The Prior Changelog: Summer Edition

Welcome to the first edition of The Prior Changelog. Europe got its first total eclipse since 1999 this month, and we shipped a few things that beat it:

Relational learning just got an upgrade

This week, we released RelArena-α alongside a family of open-source tools for prediction tasks over relational data. You can now:

• Benchmark relational models with RelArena-α. It standardizes data loading, tuning, and evaluation, making results easy to reproduce and compare.
• Apply TabPFN to relational prediction tasks using TabPFN-Rel. As our first relational model harness, it’s still evolving - and we’d love your feedback. Get started with our cookbook.‍

Read about why we’re focusing on relational data and give RelArena a star.

~10x faster inference with KV-cache

Large prediction volumes and interpretability workloads can now run much faster. Pass fit_with_cache when calling fit(), and we save TabPFN’s attention state, removing the need to run the full forward pass on each prediction. Available through both REST API and tabpfn-client. Get started with KV-cache.

See which training rows drive a prediction

To add to our growing interpretability toolkit, we now support attention readout - showing which training rows or features influence a prediction. Best for explaining borderline predictions, investigating unexpected outputs, and seeing how training set affects the results. Get started with interpretability.

These are not all things we’ve shipped - we also released batched regression support, fit speedups, better calibration for uncertainty estimates - and more.

A few other things…

• Wondering how TabPFN applies to your use case? Browse the TabPFN Cookbooks, and contribute your own.
• Doing research with TabPFN? We’ve just featured 100+ scientific publications using TabPFN - check them out.
• If you missed the “Get The Most Out of TabPFN-3” webinar - the recording is live on our YouTube channel.

Read the full Changelog →