Reproduce “VLA-Retain: Benchmarking Cross-Dimension Effects of Vision Tuning in VLAs” 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
Borui Xu, Han Chen, Bingsheng He, Ming Wu, Jason Zeng, Michael Heinrich
The authors introduce VLA-Retain, a paired benchmark of how visual tuning changes VLA performance across conditions. They compare five models with their own base checkpoints on LIBERO-Plus episodes, covering camera, sensor-noise, and perturbations. Camera-view tuning raises X-VLA’s target success by 69.4 percentage points, but its mean off-target gain is −4.6; effects vary across models and do not establish one cause.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.