TabPFN and TabICL vs. tuned XGBoost: the model that doesn't train won 14/14

Perhaps a more interesting exploration of when it does well, and when it doesn't against xgboost and also tabm

https://engineering.block.xyz/blog/blocktabbench-evaluating-...

Tabular foundation models are one of those things where when you first look into it, it does not make sense as to why they would work so well but it does.

In drug property prediction domain, tabular foundation models coupled with another foundation model for molecules are pretty close to being the state of art.

Btw the article makes heavy use of AI or is written in that way A lot of unnecessary dramatic flair that gets very tiring

Nice little test, but man this AI writing is a pain. I understand that you want to churn out blog posts, but please don't write in this breathless style. Tell your LLM that it is writing a lab report.

This AI slop is tiring.

Edit: I got downvoted, and it was most likely by the AI slop instigator himself. Imagine being so proud of your slop, that you take it personally when someone critiques it :D

Now I am curious as to how much of his blog is exactly like this.

Edit 2: All of it, it seems. Like; that article about Intern-Decision the 26. September 2026. The author installed it, fixed three failures, tested it on 27,256 days of weather, and published the write-up that same day as it was released.

And then manage to get cranky when someone points it out :D

Edit 3: Misunderstand me correctly. I'm all in AI, and I produce my fair share of AI slop as well, but it is almost exclusively for internal use and I can't imagine treating it as premium content. And if I felt the need to publish any of it, I would ensure that it was clearly marked; AI generated.

I have never seen tabular transformer models beat xgboost/catboost in industrial context where datasets is gigantic. They most produce these results on relatively small datasets, clearly not in the tens of millions of rows.

If you want a better (and Biomedical) benchmark that evaluates a whole range of models: https://tabbench-bio.eu

or a general one: https://tabarena.ai

  XGBoost’s search optimized accuracy, and afterwards I also compare by area under the curve. Which means the “fourteen of fourteen on AUC” is against a boosting model that was not tuned for that metric. Tuning it for AUC would probably improve it there; I did not measure that.

I also did not get why Xgboost had to count its training time for the inference. You only train once. I guess in some scenario, where someone says, "I need the best model now, you have five minutes on this singular dataset", but I have never been in that situation.

I would feel better if scripts were released, because I am fairly dubious. I take it as a given that a tabular model has been pre-trained on all of the public benchmark datasets, but that is what it is.

The slop was so meandering, I am not sure what is truth or not.