PC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11Crashes, No Sound, or Screen Glitches?
Random freezes, missing sound and display glitches usually trace back to one bad driver. Find and replace yours safely.Free scan · under a minuteIn a 2026 benchmark by Efrain Garay, TabICL—not TabPFN—had higher ROC AUC than tuned XGBoost on all 14 selected classification datasets, including after XGBoost’s search was rerun to optimize AUC. That is a result from a specific small-table experiment, not evidence that either in-context model always beats XGBoost. TabPFN’s results varied, and accuracy comparisons were less consistent.
What the 14-dataset benchmark found
Garay compared TabICL 2.x, TabPFN 2.2.1, default XGBoost, and XGBoost tuned with a randomized search. The benchmark used 14 classification datasets from the Grinsztajn tabular benchmark, capped each at 3,000 rows, and reported medians over five seeds. Garay describes this collection and size range as a favorable setting for in-context tabular models.
As an Amazon Associate I earn from qualifying purchases.
In the initial run, XGBoost’s 25-iteration search used three-fold cross-validation but optimized accuracy, even though the headline comparison focused on AUC. After that metric mismatch was identified, Garay reran the search using ROC AUC as its scoring metric. TabICL still had higher AUC on all 14 datasets; the reported mean gap was 0.0106 in the AUC-scored rerun, down from 0.0114 in the initial comparison. The reproducibility materials are available in Garay’s benchmark results and script, and the benchmark article explains the setup.
Do these 3 things before closing this tab:
1Clear out junk files and repair common Windows errors2Fix the driver behind crashes, sound loss and screen glitches3Repair Windows errors before they cause bigger problemsThe scope of the claim matters: it is TabICL’s reported AUC direction across these 14 datasets, not a 14-for-14 win for both TabICL and TabPFN, nor a universal result for foundation models against XGBoost.
#1 Best Overall
Why AUC and accuracy tell different stories
AUC result
ROC AUC measures how well a model ranks positive examples above negative ones across possible classification thresholds. Garay reports TabICL ahead of tuned XGBoost on AUC in all 14 datasets, including the rerun that optimized the tree model for AUC. The average gap is a benchmark-level summary; it does not mean every dataset had a large or practically important difference.
Accuracy result
Accuracy counts the share of predictions that are correct at a chosen threshold. Garay reports TabICL had higher median accuracy on 12 of the 14 datasets, but only about seven of those comparisons remained outside the seed-to-seed spread. That makes the accuracy evidence less decisive than the consistent direction of the AUC comparisons.
Metric choice should follow the task. For example, if the practical question is whether to prioritize finding likely responders over non-responders, ranking quality may be useful even before a decision threshold is chosen. If the task requires a fixed yes-or-no prediction, accuracy alone may still conceal the costs of false positives and false negatives. Neither metric by itself establishes that a model is suitable for consequential decisions.
What the individual examples show
The benchmark’s displayed seed-0 examples show why the aggregate result should not be mistaken for identical behavior on every table. These are individual-seed scores, not the five-seed medians:
| Dataset | TabICL AUC | TabPFN AUC | Tuned XGBoost AUC |
|---|---|---|---|
| Credit | 0.7667 | 0.7578 | 0.7533 |
| HELOC | 0.7222 | 0.7300 | 0.7078 |
| Default-of-credit | 0.6956 | 0.6967 | 0.6944 |
| Bank marketing | 0.7944 | 0.7967 | 0.7833 |
TabPFN is slightly ahead of TabICL in the bank-marketing example and in two of the other displayed examples. Those cases do not overturn the aggregate finding about TabICL versus XGBoost; they illustrate that the models’ relative scores depend on the dataset.
How to read the uncertainty and limits
Garay notes that the test sets contained 900 rows and estimates an AUC standard error near 0.01. He argues that the direction across seeds—TabICL ahead in 68 of 70 per-seed comparisons—is more informative than any single margin. Those are the benchmark author’s caveats and analysis, not an independent uncertainty assessment.
- Selected tasks: The datasets came from a named benchmark suite; they are not a random sample of all tabular classification problems.
- Restricted size: Each dataset was trimmed to 3,000 rows. The results do not settle how the models compare on larger datasets.
- Specific implementation: The reported versions were TabICL 2.x, TabPFN 2.2.1, and XGBoost 3.4.1, with the preprocessing and search budget used in Garay’s experiment.
- Unanswered settings: The experiment does not establish performance for different feature types, preprocessing, deployment requirements, or larger XGBoost tuning budgets.
Use the result as a useful case study with reproducible materials, not as an independent multi-lab consensus or a replacement for evaluation on the data and operating conditions that matter to you.
Free tools Windows power users keep installed
One-click scans. No signup required.
“Does not train” means no ordinary task-specific weight fitting
TabPFN and TabICL use in-context learning. Their underlying models are pretrained; for a new table, training rows are supplied as context for producing predictions rather than used for ordinary gradient-based fitting of new model weights. A software API may still expose a method called fit, so the method name alone does not establish that gradient descent is occurring on the new task. Nor does the phrase mean there was no prior training or no computation at prediction time.
Garay characterizes the approach this way: “A tabular foundation model is pretrained on millions of synthetic tables generated on purpose.” That is his conceptual explanation and should not be treated as a precise description of every TabICL or TabPFN version. For background on TabPFN’s results against tuned baselines on its own small-tabular benchmarks, see the TabPFN paper in Nature. That paper is separate evidence from Garay’s 14-dataset comparison.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Where the compute goes—and why timing matters
With these pretrained in-context methods, conventional dataset-specific fitting is reduced, while prediction can require more work because the training rows condition predictions at inference time. So a short or unusual fit phase is not enough to conclude that a model is faster overall: fit and prediction time need to be considered separately.
In Garay’s displayed Bioresponse example, which had 419 columns, TabICL’s AUC was 0.8667 and prediction took 6.0 seconds in that setup. Several other displayed examples had prediction times around 0.6–0.8 seconds. These are measurements from the reported experiment, not general latency guarantees or a rule that wider tables always take longer.
Garay’s setup used a 16 GB NVIDIA GeForce RTX 4070 Ti SUPER and allocated 14 CPU cores to XGBoost. Those details help readers understand the conditions behind the comparison; they do not show that this GPU is required or optimal. Reproducing the results also depends on the stated software versions, hardware, preprocessing, seeds, and tuning procedure.
Quick Recap
What to check before choosing a model
- Match the metric to the decision: distinguish ranking performance such as AUC from threshold-dependent measures such as accuracy.
- Test on representative data: the benchmark’s 3,000-row cap and selected datasets do not cover every table size or feature type.
- Measure the full workload: record fitting and prediction separately, especially if predictions will be made repeatedly or latency is constrained.
- Make comparisons reproducible: document model versions, preprocessing, random seeds, hardware, and the XGBoost search budget and scoring metric.
- Check current project details: consult the official TabICL project and official TabPFN project for current documentation, installation, supported limits, and license information; the benchmark’s versions are not necessarily current defaults.
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
