Pre-trained relational foundation models that predict directly on your multi-table database.
Ask a question in one line - no joins, no feature engineering, no per-task training.
Our first relational foundation model. It reads an entire database - many tables, linked by keys, ordered in time - and answers predictive questions in a single forward pass. It learns each new task from in-context examples, not retraining.
>>> rfm = RFM.connect("postgres://.../retail")
>>> rfm.predict("""
... PREDICT COUNT(orders.*, 0, 90, days) = 0
... FOR EACH customers.customer_id
... """, cutoff="2026-09-01")
customer_id churn_prob
48213 0.87
10977 0.12
66530 0.64
>>> rfm.recommend("products", "FOR customers.customer_id = 48213", k=3)
['SKU-2201', 'SKU-0914', 'SKU-7730']Illustrative preview API.
Mean gain on held-out classification tasks at equal compute
Median gap of 0.45 AUROC to the public RelBench leaderboard
have no published foundation-model results. LS-RFM is built to cover them all, including recommendation
customers ⟶ orders ⟶ order_items ⟶ products ⟶ reviews ⟶ sessions ⟶ returns ⟶ suppliers ⟶ customers ⟶ orders ⟶ order_items ⟶ products ⟶ reviews ⟶ sessions ⟶ returns ⟶ suppliers ⟶ customers ⟶ orders ⟶ order_items ⟶ products
customer_id=48213 order_id=900127 sku_id=2201 store_id=17 order_id=900311 sku_id=0914 visit_id=55120 order_id=901774 sku_id=7730 customer_id=48213 order_id=900127 sku_id=2201 store_id=17 order_id=900311 sku_id=0914 visit_id=55120 order_id=901774
2024-01-03 09:12 2024-02-17 14:40 2024-04-08 11:05 2024-06-21 19:33 2024-09-02 08:01 2025-01-14 16:22 2025-03-30 10:47 2025-07-19 13:58 2026-01-06 18:09 2026-04-25 07:36 2026-08-31 23:59 | cutoff | future rows never seen
One forward pass over the database. No joins, no feature engineering.
Unsorted adjacency lists let future rows into the model's context in 17-97% of lists across RelBench databases. We quantify the effect and release a leak-free protocol.
Masking many cells at once makes the model reason across rows and tables, lifting held-out classification by +8.0 AUROC at equal compute.
Classification, regression, recommendation and autocomplete from one model, with in-context learning of new tasks and changes to the architecture only where the evidence supports them.
We're preparing the audit preprint, the leak-free RelBench evaluation protocol, and LS-RFM checkpoints for public release.
LS-RFM fits into your stack - API, private cloud, or agents. Same foundation model, on your terms.
Customer histories, transactions, clinical records, supply chains - they live in databases of linked tables, not in single spreadsheets. Yet every prediction on them still starts with weeks of joins and feature engineering.
Latent Space is a research lab building foundation models that understand relational data natively: one pre-trained model that answers predictive questions over any database, without per-task training.
We're hiring researchers and engineers who care about rigorous evaluation and real-world data. Get in touch →
Or write to us at surajprasad@latentspaces.in