KNOWWFrom pre-training to a prediction.
This is how the tabular foundation models the Workbench runs today work, and the design LTFM keeps.
- 01 · Pre-training
Learn shared patterns
Train a transformer across synthetic datasets to learn patterns that transfer between tasks.
What changesModel weights are learned.
- 02 · Inference
Use history as context
Supply observed states, transitions and related records. The pre-trained model uses that context to answer the question.
What changesThe input changes; weights stay fixed.
- 03 · Output
Report uncertainty
Return a prediction with an uncertainty band and reasons. When evidence is too weak, pass the case to the existing process.
What you receiveA score, its uncertainty and an explanation.
Entities, their state, their history and every transition.
A longitudinal table connects records through time. This synthetic supply chain shows how a change in one entity can affect related entities.
A PFN, out of the box, beat four hours of expert tuning in 2.8 seconds.
Hollmann et al., Nature 637 (2025)
Published PFN result. This does not measure LTFM performance.
Two stages to a new foundation model.
One model, built from the inside out. Each stage ships something real before the next begins.
Longitudinal state
A model of each entity's state and how it moves between states. The internal core of LTFM, tested first on financial-services tables through the Workbench.
LTFM
The same core, generalised to every industry.
Three open questions, each with a preregistered design.
Written plans, with the controls and outcome rules fixed before any model runs. None has been run yet; results will be published here either way.
- 01 · Time
Does a frozen temporal encoder beat an analyst's features?
When an entity's history is too wide to fit a tabular model's input, something has to compress it. A frozen time-series encoder is compared with a bank of handcrafted summaries at the same width, across three frozen tabular foundation models.
StatusPlan and code scaffolded. Not yet run.
- 02 · Explanation
Can a foundation model teach a simple, auditable equation?
A tabular foundation model acts as a teacher; a short logistic equation and a symbolic-regression student learn from its probabilities. The test is whether the student beats the same equation fit directly on labels, at equal complexity, on credit datasets.
StatusPlan written with non-foundation-model controls. Not yet run.
- 03 · Relations
How much of a database can automatic features carry into a flat table?
A pipeline that connects to a relational database, discovers keys and timestamps, generates point-in-time features across tables and selects a budget of them for a tabular foundation model, measured against the published benchmarks for the same tasks.
StatusPlan written. Starting with public relational benchmarks.
Tabular models are three years behind the LLM curve, and accelerating faster.
Model releases over time. The tabular timeline is shifted three years to sit on the same axis.
Tabular foundation models: same curve, three years later.
PRAGMA FM
A BERT-style foundation model on longitudinal banking transactions.
LTFM: longitudinal, calibrated, opinionated.
Every version makes the next one smarter.
Deployments show which patterns matter in real tasks. Those become better priors for the next round of pre-training.
25
papers, read & connected before we wrote a line of LTFM.
- PFNPrior-data fitted networks · Müller et al., 2022
- TabPFNTabular foundation model · Hollmann et al., Nature 2025 · acquired by SAP, 2026
- TimesFM · Chronos-2Time-series foundation models · Das et al., 2023 · Ansari et al., 2025
- KumoRFMRelational foundation model · Fey et al., 2025
…and more. We open research intake every 8 weeks. Researchers working on tabular or longitudinal models: write to hello@knoww.ai.
What do you want to predict?
Five runs, 50,000 rows each, free for a month. Tell us the question and what you use today. We review every request by hand and email you when your workspace is ready. If you already know you want a pilot next to your current model, say so below.