 Test 5: Which method should a practitioner pick, given their data regime?

• TFMs lead on IID, tiny/small, and low-dimensional data; RealMLP leads on temporal, grouped, medium/large, high-dimensional, and text data, while CatBoost leads on high-cardinality data.All 11 subset-level Friedman tests are significant; only 5.1% of grouped-subset pairwise comparisons survive correction, so only its top method is robustly established.Source: §E.5 · p. 49, Table E.4 · p. 50

F Ablations

• BeyondArena’s ablations find that split count, outer-split correctness, inner validation protocol, grouped preprocessing, text encoding, and calibration can materially change benchmark outcomes.The reported checks cover fewer splits, IID splits for grouped data, 5 × 5 versus 8-fold validation, task-appropriate inner splits, grouped preprocessing, Qwen3 versus TF-IDF, and probability calibration.Source: F Ablations · p. 51

F.1 (Outer Splits, A.1) Using Fewer Splits with BeyondArena-Core

• Core (Φ = 0.8) keeps every dataset’s mean winrate delta below 10%, costs 60% more than Lite, and is 5× cheaper than Full for per-dataset analyses.Aggregate mean winrate deviation remains below 1% for every stability target, while Lite is sufficient for aggregate-only metrics and is 9× faster than Full.Source: §F.1 · pp. 51–52, Figure F.1 · p. 52

F.2 (Outer Splits, A.2) IID Splits for non-IID Grouped Data

• Using IID outer splits for grouped data makes musk nearly trivial, simplifies sat11_hand_algo_runtime, and distorts fold variance and model rankings.The ablation therefore concludes that appropriate outer splits are crucial for valid model comparisons on non-IID data.Source: §F.2 · p. 52

F.3 (Inner Splits, B.1) Using 8-fold Cross-Validation for Tiny Data

• For datasets with at most 500 training samples, 5 × 5-fold validation outperforms 8-fold validation for every tested model, especially RealMLP, while rankings remain comparable.The comparison covers Linear, ExtraTrees, LightGBM, and RealMLP on IID datasets with at most 100 features and three outer folds; TFMs are unaffected.Source: §F.3 · p. 52

F.4 (Inner Splits, B.2) Using IID Splits for non-IID Data

• Changing grouped outer splits to IID drastically altered metric values, variances, and model ranks, with Kendall τ = 0.60 for musk and τ = 0.49 for sat11_hand_algo_runtime.For tiny datasets, 5 × 5 cross-validation consistently outperformed 8-fold cross-validation, while tuned RealMLP received a drastic boost.Source: Figure F.2 · p. 53, Figure F.3 · p. 53
• Non-IID inner splits improved performance in almost all cases, significantly benefiting Linear and LightGBM while leaving ExtraTrees largely unaffected.Using random inner splits could also distort rankings, allowing tuned and ensembled ExtraTrees to beat default LightGBM.Source: §F.4 · p. 54, Figure F.4 · p. 54

F.5 (Grouped Data Preprocessing, C.1a/b) Disabling Preprocessing for Grouped Data.

• Grouped preprocessing improved average performance, but rankings changed substantially and TabPFN-2.6 performed better without preprocessing.The ablation covered Linear, ExtraTrees, LightGBM, RealMLP, TabM, TabPFN-2.6, and TabICLv2 on datasets with ≤100k training samples.Source: §F.5 · p. 54, Figure F.5 · p. 55

F.6 (Text Data Preprocessing, C.2a/b) TF-IDF for Text Encoding.

• Text encoding favored Qwen3-Embedding-8B for short text and TF-IDF with SVD for long text, without changing model rankings.Both encoders produced 32-dimensional output vectors; short text was defined as ≤50 characters on average across text cells.Source: §F.6 · p. 54, Figure F.6 · p. 55

F.7 (Post-processing, D) Using Probability Calibration for Log-loss.

• SMS calibration improved most methods’ log-loss, significantly benefiting tree-based models, while TabPFN-2.6 and RealMLP performed worse.Calibration also yielded significant average-rank gains for several models and changed the model ranking order.Source: §F.7 · p. 56, Figure F.7 · p. 56
• The calibration ablation recomputed multiclass log-loss from saved benchmark predictions using structured matrix scaling (SMS), the probmetrics framework default.BeyondArena otherwise does not use post hoc calibration, despite its potential to improve log-loss.Source: §F.7 · pp. 54, 56

Per Dataset Results

• Per-dataset tables compare metric error and fold-level standard deviations for default, tuned, and tuned-plus-ensemble configurations.The accompanying figures cover datasets including acquire_valued_shoppers_challenge through lung_cancer_epithelial_genexp in this section part.Source: G Per Dataset Results · p. 57, Figure G.1 · p. 57, Figure G.88 · p. 79
• Best-method annotations use average performance and paired Wilcoxon tests with Holm-Bonferroni correction at α = 0.05, while single-split datasets receive no bold or underline annotations.Green marks the best average method; bold and underlining identify methods not significantly worse overall or within the same pipeline regime, respectively.Source: G Per Dataset Results · p. 57

HPO Pareto Trajectory

• HPO Pareto trajectories compare tuned and ensembled configurations by validation error versus cumulative training time as the number of trials increases.Points further down and left dominate, allowing comparison of compute required to reach a given error and the achievable error beyond a single endpoint.Source: G Per Dataset Results · p. 57
• Each trajectory includes ensembles of {1, 2, 5, 10, 15, 20, 26} configurations, with randomly selected configurations averaged across 10 seeds alongside the always-present default.The trajectories therefore represent progressively larger ensemble budgets rather than only one tuned endpoint.Source: G Per Dataset Results · p. 57

Per-dataset evaluation

• Each dataset figure pairs per-method test error with an HPO Pareto trajectory, providing a consistent two-part view of model performance and tuning behavior.This format is repeated across the dataset-specific figures in this section.Source: Figure G.89 · p. 79, Figure G.90 · p. 79, Figure G.91 · p. 80

Application-domain coverage

• The figures include application datasets such as marketing_campaign, maternal_health_risk, miami_housing, and porto_seguro.These examples span marketing, healthcare, housing, and insurance-related prediction settings.Source: Figure G.90 · p. 79, Figure G.91 · p. 80, Figure G.94 · p. 80

Broader dataset coverage

• Additional figures extend the same analysis to large-scale, financial, transactional, behavioral, and infrastructure-related datasets.Examples include sepsis_prediction_1m, santander_customer_transaction_prediction, online_shoppers_purchasing_intention_dataset, and 5g_energy_consumption.Source: Figure G.102 · p. 82, Figure G.118 · p. 86, Figure G.119 · p. 87

Scientific and biomedical datasets

• The section also covers scientific and biomedical datasets, including superconductivity, pancreatic_cancer_mouse_detection, physiochemical_protein, and prostate_cancer_detection.The superconductivity dataset is explicitly identified alongside its corresponding per-method comparison figure.Source: Figure G.104 · p. 83, Figure G.106 · p. 83, Figure G.110 · p. 84

• The paper claims and delivers BeyondArena, DataFoundry, 142 curated datasets, diverse task and feature settings, and the reported model-performance conclusions.These claims are addressed in Sections 4–6.Source: H NeurIPS Paper Checklist · p. 93
• Limitations and societal impacts are discussed in Section 7, while the released code and curated data artifacts are described as posing no perceivable misuse risk.Source: H NeurIPS Paper Checklist · pp. 93, 96, 97
• The paper presents no theoretical results, so assumptions and proofs are not applicable.Source: H NeurIPS Paper Checklist · p. 94
• Reproducibility is supported by documented dataset curation and experimental procedures, Appendix C.2 details, shared code, and released benchmarking and dataset artifacts.Source: H NeurIPS Paper Checklist · pp. 94–95
• The authors report all results with error bars, provide statistical analysis in Appendix E, specify compute resources in Sections 5 and Appendix C.2, and give training and test details in Section 5.Source: H NeurIPS Paper Checklist · pp. 95–96
• The work conforms to the NeurIPS Code of Ethics, credits datasets and respects asset licenses, involves no human-subject research, and uses no LLMs in its contributions.Source: H NeurIPS Paper Checklist · pp. 96–99