Arc researcher disputes interpolation claim: optimized for graceful degradation across cell lines
kjaganatha · x · 2026-11-20
kjaganatha responds to Anshul Kundaje's probing questions, denying any claim of invariance to train-test similarity and calling it a straw man. He reiterates that they explicitly optimized for graceful performance degradation via careful holdout splits across roughly 56 cell lines. Largely overlapping with the prior post, this is the opposing side's view in the same academic dispute.