1st Workshop on Trustworthiness of Tabular Foundation Models (TrustTFM)

Overview

Tabular foundation models (TFMs) are becoming an increasingly important approach for learning from structured data. Models such as TabPFN and TabICL can make predictions on new datasets with little or no task-specific training, often by conditioning on a set of examples at inference time. This makes them attractive for many domains where tabular data remains central, including healthcare, finance, cybersecurity, and the natural and social sciences.

At the same time, TFMs raise new questions about trustworthiness that are still poorly understood. Their predictions may depend not only on a query point, but also on the inference-time context and on priors learned during pretraining. This creates failure modes that differ from those of conventional tabular models. Recent work has started to study adversarial robustness, privacy, distribution shift, noisy or manipulated contexts, fairness, calibration, and model invariances, but these directions are still relatively fragmented.

The goal of TrustTFM is to bring together researchers working on these questions and to help structure the emerging area of trustworthy TFMs. We are particularly interested in understanding which problems are inherited from classical tabular machine learning, and which arise specifically from foundation-model pretraining and in-context inference.

Within SaTML, TrustTFM will provide a focused venue for discussing the specific trustworthiness challenges raised by tabular foundation models. We will hold a half-day workshop combining an invited keynote, contributed presentations or spotlight talks, a panel, and an interactive discussion on open problems, with the goal of identifying missing benchmarks, useful threat models, and promising directions for future work.