Preprint
Aug 2026
Training Crossroads for Recurrent Vision Transformers: Recurrence, Neural ODEs, and Deep Supervision
It is found that standard ViTs remain preferable when FLOPs are the primary constraint, whereas recurrent ViTs offer a better accuracy--parameter trade-off under memory constraints, whereas naive recurrence collapses to near-random performance.
Grzegorz Gruszczynski, P. Olszowiec, Michal Byra et al.
· 1 citation