Reproduce “T3R: Training-Free Two-Stage Token Refinement Towards Efficient and Robust VLA Models” 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

Mohammed Hassan, Zhenyang Chen, Zheng Wang, Zhixin Zhu, Tianlong Chen, Yingyan, Chaojian Li

T3R reduces visual tokens before a robot-action decoder and steers attention among those retained. It uses instruction-weighted image-patch similarity and segmentation to prune patches, then biases action queries toward salient instruction words and retained patches. In simulated OpenVLA-OFT LIBERO-Spatial, it prunes 85% of scene-camera tokens, reaches 96% success versus 97% baseline, and lowers per-call latency from 107.5 to 92.5 milliseconds.

Summary by alphaXiv.

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