Tabular data—data organised in spreadsheets in rows and columns, relational databases, or time series—remains the most widespread and economically most valuable data modality in the world. It underpins decision-making across nearly every sector—finance, insurance, healthcare, energy, manufacturing, logistics, and the public sector—where structured records of entities, transactions, and measurements are the norm. It represents a trillion dollar market.
This proposal aims at no less than entirely revolutionising the field of tabular data science, leading the foundation model revolution in this massive and underexplored market. Its success will determine whether the German startup Prior Labs (www.priorlabs.ai), currently leading the worldwide research effort on tabular foundation models, will be able to defend its lead against extremely well-funded US and Chinese companies. The company aims to shape this field with European values, focusing on trustworthy methods, making causality, explainability, and robust out-of-distribution generalisation first-class citizens.
Prior Labs' TabPFNv2 started this revolution. Even though at its release in late January 2025 it was limited to prediction on tables with less than 10,000 rows, it has already been downloaded 2 million times, and been cited by over 550 scientific papers, with >100 use cases in Healthcare, Finance, and Energy. In November 2025, we released TabPFN-2.5, pushing state-of-the-art performance to 50,000 rows. However, to scale to millions of rows and complex relational tasks, a "Production Scale" compute budget is required.
We now detail Prior Labs' ambitious research agenda for 2026, executing 10 Work Packages (organised into four pillars):
Pillar 1: Additional modalities. Extending TabPFN to Time Series (WP 1), Relational Databases (WP 2), and Semantic Problem Understanding by Multimodal Text-Tabular Models (WP 3).
Pillar 2: Core improvements. Linear attention for handling long contexts natively (WP 4), deriving 4D scaling laws (WP 5), and developing a single generalist foundation model for many different varieties of tasks (WP 6).
Pillar 3: Trustworthiness. Enforcing counterfactual fairness (WP 7) and native interpretability via amortized SHAP (WP 8) to be compliant with the EU AI Act by design.
Pillar 4: Agents. Using test-time compute to wrap a model with retrieval, fine-tuning, and ensembling (WP 9) and enabling autonomous Agentic Data Science using TabPFN as a fast critic (WP 10).
This project is also in line with the HORIZON Action ELLIOT, where Prior Labs leads tabular and time series foundation model building. As discussed below, the outcomes will unlock thousands of high-impact use cases and enable EU leadership in the trillion dollar AI enterprise economy.
Principal Investigator, Company and Country
Frank Hutter, Prior Labs, Germany