Reproduce “JevSpawn: Adaptive Agentic Inference through Compositional Action Spaces” 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
Haoyang Su, Weiran Huang
JevSpawn lets LLM agents infer and revise finite choices for actions from task instructions and interaction, rather than fixing them in advance. It combines reusable action fields into alternatives, scores them in parallel, and uses environment observations to select branches or revise the choices. On Maze, it reaches 0.96 success versus FoldAgent’s 0.52, with 40.91-second mean latency; this is one navigation task.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.