Reproduce “Octopus: One Brain, Many Branches for Latent-KV Multi-Agent Reinforcement Learning” 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

Tianming Sha, Junnan Li, Siqi Zhao

Octopus organizes long-horizon search as GPU-resident branches that share cached key/value representations (KV), roll back dead ends, and read sibling findings without copying text. Each branch retains a 32K-token growth budget. On BrowseComp-Plus, it reports 49.6% accuracy versus 32.1% for a same-size 143K single-agent baseline; strict counting yields 49.3% across 830 queries. Results concern this sparse-BM25 evaluation.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.