SkillTFM Adapts Tabular Foundation Models With Zero Parameter Updates, Lifting Nonlinear-Boundary AUC From 0.699 to 0.898

arXiv 2608.06137·medium signalXBlueskyLinkedInCopy link

SkillTFM (2608.06137, submitted 2026-08-06) shifts tabular-foundation-model adaptation from fine-tuning to a training-free, verifiable skill bank: boundary evidence identification characterizes task structure and where the base model fails, then gated skill evolution retrieves and extends reusable skills only when they pass explicit validation. Across simulated boundary settings and real-world electricity-price forecasting it improves AUC by 0.128-0.142 and raises nonlinear-boundary AUC from 0.699 to 0.898, with the gains holding across different TFM backbones. The practical read is that agentic skill-bank machinery is now showing up outside coding agents — as a substitute for costly fine-tuning under distribution shift.