Weighing mainstream and alternative accounts…
Two lenses on the same evidence. Source weight and the primary source ratio show what each rests on.
Deeper threads worth pulling on next.
Investigated
Yes, interpolation can be robust to some distribution shifts, but it is not automatically robust. Results depend on shift type and data geometry: transfer can help under favorable conditions, while model shift or robust-risk objectives can make interpolation harmful. Methods that interpolate data distributions or selectively augment examples offer practical ways to improve out-of-distribution performance, but their guarantees and benefits remain conditional rather than universal.
Two lenses on the same evidence. Source weight and the primary source ratio show what each rests on.
Lens adapted to this topic: Why interpolation may fail under distribution shift
A skeptical reading emphasizes that interpolation’s favorable behavior is regime-dependent and can disappear when the test environment differs in the wrong way or when robustness is evaluated against perturbations. The strongest caution is that zero training error does not guarantee robust generalization: even without label noise, regularization can outperform interpolation for robust risk. Other work likewise finds that additional or shifted data can hurt.
Deeper threads worth pulling on next.