Frontier research lab · Heidelberg

Putting science
at the core of AI.

We teach AI to learn from first principles — and to adapt when the world changes. One foundation model for dynamic systems, instead of a new model for every task.

Join the lab See the progress
Latest result↗

TipPFN flagged hidden instability in a recorded grid failure two minutes before it became apparent.
01Our bet

Intelligence is not how much you remember. It is how well you handle what you have never seen.

How AI scales today

Today’s models scale by memorising more of the world.

Trained on large, static datasets, they generalise well inside that distribution. Teaching them something new means retraining — costly, slow and brittle.

Our bet

We scale by increasing the space of problems a model can infer over.

A model that understands the mechanics of the system it perceives can predict, plan and act in situations it has never encountered — without constant retraining.

The missing quadrant

General and adaptable.

Most AI is either broad but frozen at training time, or adaptive but built for one system. The hardest problems need both: they are future-facing, ill-defined and come without training data. Their structure has to be discovered — through reasoning and iteration.

01 · Prior

Learn from worlds we understand.

We train a prior-fitted network only on synthetic, custom-built systems whose cause-and-effect we know. No data scraping — and every claim stays testable.

02 · Evidence

Infer from a few observations.

At inference time, the model reads a short record from the system in front of it — recent measurements, episodes, simulations — and infers how this system behaves.

03 · Inference

Answer without retraining.

Calibrated forecasts, early warnings and — increasingly — what-if answers about interventions. When the data in context changes, the answer changes with it.

The scientific question underneath: can a model learn inference itself — so that new problems become a matter of context, not of training?

Read the scientific approach ↗
02Progress

From research vision to proof points in months.

Every result comes with a paper, code where we can share it, the people responsible — and the question it leaves open.

Featured result
Preprint · 2026Early warningCross-domain
Schematic · not data from the paper

A warning before the sudden change.

Can a short measurement record reveal that a system is moving towards a critical transition? TipPFN estimates the distance to a tipping point — in systems it was never trained on.

2min
Earlier than the failure itself

In a recorded grid-failure example, the paper reports hidden instability two minutes before the failure became apparent.

Trained on

Synthetic systems only

Evaluated across

15+ domains, from grid blackouts to early seizure detection

Still open

How far can this ability travel to systems the model has never encountered?

YSKU
Yunus Sevinchan, Kai Ueltzhöffer et al.
Paper ↗Results ↗
NeurIPS 2025Open benchmark

CausalDynamics: what is driving the change?

Test systems whose workings are known, so methods for causal discovery can be checked against the truth.

Still open: can we find those relationships when evidence is incomplete?

BH
Benjamin Herdeanu et al.
Paper ↗Code ↗
Adaptive actionTo confirm

BobPFN: inferring a world model while acting.

10,000×more efficient than reinforcement learning

State-of-the-art control in environments the model has never seen — a first step from prediction to action.

Paper in preparation
Authors · link
CounterfactualsTo confirm

DynaForecastPFN: the effect of a decision, before you make it.

Calibrated forecasts — and how they change when a lever such as price or lead time moves, with confounders held fixed.

In development
Authors · link
All research & publications ↗ kausable on GitHub ↗
03If we are right

From a cost you cut to a capability you gain.

If a model can adapt from context instead of retraining, the consequences reach beyond a better forecast: it changes which problems are practical to tackle at all.

Efficienciesdo today’s work faster
New horizonsdo what was impossible
↓ time to value

Live in days, not months.

No per-case training — the system is configured with data, levers and goals.

↓ maintenance

Always current.

The model adapts the moment the data in context changes. No retraining loop.

↑ coverage

Predict where it was impossible.

Calibrated predictions on sparse, intermittent, small-history data.

↑ new capabilities

Ask what-if.

Causal what-if, tipping-point early warning and decision support — from one model.

Energy & grids

Grid stability

Estimate proximity to instability before visible failure.

First proof · blackout and AMOC transition nowcasts

Industry

Process stability

Detect drift towards degradation before thresholds become reliable.

Research direction · multivariate telemetry

Digital twins

Simulation transfer

Use simulated episodes as context for real-world monitoring.

First proof · sim-to-real transition prediction

Life sciences

Personalised warning

Adapt warning signals to individual dynamics from limited context.

First proof · seizure prediction datasets

Where the work is heading

Results today · research direction · long-term ambition
Now · results

Prediction

Forecasting, early warning of tipping points, simple control.

Next · in research

Counterfactuals

Interventions and planning: what happens if we change this lever?

Ambition

Autonomous decisions

Automated discovery — models that help choose the next most informative experiment.

04Why here

Help define a research direction while it is still open.

kausable was founded in 2025. The questions that will shape this field are unsolved — and the people who join now will own them.

€12M

Seed funding to go deep on the science

2025

Founded — early enough to shape the agenda

15+

Domains where TipPFN has been evaluated

4

Disciplines in one team: physics, mathematics, causal inference, ML

i.

Own an open question.

Take responsibility for a research problem end to end: from the synthetic prior to the model, the paper and its first real application with a design partner.

ii.

Publish and open your work.

Results go to peer review and arXiv, benchmarks and code into the open — so the field can check them.

NeurIPS 2025arXiv 2026GitHub
iii.

Resources to go deep.

A €12M seed round backed by UVC Partners, Entourage, HTGF and Mätch — and a synthetic-data approach that does not depend on scraping the web.

Compute & budgets
iv.

Work across disciplines.

Physicists, mathematicians, causal-inference and machine-learning researchers on one team in Heidelberg — with a direct line from model to application.

“The hardest problems have no training data. Their structure must be discovered.”
What joining looks like
05The researchers

Know whose thinking you’ll be challenging.

Get to know the lab through the work it puts into the world — and the questions it is still working on.

The kausable team in Heidelberg
The lab · Heidelberg
06Principles

Capability and control, designed together.

Flexible

Works with the signals a system actually emits.

Adaptable

Learns from new context without task-specific retraining.

Steerable

People shape behaviour through data, goals and constraints.

Sovereign

Synthetic training and local deployment — no scraped data.

Responsibility · Europe

Control matters alongside capability.

Greater autonomy changes work and responsibility: people need ways to question assumptions, evaluate decisions and set limits. Building these models, methods and the talent behind them in Europe strengthens our ability to make our own technical choices.

€12M seed · backed by
UVC Partners Entourage HTGF. Mätch VC
07Journal

Inside the lab.

All stories ↗
Explainer↗

What does “without retraining” mean?

A closer look at learning from new evidence.
Story↗

Why build worlds to learn from?

Known mechanisms. Questions we can test.
Story↗

Why are rare events so hard to learn?

When the observations we need are hard to find.

Help shape
what comes next.

Bring a hypothesis, a result, a way of building. The next contribution to this research could be yours.

Partner with us ↗Join the lab