Prospector makes scientific graph reasoning local
MIT’s Prospector is a post-trained compact model that traverses scientific knowledge graphs with find, expand, and connect tools, producing verifiable evidence chains. Trained on semiconductor literature, it transfers to three other domains and runs privately on a single GPU.
Prospector’s real breakthrough is systems efficiency: it turns graph navigation into a learnable policy instead of outsourcing every reasoning step to a frontier model.
- –Reaches near-frontier accuracy on the semiconductor benchmark while using a fraction of hosted-model cost
- –Keeps proprietary graphs on local hardware, with reported support for GPU and Mac-based offline serving
- –Frozen weights transfer to bio-inspired materials, PFAS, and materials-property graphs
- –The released code, adapters, graphs, and evaluation harness make the approach unusually reproducible
- –Results remain bounded by LLM-extracted graphs and a citation gate that checks retrieved-edge membership more than full semantic correctness
DISCOVERED
1h ago
2026-10-30
PUBLISHED
1h ago
2026-10-30
RELEVANCE
AUTHOR
Discover AI