Structurally-bounded Agentic Graph Exploration for Evidence-Grounded Scholarly DeepSearch
2026-08-25 • Computation and Language
Computation and LanguageInformation Retrieval
AI summaryⓘ
The authors introduce Crase, a new tool for academic paper search that works differently from typical deep learning agents. Instead of searching endlessly, Crase first finds a few important papers, then looks at related papers nearby in the citation network. It removes connections that don’t have clear support and ranks papers based on how recent they are. Their tests showed Crase finds more relevant papers more efficiently and cheaply than some existing deep research tools. This approach makes the decision process understandable and fixed before it starts searching.
scholarly searchcitation networkrandom walkrecency-aware rankingrecall@50arXiv corpusdeep research agentscitation pruningentailment support
Authors
Rima Hazra, Sayan Layek, Somnath Banerjee, Soumen Chakrabarti, Animesh Mukherjee
Abstract
We present Crase, a bounded and inspectable alternative to deep research agents for scholarly search. Instead of an open-ended search loop, Crase 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, and ranks the remaining papers with a recency-aware random walk. This makes the candidate set, the reason each paper is kept, and the stopping condition explicit and fixed before inference. On LitSearch and one further benchmarks over a 500K-paper arXiv corpus, Crase outperforms deep research agents built on proprietary models by up to 3$\times$ recall@50 at roughly a third of the cost.