Why it is a question
THEORY-246 holds what Kornblith et al. (LIT-1044) measured. Among ResNet-50s trained on ImageNet with nine objectives and early-stopped, fixed penultimate features with more class separation transfer worse to datasets whose classes are mostly not ImageNet’s (Spearman ρ = −0.93 across the nine). One temperature knob moves both, and full fine-tuning removes most of the gap. Two things that study cannot settle were THEORY-246’s promote_when until 2026-10-20. One is generality: it is one pretraining set, two architectures, and networks stopped near 92% training accuracy. The other is mechanism: across losses the relation is a correlation of nine points, and no experiment holds separation fixed while changing the loss. The authors themselves expect a larger, richer label space might soften or reverse the trade (§5).
What would answer it
The condition as it stood, unchanged:
- For generality. The same comparison, fixed-feature readout of new-class tasks against penultimate class separation, in a pretraining setting LIT-1044 did not test: a much larger label space, a transformer, or a network trained past zero training error, with the inverse relation holding.
- For the mechanism. An intervention that changes within-class spread in the last layers with the loss held fixed, or holds it fixed while the loss changes, and shows new-class transfer moving with the spread and not with the loss.
Against it. At matched class separation, different losses giving clearly different new-class transfer, or, with a large and diverse pretraining label space, more separation ceasing to cost new-class transfer.
What cannot settle it. More losses on ResNet-50 and ImageNet repeat the measured correlation. More same-class shift benchmarks answer a different question (THEORY-104).