
See What Matters: Differentiable Grid Sample Pruning for Generalizable Vision-Language-Action Model
ICML 2026 CCF-A · CORE A* · First author · Accepted
A plug-and-play differentiable grid sampler for VLA models. With only 16 visual tokens it reduces VLA FLOPs by about 76%, while improving LIBERO success and real-robot performance.
Read paper ↗ · Paper page ↗





