SkillTFM Enables Training-Free Adaptation for Tabular Foundation Models

Tabular data is pervasive across many industries, driving predictions and decisions in science, finance, healthcare, and more. Tabular Foundation Models (TFMs) offer a promising paradigm for general-purpose tabular learning, aiming to provide reusable predictors that reduce the need for extensive task-specific training and tuning. However, their practical deployment is often hindered by challenges like distribution shifts, varied feature semantics, and unique task patterns that are difficult to capture without costly fine-tuning or additional labeled data. To address these limitations, researchers propose SkillTFM, a training-free system that redefines TFM adaptation. Instead of updating model parameters, SkillTFM focuses on the gated evolution of "agentic skills." At its core, SkillTFM features a verifiable and extensible skill bank. This bank combines boundary evidence identification, which characterizes task structure and identifies base-model failure patterns, with gated skill evolution, which retrieves and extends reusable skills under explicit validation. Experiments conducted across simulated boundary settings and real-world electricity-price forecasting demonstrate SkillTFM's effectiveness. It improved AUC by 0.128–0.142 and raised nonlinear-boundary AUC from 0.699 to 0.898. Furthermore, its efficacy and generality were confirmed across various TFM backbones, validating the power of skill-based adaptation for tabular data.