MIT’s Prospector achieved a gated success rate of 0.839 on 195 held-out semiconductor questions, against GPT-6 Astra’s 0.877 and Claude Opus 5.5’s 0.846, with the same graph-search tools. The findings appear in a 27 October preprint.[1]
The four-billion-parameter model learns to locate entities, expand connections and join paths through a graph extracted from papers. Each search attempt had a budget of 20 tool calls and 400 retrieved connections. Without retraining, its search policy also transferred to bio-inspired materials, PFAS and materials-property graphs, scoring 0.766 on the PFAS questions.[1]
Selecting among 32 sampled attempts with a trained verifier raised the success rate to 0.913, using additional inference work. Questions and accepted answers came from the graph. Scoring requires a correct answer within cited connections, all retrieved during the attempt; the graph’s own accuracy remains a separate question.[1]
Graph-based retrieval already serves a different purpose in Microsoft’s earlier GraphRAG research. That system extracts entities and relationships from documents, groups connected entities into communities and prepares summaries of those groups. When a question arrives, it generates partial answers from the summaries and combines them into a response.[2]
The design addresses questions about broad themes across a document collection. The GraphRAG paper evaluates how comprehensive and varied those answers are, using model judges; it does not use a fixed answer set for scoring. Consequently, the useful comparison is architectural, rather than a shared leaderboard: some questions need a broad overview prepared from many sources, while others need a precise route through connected facts.[2]
Sources & further reading
- Prospector: learning a compact graph traversal policy for scientific reasoning Yang and Buehler / arXiv · 27 Oct 2026 · Accessed 29 Oct 2026
- From Local to Global: A GraphRAG Approach to Query-Focused Summarization Microsoft researchers / arXiv · 19 Feb 2025 · Accessed 29 Oct 2026