Amazon Science introduced Mitra-v2 on 10 November, a model for predicting values in unfamiliar tables, and said it leads the single-model field on TabArena. The release includes classification and regression models, plus inference and fine-tuning code, under the Apache-2.0 licence.
Mitra-v2 learns from synthetic tables rather than real or customer data. Its Transformer reads across both rows and columns, using labelled rows in a new table as examples for predicting unlabelled ones. Amazon retained Mitra-v1’s architecture, while Mitra-v2’s pretraining examples contained 10 times the row count and roughly three times the column count of its predecessor’s. Its synthetic training data also mixes neural networks, decision trees, Gaussian processes and time-dependent processes within causal graphs.
Forecasting demand from a table of past figures could therefore use labelled rows as examples without a separate training run, although Amazon says Mitra-v2 briefly fine-tunes itself on each new table by default.
Amazon reports that Mitra-v2 outperformed Mitra-v1 on 32 of the 33 TabArena datasets both could run natively. It completed all 51 datasets within the benchmark’s one-hour budget, while Mitra-v1 missed that limit on more than a third. TabArena, maintained by an independent academic team, rates methods through head-to-head results on real-world datasets. Amazon Connect Customer, which led development, is adopting Mitra-v2 to predict outcomes from structured business and customer-experience data.
