Reproduce “Bring Back Language to Actions: Restoring VLM capabilities in VLA Models via Parameter Merging” 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
Mohamed Salim Aissi, Olivier Sigaud, Laure Soulier, Nicolas Thome
The authors introduce a post-hoc method that merges a vision-language model with its action-trained version, learning coefficients for individual weight-change directions. Compared with AdaMerging, it scores 29.5 ROUGE-L on COCO versus 28.1, and 89.3, 88.1, and 93.3% success across LIBERO suites versus 6.2, 2.9, and 3.7%. These results cover tested simulation tasks, not physical-robot deployment.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.