7a9f8077b3f461f4239df3e0449c701c26213172e50fefdb44ec1a833a428d7c9a62f228e04afa52204e99913551afa170e47e5e

7eef8605c6836af423ea8fe6469c701b903dd2bf3e0642e1e817633b5a6be88f5f9ea3570a1bd9ca41aa1ba5154bc86d08e6dd19


-

A new version of TabPFN is ready!ย 

When we released TabPFNv1 over three years ago, I didnโ€™t expect at all the hundreds of comments and reposts we would see. Tabular data had been a field getting little love from AI researchโ€”but we immediately felt that this was a topic that data scientists, scientists, financial analysts, and enterprise users deeply cared about, and that using in-context learning could be a breakthrough for it.
We kept pushing: v2 landed in Nature earlier this year, and we started Prior Labs so we could obsess over this full-time. ย 
Now, with TabPFN-2.5, we were honestly surprised by how much further we could go. ย 
On datasets up to 50,000 datapoints and 2000 features, its untuned performance now far outperforms tuned XGBoost and CatBoost. It even matches a 4-hour tuned AutoGluon ensemble - an ensemble that includes our previous v2 model.ย 
We also focused a lot on making models deployment-ready: We show how to export our models to a tree or MLP representation (TabPFN-as-MLP in the plot below), making them fast in inference and easy to deploy. At the same time, we improved our Python SDK (a lot!) running TabPFN in the cloud, and added a REST API for developers to access the model from anywhere. ย 
We are incredibly proud by the ways TabPFN has already helped in science and decision-making, from oncology to climate research (over 400 citations and 100 published use cases since beginning of the year!).ย 
This new release will immediately boost all those applications - for example, our report shows v2.5 is much stronger for causal inference tasks. ย 
It's amazing to see the field accelerating. We can't wait to show you what's next.

TabPFN์˜ ์ƒˆ ๋ฒ„์ „์ด ๋‚˜์™”์Šต๋‹ˆ๋‹ค!

3๋…„ ์ „ TabPFN v1์„ ๊ณต๊ฐœํ–ˆ์„ ๋•Œ, ์ด๋ ‡๊ฒŒ ์ˆ˜๋ฐฑ ๊ฐœ์˜ ๋Œ“๊ธ€๊ณผ ๋ฆฌํฌ์ŠคํŠธ๊ฐ€ ๋‹ฌ๋ฆด ์ค„์€ ์ „ํ˜€ ์˜ˆ์ƒํ•˜์ง€ ๋ชปํ–ˆ์Šต๋‹ˆ๋‹ค. ๋‹น์‹œ ํ‘œ(tabular) ๋ฐ์ดํ„ฐ๋Š” AI ์—ฐ๊ตฌ์—์„œ ๊ฑฐ์˜ ์™ธ๋ฉด๋ฐ›๋˜ ๋ถ„์•ผ์˜€์ง€๋งŒ, ์šฐ๋ฆฌ๋Š” ์ฆ‰์‹œ ๋ฐ์ดํ„ฐ ์‚ฌ์ด์–ธํ‹ฐ์ŠคํŠธ, ๊ณผํ•™์ž, ๊ธˆ์œต ๋ถ„์„๊ฐ€, ๊ธฐ์—… ์‚ฌ์šฉ์ž๋“ค์ด ์ด ์ฃผ์ œ์— ์–ผ๋งˆ๋‚˜ ๊นŠ์€ ๊ด€์‹ฌ์„ ๊ฐ€์ง€๊ณ  ์žˆ๋Š”์ง€ ๋А๋‚„ ์ˆ˜ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋ฆฌ๊ณ  โ€˜์ธ์ปจํ…์ŠคํŠธ ๋Ÿฌ๋‹โ€™์„ ํ†ตํ•ด ์ด ๋ถ„์•ผ๋ฅผ ์™„์ „ํžˆ ๋ฐ”๊ฟ€ ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฐ€๋Šฅ์„ฑ๋„ ๋ณด์˜€์ฃ .

์šฐ๋ฆฐ ๋ฉˆ์ถ”์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค.
์˜ฌํ•ด ์ดˆ Nature์— ์‹ค๋ฆฐ v2๋ฅผ ํ†ตํ•ด ํ•œ ๋‹จ๊ณ„ ๋” ๋‚˜์•„๊ฐ”๊ณ , ์ด ํ”„๋กœ์ ํŠธ์— ์ „๋…ํ•˜๊ธฐ ์œ„ํ•ด Prior Labs๋ฅผ ์„ค๋ฆฝํ–ˆ์Šต๋‹ˆ๋‹ค.

๊ทธ๋ฆฌ๊ณ  ์ด์ œ TabPFN-2.5๋ฅผ ๊ณต๊ฐœํ•ฉ๋‹ˆ๋‹ค.
์†”์งํžˆ, ์šฐ๋ฆฌ๊ฐ€ ์˜ˆ์ƒํ•œ ๊ฒƒ๋ณด๋‹ค ํ›จ์”ฌ ๋” ํฐ ๋„์•ฝ์ด์—ˆ์Šต๋‹ˆ๋‹ค.
5๋งŒ ๊ฐœ์˜ ๋ฐ์ดํ„ฐ ํฌ์ธํŠธ์™€ 2000๊ฐœ์˜ ํ”ผ์ฒ˜๋ฅผ ๊ฐ€์ง„ ๋ฐ์ดํ„ฐ์…‹์—์„œ๋„, ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ ํŠœ๋‹์„ ํ•˜์ง€ ์•Š์€ ์ƒํƒœ๋กœ๋„ ํŠœ๋‹๋œ XGBoost์™€ CatBoost๋ฅผ ๋Šฅ๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ์‹ฌ์ง€์–ด, 4์‹œ๊ฐ„ ๋™์•ˆ ํŠœ๋‹ํ•œ AutoGluon ์•™์ƒ๋ธ”(๊ฑฐ๊ธฐ์—” TabPFN v2๋„ ํฌํ•จ๋ผ ์žˆ์Œ)๊ณผ ๊ฑฐ์˜ ๋™์ผํ•œ ์„ฑ๋Šฅ์„ ๋ƒ…๋‹ˆ๋‹ค.

์ด๋ฒˆ์—๋Š” ๋ฐฐํฌ ์ค€๋น„์—๋„ ์ง‘์ค‘ํ–ˆ์Šต๋‹ˆ๋‹ค.
๋ชจ๋ธ์„ ํŠธ๋ฆฌ๋‚˜ MLP ํ˜•ํƒœ๋กœ ๋‚ด๋ณด๋‚ด๊ธฐ(TabPFN-as-MLP) ๊ธฐ๋Šฅ์„ ์ถ”๊ฐ€ํ•ด, ์ถ”๋ก  ์†๋„๋ฅผ ๋†’์ด๊ณ  ๋ฐฐํฌ๋ฅผ ๋‹จ์ˆœํ™”ํ–ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ Python SDK๋ฅผ ๋Œ€ํญ ๊ฐœ์„ ํ•ด ํด๋ผ์šฐ๋“œ์—์„œ TabPFN์„ ์‹คํ–‰ํ•  ์ˆ˜ ์žˆ๊ฒŒ ํ–ˆ๊ณ , REST API๋„ ์ƒˆ๋กœ ๋งŒ๋“ค์–ด ์–ด๋””์„œ๋“  ๋ชจ๋ธ์„ ์“ธ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

์˜ฌํ•ด์—๋งŒ ์ด๋ฏธ 400ํŽธ ์ด์ƒ์˜ ์ธ์šฉ๊ณผ 100๊ฑด์ด ๋„˜๋Š” ์‹ค์ œ ์‘์šฉ ์‚ฌ๋ก€๊ฐ€ ๋‚˜์™”์Šต๋‹ˆ๋‹ค โ€” ์•” ์—ฐ๊ตฌ๋ถ€ํ„ฐ ๊ธฐํ›„ ์—ฐ๊ตฌ๊นŒ์ง€ ๋‹ค์–‘ํ•˜์ฃ .
์ด๋ฒˆ v2.5๋Š” ํŠนํžˆ ์ธ๊ณผ์ถ”๋ก (causal inference) ์ž‘์—…์—์„œ ํ›จ์”ฌ ๋” ๊ฐ•๋ ฅํ•œ ์„ฑ๋Šฅ์„ ๋ณด์ž…๋‹ˆ๋‹ค.

์ด ๋ถ„์•ผ๊ฐ€ ๊ฐ€์†ํ•˜๊ณ  ์žˆ๋‹ค๋Š” ๊ฒŒ ์‹ค๊ฐ๋‚ฉ๋‹ˆ๋‹ค.
๊ทธ๋ฆฌ๊ณ  ์•„์ง ๋ณด์—ฌ๋“œ๋ฆด ๊ฒŒ ๋งŽ์Šต๋‹ˆ๋‹ค.

- dc official App