Amazon releases Mitra-v2 for tabular data prediction
In short: Amazon introduced Mitra-v2, a tabular foundation model trained entirely on synthetic data. This version improves upon its predecessor by scaling the pretraining data size and introducing a hybrid structural causal model prior to better reflect real-world data relationships.
This summary was generated automatically by AI from Amazon Science's publication. It is our own text, not a copy of the original — facts, figures and quotes belong to the source, linked above and below.
What changed?
- 1Increased pretraining data: 10x more rows and 3x more columns compared to Mitra-v1.
- 2New hybrid SCM prior combining neural networks, decision trees, Gaussian processes, and autoregressive mechanisms.
- 3Support for hierarchical classification for tasks with more than 10 classes.
- 4Singular-value decomposition wrappers to handle very wide input tables.
| Parameter | Before | Now |
|---|---|---|
| Pretraining rows | Mitra-v1 scale | 10x Mitra-v1 |
| Pretraining columns | Mitra-v1 scale | 3x Mitra-v1 |
Why it matters
Mitra-v2 replaces the need for custom gradient-boosted tree pipelines for every new table, allowing for rapid predictions on unseen structured data. It ranks first in the single-model category on the TabArena benchmark and excels in uncertainty-aware regression.
What it means for AI agents and contact centers
If you are building systems that require automated insights from customer interactions or operational logs stored in tables, Mitra-v2 offers a zero-shot/few-shot approach that could reduce pipeline complexity. Test it against your current models for tasks like propensity scoring or demand forecasting, specifically evaluating the time-to-result vs. accuracy trade-off.
🧪 Worth Testing
It offers a highly efficient way to handle structured data prediction without needing to train separate models for every new dataset.
Sources
- Amazon ScienceOfficialPrimary sourceOriginal article →„Mitra-v2: Scaling synthetic pretraining for tabular foundation models“10 Nov 2026, 17:37
- Published by source
- 10 Nov 2026, 17:37
- Found by our system
- 10 Nov 2026, 17:54
- Summary generated
- 10 Nov 2026, 23:28
This article was written by AI from the original source. Facts, numbers and prices come from the source; missing values are marked “Not specified”. Legal notice, copyright and privacy