Xiaomi introduces TabLDM, a tabular large data foundation model for classification and regression via in-context learning that needs no task-specific fine-tuning. The model is pretrained exclusively on synthetic data generated from structural causal models, and the technical report claims a new performance standard.
Two arXiv papers (2026-09-29) present TabFM, a 400M-parameter tabular foundation model doing zero-shot tabular prediction as in-context learning, trained entirely on synthetic tables from structural causal models. Zero-shot TabFM ranks first among default tabular foundation models across all 51 TabArena datasets and beats tuned AutoML pipelines. A companion system, TabFM-Auto, pairs TabFM with an LLM agent that evolves the data pipeline, sweeping the top five TabArena positions and ranking first among MLE agents on MLE-Bench.
Two arXiv papers (2026-09-29) present TabFM, a 400M-parameter tabular foundation model doing zero-shot tabular prediction as in-context learning, trained entirely on synthetic tables from structural causal models. Zero-shot TabFM ranks first among default tabular foundation models across all 51 TabArena datasets and beats tuned AutoML pipelines. A companion system, TabFM-Auto, pairs TabFM with an LLM agent that evolves the data pipeline, sweeping the top five TabArena positions and ranking first among MLE agents on MLE-Bench.
LG AI Research released EXAONE Tabular 1.0, a compact tabular foundation model family for classification and regression via in-context learning, eliminating the need for dataset-specific gradient updates. The model is pretrained exclusively on a synthetic structural-causal-model (SCM) prior and introduces an architecture-centered redesign of tabular in-context learning.
LG AI Research released EXAONE Tabular 1.0, a compact tabular foundation model family for classification and regression via in-context learning, eliminating the need for dataset-specific gradient updates. The model is pretrained exclusively on a synthetic structural-causal-model (SCM) prior and introduces an architecture-centered redesign of tabular in-context learning.
A technical report introduces Mitra-v2, a tabular foundation model claiming state-of-the-art performance on real-world classification and regression tasks including credit-risk scoring, clinical prediction, and house-price estimation. It is trained entirely on synthetic data using a small 2D Transformer backbone with longer contexts and larger feature spaces than Mitra-v1.
A technical report introduces Mitra-v2, a tabular foundation model claiming state-of-the-art performance on real-world classification and regression tasks including credit-risk scoring, clinical prediction, and house-price estimation. It is trained entirely on synthetic data using a small 2D Transformer backbone with longer contexts and larger feature spaces than Mitra-v1.
An arXiv paper introduces SOMTab, a Set-Order Mamba architecture for efficient tabular in-context learning. It challenges attention-heavy tabular foundation models by asking whether attention is needed at every stage of the pipeline. SOMTab separates representation construction from query-conditioned retrieval, offering practitioners a potentially cheaper alternative to attention-based tabular models.
An arXiv paper introduces SOMTab, a Set-Order Mamba architecture for efficient tabular in-context learning. It challenges attention-heavy tabular foundation models by asking whether attention is needed at every stage of the pipeline. SOMTab separates representation construction from query-conditioned retrieval, offering practitioners a potentially cheaper alternative to attention-based tabular models.
GEAR is a two-stage framework that distills tabular foundation models into lightweight MLP or tree-based predictors deployable on commodity CPUs, addressing the latency and memory costs of context-dependent inference. Stage 1 expands coverage with synthetic covariates, and Stage 2 re-anchors the student to real targets.
GEAR is a two-stage framework that distills tabular foundation models into lightweight MLP or tree-based predictors deployable on commodity CPUs, addressing the latency and memory costs of context-dependent inference. Stage 1 expands coverage with synthetic covariates, and Stage 2 re-anchors the student to real targets.