Crase replaces the open-ended search loop of agents with a bounded, inspectable graph walk. It queries a search engine once for seed papers, expands them within their 1.5-hop citation neighborhood, prunes citation edges whose claims lack entailment support, and ranks survivors with a recency-aware random walk. The candidate set, the reason each paper is kept, and the stopping condition are all fixed before inference.
On LitSearch and a second benchmark over a 500K-paper arXiv corpus, Crase achieves up to 3 times the recall@50 of deep research agents built on proprietary models, at roughly a third of the cost. The design makes evidence auditable: the model only supplies local judgments such as plans, claims, and entailment scores, while the graph structure determines what is searched and when search stops.