Bring deterministic predictions to AI agents with tabular foundation models

Transcript

1. Bring deterministic predictions to AI agents with tabular foundation models Tuana Çelik, Prior Labs AI Engineer Paris, 24 September 2026
2. Tuana Çelik Developer Relations @ Prior Labs [Photo, 300 x 300, drop in here]
3. Agenda 01 What language models do with tables Where they are excellent, and where the answer is not in the data 02 Tabular Foundation Models What they are and how in-context learning over rows works 03 Prediction as an agent tool Why no training loop matters, and what the tool call looks like 04 TabPFN-3.5 TabPFN-3.5 and the hackathon
4. Agenda 01 What language models do with tables Where they are excellent, and where the answer is not in the data 02 Tabular Foundation Models What they are and how in-context learning over rows works 03 Prediction as an agent tool Why no training loop matters, and what the tool call looks like 04 TabPFN-3.5 TabPFN-3.5 and the hackathon
5. The gist of this talk What are we getting at..
6. Language models read structured data well Parsing, lookup and extraction are language tasks. Give a model a CSV and ask. USER SALES.CSV ADDRESS BEDS M2 BUILT SOLD_FOR 1 A Street 3 112 1962 485,000 7 A Street 2 84 1975 362,000 12 A Street 4 141 2004 598,000 3 C Street 3 105 1988 441,000 9 C Street 5 176 2011 745,000 How much did the house at 1 A Street sell for?
7. Language models read structured data well Parsing, lookup and extraction are language tasks. Give a model a CSV and ask. USER SALES.CSV ADDRESS BEDS M2 BUILT SOLD_FOR 1 A Street 3 112 1962 485,000 7 A Street 2 84 1975 362,000 12 A Street 4 141 2004 598,000 3 C Street 3 105 1988 441,000 9 C Street 5 176 2011 745,000 How much did the house at 1 A Street sell for? LLM 1 A Street sold for 485,000. It is a 3-bedroom, 112 m2 house built in 1962. Correct, grounded in the context, reproducible. The answer was in the data and the model found it.
8. What about unseen data? Now the answer is not in the data. The model produces a number anyway. USER SALES.CSV ADDRESS BEDS M2 BUILT SOLD_FOR 1 A Street 3 112 1962 485,000 7 A Street 2 84 1975 362,000 12 A Street 4 141 2004 598,000 3 C Street 3 105 1988 441,000 9 C Street 5 176 2011 745,000 4 B Street 4 138 1998 ? What is the likely sale price for 4 B Street?
9. What about unseen data? Now the answer is not in the data. The model produces a number anyway. USER SALES.CSV ADDRESS BEDS M2 BUILT SOLD_FOR 1 A Street 3 112 1962 485,000 7 A Street 2 84 1975 362,000 12 A Street 4 141 2004 598,000 3 C Street 3 105 1988 441,000 9 C Street 5 176 2011 745,000 4 B Street 4 138 1998 ? What is the likely sale price for 4 B Street? LLM Based on comparable properties, 4 B Street would likely sell for approximately 560,000. 07
10. What about unseen data? Now the answer is not in the data. The model produces a number anyway. USER SALES.CSV ADDRESS BEDS M2 BUILT SOLD_FOR 1 A Street 3 112 1962 485,000 7 A Street 2 84 1975 362,000 12 A Street 4 141 2004 598,000 3 C Street 3 105 1988 441,000 9 C Street 5 176 2011 745,000 4 B Street 4 138 1998 ? What is the likely sale price for 4 B Street? LLM Based on comparable properties, 4 B Street would likely sell for approximately 560,000. Plausible, but nothing was actually learned from the five rows. How certain can we be about the answer? What’s the statistical evaluation of the prediction?
11. Historical context Five rows with a known outcome, one row without. That is a supervised learning problem. SALES.CSV ADDRESS BEDS M2 BUILT SOLD_FOR 1 A Street 3 112 1962 485,000 7 A Street 2 84 1975 362,000 12 A Street 4 141 2004 598,000 3 C Street 3 105 1988 441,000 9 C Street 5 176 2011 745,000 4 B Street 4 138 1998 ?
12. Historical context Five rows with a known outcome, one row without. That is a supervised learning problem. LABELLED ROWS: THE TRAINING SET SALES.CSV ADDRESS BEDS M2 BUILT SOLD_FOR Features (beds, m2, built) and a known target (sold_for). Every pattern the answer can be based on lives in historical 1 A Street 3 112 1962 485,000 7 A Street 2 84 1975 362,000 12 A Street 4 141 2004 598,000 3 C Street 3 105 1988 441,000 9 C Street 5 176 2011 745,000 4 B Street 4 138 1998 ? tabular context.
13. Value prediction is a different job WHAT AN LLM IS TRAINED TO DO The house sold for ? approximately .34 485,000 .22 about .19 a .08 A distribution over words. Good at: Reason Plan
14. Value prediction is a different job WHAT AN LLM IS TRAINED TO DO The house sold WHAT A PREDICTION TASK ASKS FOR for ? beds=4 approximately .34 485,000 .22 about .19 a .08 480k m2=138 built=1998 › sold_for = ? P(sold_for | features): 556k, 90% interval 510k to 600k A distribution over words. Good at: A distribution over values, calibrated on the rows. Needed for: Reason Plan Forecast Predict 640k
15. Value prediction is a different job
16. Real tables are wide, mixed and long Numbers, categories, dates, free text, gaps. Thousands of rows and hundreds of columns. A REALISTIC TABLE ID BEDS M2 BUILT HEATING AGENT_NOTES EPC int int float date category text ordinal 10021 3 112.0 1962-04 gas south-facing garden C 485,000 10022 2 84.5 1975-11 electric — D 362,000 10023 4 141.2 2004-06 heat pump renovated 2021, loft B 598,000 10024 3 105.0 — gas needs roof work E 441,000 10025 5 176.4 2011-09 district corner plot A 745,000 10026 1 48.0 1930-01 gas ground floor — 221,000 … 1,000,000 rows SOLD_FOR target 200 columns …
17. PART TWO Tabular Foundation Models
18. What a tabular foundation model is: TabPFN A pre-trained transformer that predicts on tabular data in-context. Your table is the “context”.
19. One model per domain, or one model Classical tabular ML trains per dataset. A tabular foundation model is trained once and reads each dataset in-context.
20. TabPFN scikit-learn compatible by design
21. TabPFN scikit-learn compatible by design
22. TabPFN - Interpret Results
23. TabPFN - Classification, Regression, Forecasting 17
24. Why ‘no training loop’ matters for agents The properties a prediction tool needs are the ones a per-task model cannot provide. A prediction is a tool call The dataset can arrive mid-conversation Same shape as search or code execution: arguments in, result out, A CSV the user uploads, a query result, rows the agent assembled itself. seconds of latency. It fits the loop the agent already runs. No model existed for it a minute ago and none needs to. One tool covers every task The output is a distribution Classification, regression, forecasting, from the same weights. One tool Probabilities the agent can reason over: act, ask for more data, or escalate definition, one deployment, no registry of per-task models. to a human. An eyeballed number gives it nothing to reason with.
25. Prediction as a tool call Twelve lines. No training script, no model artefact, no schedule. fit() Akin to providing context to an LLM predict_proba() One forward pass over context and query. The agent gets a grounded prediction Not a sentence with a number in it.
26. LIVE DEMO An agent that predicts
27. How the field has been evolving TabPFN-3.5, released 15 September, ranked first on TabArena and BeyondArena at release Source: TabPFN-3.5 technical report and release notes, 15 September 2026. Verify against the live leaderboards before the talk.
28. How the field has been evolving A Jev detour: “a new class of frontier models built to make fast, structured decisions that software c use directly.”
29. How the field has been evolving A Jev detour
30. TabPFN-3.5 shipped on 15 September THREE VARIANTS ACCESS Plus, Thinking and Fast Plus for text-rich tables. Thinking, the most accurate, for grouped and temporal data. Fast (alpha) runs up to 6× quicker than base for latency-sensitive tools. pip install tabpfn Open weights for base and Fast. Plus and Thinking via API, MCP, SAP AI Core and AWS SageMaker. 50% off 3.5 token rates until 29 September. Release notes: docs.priorlabs.ai/changelog/tabpfn-3.5, 15 September 2026.
31. Where it ranks on TabArena and BeyondArena Source: TabPFN-3.5 technical report, 15 September 2026. 21
32. Build with TabPFN-3.5, win an NVIDIA GPU A hackathon for exactly the kind of thing in this talk: agents that predict instead of guess. Join the hackathon
33. Build with TabPFN-3.5, win an NVIDIA GPU A hackathon for exactly the kind of thing in this talk: agents that predict instead of guess. 🥇 Nvidia DGX Spark 🥈Nvidia Jetson AGX Orin 64GB 🥉Nvidia GeForce RTX 4090 Join the hackathon
34. Give your agent a prediction tool Install pip install tabpfn Docs docs.priorlabs.ai Learn about TabPFN-3.5