EXAONE Tabular, a reusable system for making predictions from tabular data, led several of the comparisons in a new arXiv preprint. It recorded the highest classification Elo on TabArena, the highest mean R2 on BCCO and TALENT regression datasets, and the best mean rank on ScoringBench for R2 and RMSE point estimates and for CRPS predictive-distribution quality.
The results were not a clean sweep. EXAONE was second to TabFM on mean accuracy in both the BCCO and TALENT classification comparisons, while its leads were concentrated in regression, ranking and inference efficiency.
Speed sharpened the advantage
On TabArena’s classification leaderboard, default single-forward-pass inference took 0.605 seconds per 1,000 samples. EXAONE sat on the benchmark’s accuracy-latency Pareto front—the set of best observed trade-offs between speed and accuracy—and had the highest Elo rating, about 125 points above TabPFN-3 at comparable latency.
On regression, the reported advantage was mainly about cost. EXAONE reached TabFM’s accuracy regime at roughly one-eleventh of the inference cost. The comparison came with overlapping 95% confidence intervals, so the efficiency result should be read alongside uncertainty about the exact accuracy difference.
Classification was close, not dominant
Across 106 BCCO classification datasets, EXAONE’s mean accuracy was 0.792, compared with 0.799 for TabFM, an absolute gap of 0.007. On the common TALENT classification set, EXAONE averaged 0.857 against TabFM’s 0.863.
The TALENT comparison covered 188 datasets after 12 of the original 200 were excluded because they had more than 10 classes. It therefore describes the common comparison set, not the full original classification collection.
For BCCO, 95% bootstrap confidence intervals were calculated from 10,000 dataset resamples. Because the reported intervals overlapped, the observed difference was not treated as statistically significant without an additional hypothesis test.
Regression led across two suites
The order changed on numeric prediction. EXAONE had the highest mean R2 across 50 BCCO regression datasets, at 0.799, followed by TabPFN-3 at 0.795. Across 100 TALENT regression datasets, it also led with a mean R2 of 0.736, ahead of TabFM at 0.733.
Built to generalize from context
The report’s central question is whether a reusable tabular prediction algorithm can make predictions on new datasets through in-context learning, using the new data as context instead of dataset-specific parameter optimization.
To train it, the authors used separate classification and regression recipes on synthetic tables. The classification setup processed approximately 30 million synthetic table instances, while regression processed approximately 10 million, with bf16 precision.
The model’s processing design, called CAST, interleaves feature-axis processing within each item with support-conditioned item-axis processing within each feature. It adds item-summary and feature-summary tokens while preserving cell-level representations.
A ranking with built-in caveats
ScoringBench used five-fold cross-validation with a fixed random seed, averaged results across folds and calculated mean ranks at the dataset level. Each dataset contributed at most 3,000 observations. The reported measures were R2 and RMSE for point estimation and CRPS for predictive-distribution quality.
EXAONE had the best mean rank for all three measures. But the comparison is metric-specific: the models eligible for the leaderboard, and the effective set of datasets, can differ between R2, RMSE and CRPS. The three top rankings should therefore not be read as one unchanged contest repeated three times.
What the report can support
Taken together, the evidence supports a strong benchmark showing, not a universal claim that EXAONE will outperform every alternative. The evaluation covered four public benchmarks—TabArena, BCCO, TALENT and ScoringBench—and the results describe performance within those benchmark settings.
The report is an arXiv version 1 preprint dated 26 August 2026.
Paper data and sources
Original title: EXAONE Tabular 1.0 : Technical Report
Authors: Moonjung Eo, Min-Kook Suh, Hye-Seung Cho et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text