---
title: "Structurally-Bounded Agentic Graph Exploration for Evidence-Grounded Scholarly DeepSearch"
source: "arXiv AI/CL/LG（论文）"
source_tier: "t2"
url: "https://arxiv.org/abs/2608.24809v1"
permalink: "https://feed.astralor.com/items/itm_4c93fcbfe455e7e2a8f9f854"
published: "2026-08-25T16:51:26.000Z"
score: 76
channel: "ai"
tags: ["arXiv","Crase","Scholarly search","Citation graph","Deep research agent"]
---

# Structurally-Bounded Agentic Graph Exploration for Evidence-Grounded Scholarly DeepSearch

## AI commentary

> What stands out most is that this paper makes 'inspectability' a core design goal, not an after-the-fact explanation. For research scenarios that require strict evidence tracing, this bounded approach may be more practical than chasing all-powerful agents. If you're building literature review tools or evaluating search agent reliability, it's worth examining the pruning and ranking details closely.

## At a glance

A new arXiv paper presents 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 benchmark over a 500K-paper arXiv corpus, Crase outperforms deep research agents built on proprietary models by up to 3x recall@50 at roughly a third of the cost.

## Key points

- Crase uses one-time query plus local graph expansion to avoid open-ended search loops, enabling bounded reasoning.
- By expanding the 1.5-hop citation neighborhood and pruning entailment-unsupported edges, it ensures evidence-grounded candidates.
- On LitSearch and other benchmarks, it outperforms proprietary deep research agents by up to 3x recall@50 at roughly one-third the cost.

## Full text

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.

---

Verify against the original before citing; curated by Astral Feed: https://feed.astralor.com/items/itm_4c93fcbfe455e7e2a8f9f854