Reproduce “Self-Improving Agents with World Model Predictive Paths” 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
Haori Lu, Tianyu Gao, Yaoyao Liu, Yingying Li
The method helps long-horizon language agents compare actions before acting: it predicts one-step outcomes for up to eight candidates, scores progress and risk, and uses cost-based weights to guide action selection. Retrieved planner and world-model memories provide context while model parameters remain frozen. On ALFWorld with Gemma-4-26B-A4B, it reaches 65.7% SR@1 versus 32.8% for ITP-I; tests cover two simulated environments.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.