Structurally-bounded Agentic Graph Exploration for Evidence-Grounded Scholarly DeepSearch
Crase narrows scholarly search by fixing the graph, pruning unsupported citation links, and stopping before inference begins.
The paper presents Crase as an inspectable alternative to open-ended deep research agents. It starts with one search-engine query for seed papers, expands through a 1.5-hop citation neighborhood, removes citation edges without entailment support, then ranks papers with a recency-aware random walk. On LitSearch and another benchmark over a 500,000-paper arXiv corpus, the authors report up to 3x recall@50 versus proprietary-model deep research agents at about one-third the cost. ArXiv · AI/CL/LG's note
The paper presents Crase as an inspectable alternative to open-ended deep research agents. It starts with one search-engine query for seed papers, expands through a 1.5-hop citation neighborhood, removes citation edges without entailment support, then ranks papers with a recency-aware random walk. On LitSearch and another benchmark over a 500,000-paper arXiv corpus, the authors report up to 3x recall@50 versus proprietary-model deep research agents at about one-third the cost. ArXiv · AI/CL/LG's note
score 5