nkh/test-time-compute-for-tabular-foundation-models-results

Test-Time Compute for Tabular Foundation Models: TabArena Results

Per-cell TabArena test errors for the 12 methods ("arms") in the main comparison of Test-Time Compute for Tabular Foundation Models: Mechanisms, Gains, and Limits. Every arm covers all 51 TabArena datasets (816 cells), so these files are enough to recompute the paper's Elo ratings and paired comparisons.

Code: GitHub · Paper: arXiv:2610.12005

Files

file	content
data/results.parquet	config results: one row per arm × cell (9,792 rows)
data/arms.parquet, data/arms.csv	config arms: one row per arm (config, Elo, ΔElo, GPU hours, file checksum)
raw/<arm>/results_per_split.csv	each arm in TabArena's format, the input to ttc.eval.elo

Arms

ΔElo is measured against TabArena's default entry for TabPFN-3 and TabICL v2, and against our frozen run for TabFM. GPU hours are fit + predict time over 816 cells on one H100.

backbone	arm	LR	Elo	ΔElo	GPU h
TabPFN-3	tabpfn3_frozen		1658	+1	0.5
	tabpfn3_native_views_32		1675	+21	1.7
	tabpfn3_full_ft	1e-6	1681	+25	7.7
	tabpfn3_diagscale	3e-2	1682	+26	6.6
	tabpfn3_aggregation_96		1721	+67	247.5
	tabpfn3_diagscale_aggregation_96	1e-2	1748	+93	253.1
TabICL v2	tabiclv2_frozen		1592	+9	1.2
	tabiclv2_full_ft	1e-5	1659	+77	10.9
	tabiclv2_diagscale	3e-2	1664	+82	10.8
TabFM	tabfm_frozen		1792	0	19.3
	tabfm_full_ft	1e-4	1820	+28	61.9
	tabfm_diagscale	1e-2	1812	+20	59.4

Columns of results

column	meaning
arm, backbone, method, learning_rate	arm identity; learning_rate is null without fine-tuning
dataset, fold	TabArena dataset and split (fold enumerates repeat × fold)
metric, metric_error	roc_auc (binary, error = 1 − AUC), log_loss (multiclass) or rmse (regression); lower is better
time_train_s, time_infer_s	wall-clock seconds per cell
problem_type	binary, multiclass or regression
other columns	TabArena bookkeeping (tabarena_method, imputed, ...); metric_error_val is not recorded (null)

Usage

from datasets import load_dataset

repo = "nkh/test-time-compute-for-tabular-foundation-models-results"
results = load_dataset(repo, "results", split="train")
arms = load_dataset(repo, "arms", split="train")

Recompute an Elo rating with the code repository:

hf download nkh/test-time-compute-for-tabular-foundation-models-results \
    --repo-type dataset --local-dir hf_results
python -m ttc.eval.elo \
    --candidate hf_results/raw/tabpfn3_diagscale/results_per_split.csv \
    --reference-field hf_results/raw/tabfm_frozen/results_per_split.csv \
    --baseline "TA-TABPFN-3 (default)"            # ΔElo +26

Notes

• Elo. Fixed field of 68 entries: the 67 official TabArena entries (fetched by TabArena, not included here) plus our frozen TabFM run. Each arm is added alone and rated with TabArena's Bradley–Terry solver, RF (default) = 1000.
• GPU fine-tuning is not bitwise deterministic; re-running an adaptation config gives slightly different cell errors.

License

The results are released under CC-BY-4.0. TabArena data and reference results, and the TabPFN-3, TabICL v2 and TabFM weights (none included), are under their own licenses.

Citation

@article{ning2026testtime,
  title   = {Test-Time Compute for Tabular Foundation Models: Mechanisms, Gains, and Limits},
  author  = {Ning, Kanghui and Bilo{\v{s}}, Marin and Wilson, James T. and Zhang, Yilang and Rasul, Kashif
             and Song, Dongjin and Schneider, Anderson and Nevmyvaka, Yuriy},
  journal = {arXiv preprint arXiv:2610.12005},
  year    = {2026}
}