Mechanism of the co-evolution advantage

Determine the factor responsible for the co-evolution advantage whereby jointly training memory with the clean signal outperforms post-hoc memory overlay, particularly for the SVHN ResNet setting.

Background

The paper reports that co-training memory together with the clean representation can outperform post-hoc overlay at matched memory load. The advantage is strongest in the SVHN ResNet experiment, where it is approximately +0.050 in validation accuracy, compared with smaller gains for CIFAR and VGG. Although the effect is measured, its determining factor is explicitly left unresolved.

References

What remains open: (i) the residual image term in $C$ ($\approx\pm0.02$ at the $\eta$50 / $K{=}12500$ caliber, beyond the feature-displacement route; Sec.~\ref{sec:C-mechanism}); (ii) determining factor of the co-evolution advantage (SVHN resnet); (iii) breadth --- more architectures/datasets/noise schedules; (iv) the dynamics of within --- why convergence reshapes the local-neighborhood graph (the reshuffle is quantified but not mechanistically explained); (v) the ``chaos region'' ($n{=}3$ cannot distinguish deterministic-effect-plus-s42-anomaly from true chaos).

— Measuring Memory and Generalization as Separable Geometric Channels: The Topo^2 Framework  (2608.30487 - Zhang et al., 31 Aug 2026) in Section Co-evolution advantage; Discussion and Outlook