Reproduce “RoleWorld: Knowledge Transfer in World Modeling with Functional Role Abstractions” 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
Boyuan Xiao, Pu Wang, Ji Feng, Yao-Xiang Ding
RoleWorld transfers interaction knowledge to changed worlds by separating behavioral role prototypes from object assignments and parameters. A teacher LLM builds prototypes, fixed transition rules, and an actor skill from a source world; in the target, feedback updates assignments and parameter estimates. On seven MARS variants, GPT-5.6 Luna RoleWorld averaged 8.1 reward versus WorldCoder’s 2.7 under the tested transfer protocol.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.