Reproduce “Prospector: learning a compact graph traversal policy for scientific reasoning” 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
Haiqian Yang, Markus J. Buehler
Prospector trains a compact model to search knowledge graphs step by step, choosing evidence from earlier tool results rather than receiving a fixed selection. On 195 held-out semiconductor questions, Prospector-4B achieves 0.839 ± 0.015 gated success, compared with 0.877 for gpt-6-astra using the tools. The evaluation covers graph-path questions on four literature-derived graphs, not private corpora or open-ended hypotheses.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.