Reproduce “Scaling in Games: Continued Pre-Training For Embodied Agents In Diverse Virtual Worlds” 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
Kuan Zhang, Yukun Chen, Zhihao Yang, Yue Su, Xiangnan Wu, Zirong Chen, Run Luo, Jinkun Hou and 3 more
The authors test whether continued pre-training on recorded gameplay helps a vision-language model reuse actions across virtual worlds. They vary model size, gameplay hours, and mixtures of game data across 120 configurations. At 1,600 hours, mixed training averages about 2.8 times Minecraft-only performance on seven unseen related 3D worlds, but gains do not reliably extend to distant games.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.