Reproduce “SATOR: Backpropagation-Free 3D Test-Time Adaptation via Sample-Adaptive Token Routing” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Inseok Kong, Geunyoung Jung, Jiyoung Jung

SATOR adapts 3D point-cloud classifiers at test time by scoring local patches through attention and geometric relationships, then setting a sample-specific pruning threshold and routing between original and pruned predictions. On ScanObjectNN-C, a real-scan-derived benchmark, it achieves 49.43% mean top-1 accuracy versus 46.78% for Purge-Gate, beating it on 13 of 15 corruption types without establishing broader real-world robustness.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.