Spurious Signal Routing in Tabular In-Context Learners
A new paper on arXiv (2607.25532) formalizes spurious routing in composite representations for tabular in-context learning (ICL) models. The authors show that when a feature bundles a true causal signal with a systematic artefact (e.g., from hospital equipment), ICL models like TabPFN may route predictions through the spurious signal, failing when deployed in new environments. They prove this routing is unavoidable under ridge ICL regardless of context size.
Key facts
- arXiv paper 2607.25532
- Formalizes spurious routing in composite representations
- Feature X bundles causal signal C and spurious signal S
- ICL models route through S instead of C
- Fails silently at new hospitals with different equipment
- Proven unavoidable under ridge ICL
- TabPFN shows consistent behavior
- Published on arXiv
Entities
Institutions
- arXiv