What they did

Crase queries a search engine once to retrieve seed papers, then expands the candidate set through their 1.5-hop citation neighborhood. It removes citation edges when the associated claims lack entailment support—meaning the cited evidence does not substantiate the claim—and ranks the remaining papers using a recency-aware random walk over the graph.

The authors evaluated this bounded pipeline on LitSearch and one additional benchmark using a corpus of 500,000 arXiv papers. They compared it with deep research agents built on proprietary language models.

Key findings

• Crase achieved up to 3× higher recall@50, the share of relevant papers recovered among the top 50 results, than the proprietary-model agent baselines.
• It delivered these results at roughly one-third of the reported cost.
• The fixed 1.5-hop expansion provides an explicit candidate set and stopping condition before inference.
• Entailment-based pruning makes the rationale for retaining citation links inspectable rather than leaving selection entirely to an open-ended agent loop.

Why it matters

The results suggest that scholarly retrieval may benefit more from constrained graph exploration and evidence checks than from repeatedly invoking a general-purpose research agent. Such structure can make search behavior easier to audit, reproduce, and budget while retaining strong retrieval coverage.

Caveats

The reported evaluation covers two benchmarks and a 500,000-paper arXiv corpus, so performance on broader, multidisciplinary collections remains uncertain. The abstract reports maximum relative recall and cost gains but not absolute scores, variation across benchmarks, latency, or the contribution of each pipeline component; it also does not establish whether better retrieval yields better final literature syntheses.