Crase replaces the deep-research loop with a bounded citation walk and gets 3x recall@50 at a third of the cost

arXiv·medium signalXBlueskyLinkedInCopy link

Crase (arXiv 2608.24809, 25 August) is a deliberately non-agentic alternative to open-ended deep research agents for scholarly search: it queries a search engine once for seed papers, expands them along their 1.5-hop citation neighborhood, prunes citation edges whose claims lack entailment support, then ranks with a recency-aware random walk. Because the candidate set, the retention reason per paper, and the stopping condition are all fixed before inference, the whole run is inspectable rather than a black-box loop. On LitSearch and a second benchmark over a 500K-paper arXiv corpus it outperforms deep research agents built on proprietary models by up to 3x recall@50 at roughly a third of the cost, which is a direct argument that a bounded pipeline beats an agentic loop when the search structure is known in advance.