Spurious Signal Routing in Tabular In-Context Learners
July 29, 2026
Research identifies that in-context learners like TabPFN struggle to distinguish between causal signals and spurious artifacts when they are bundled within the same feature. The study proves that for ridge in-context learners, this routing error is unavoidable regardless of the context size provided.
HOW THIS AFFECTS YOU
●
builderThis warns you that increasing context size will not solve feature entanglement issues in tabular models.
●
researcherYou should account for unavoidable spurious signal routing when evaluating tabular ICL performance.