# Mitra-v2: Scaling Synthetic Pretraining for Tabular Foundation Models Amazon introduces Mitra-v2, a synthetic-data-trained model for complex tabular prediction tasks. - **Source:** [Amazon Science](https://www.amazon.science) - **Published:** 2026-11-10 - **Topic:** [technology](https://brifing.si/c/technology) - **Original article:** - **On Brifing:** https://brifing.si/s/8feffb93-44e4-4297-999d-d03d5f06ba81 ## Advancing Tabular Machine Learning Mitra-v2 represents the second generation of tabular foundation models designed to eliminate manual feature engineering. Unlike traditional gradient-boosted decision trees that require specific tuning for every new dataset, this model predicts attributes directly from labeled rows. It is now available under an Apache-2.0 license on the Hugging Face platform. ## Scaling Synthetic Data Pretraining The model is pretrained entirely on synthetic tables, ensuring no real customer data is used during the process. Mitra-v2 utilizes tables with ten times as many rows and three times as many columns as its predecessor. This increased scale allows the model to better generalize across diverse, real-world business datasets. ## Hybrid Causal Graph Architecture Mitra-v2 employs a new hybrid SCM prior that builds complex causal graphs. Each relationship within these graphs can represent neural networks, decision trees, or Gaussian processes. This structural diversity enables the model to capture the intricate mix of relationships typically found in actual business data, improving overall prediction accuracy. ## Overcoming Technical Limitations To handle exceptionally wide tables, the system uses singular-value decomposition to compress data while retaining variance. For classification tasks exceeding ten classes, the model routes outputs through a hierarchical tree structure. These lightweight wrappers ensure the model remains functional even when datasets exceed the standard pretraining size limits during deployment. ## Benchmarking Performance Gains Mitra-v2 leads the single-model field on the TabArena benchmark, which evaluates performance across 51 real-world datasets. It outperformed the original Mitra model on 32 out of 33 shared datasets. Furthermore, it consistently meets the strict one-hour computational budget, a milestone the previous version failed to achieve on many tasks. ## Regression and Calibration Capabilities For regression tasks, Mitra-v2 provides prediction intervals alongside point estimates, offering a range of likely outcomes. This approach secured the top rank on ScoringBench, a public benchmark for calibrated regression models. By predicting ranges rather than single numbers, the model provides users with more nuanced insights into potential business events. ## Operational Adoption at Amazon Amazon Connect Customer is integrating Mitra-v2 to enhance its agentic-AI systems. The model estimates the likelihood of business-relevant events based on customer interaction history and operational context. This adoption demonstrates the model's practical utility in automating complex decision-making processes within large-scale, structured business environments and customer experience workflows. ## Future of Foundation Models Mitra-v2 proves that increasing the diversity of synthetic pretraining yields substantial gains without requiring deeper network architectures. The team expects that future iterations will utilize even richer synthetic priors and adaptive mixtures. Researchers and practitioners are invited to access the fine-tuned models and code to test their own data. ## Sources covering this story - https://www.amazon.science/blog/mitra-v2-scaling-synthetic-pretraining-for-tabular-foundation-models — --- _Summary written by Brifing from the original reporting. Copyright in the original article belongs to its publisher; read the full piece at the link above._