Recurrent-Depth VLA: Implicit Test-Time Compute Scaling of Vision-Language-Action Models via Latent Iterative Reasoning
IntermediateYalcin Tur, Jalal Naghiyev et al.Feb 8arXiv
Robots often use the same amount of thinking for easy and hard moves, which wastes time on easy steps and isn’t enough for tricky ones.
#Recurrent depth#Latent iterative reasoning#Vision-Language-Action