Agentic search over large document collections often misses evidence spread across multiple documents because each document is treated as an isolated file. CorpusMap addresses this by building an offline navigation layer that recognizes recurring entities, such as people, projects, or products, and connects every document that mentions them through entity pages. At query time an agent follows links between entities and documents to collect complementary evidence instead of rediscovering relationships by repeatedly searching. Across 7 models and 3 benchmark datasets, CorpusMap improved evidence discovery and answer quality over raw-corpus agentic search while using fewer tokens on average, and it also outperformed 4 alternative navigation layers. The result is notable because it suggests that reusable, query-independent structure can make agentic search both more effective and cheaper, rather than relying on more retrieval steps at inference time.
Entity maps help search agents connect evidence across documents
CorpusMap links recurring entities to their source documents, improving evidence discovery and answer quality while reducing average token use.
Big Tech
Soyeong Jeong · Sujay Kumar Jauhar · Sung Ju Hwang · Andrew Joohun Nam
KAIST · Microsoft
Research Digest··3 min read
In one line
Organizing a document corpus into entity pages linked to source documents improves agentic search evidence discovery and answer quality while reducing token use.
Before you read this
Two ideas this paper builds onagentic scaffoldingExternal supports like training curricula, tool selection, and success verifiers enable agents to handle long-horizon or out-of-distribution tasks.33 papers →evidence retrievalPulling specific, relevant information from a large corpus to support an answer, decision, or reasoning step.115 papers →
Read it as
Jeong et al. introduce CorpusMap, an offline-built navigation layer for agents searching large, heterogeneous document collections.
The authors extract recurring entities from a document collection, resolve mentions that refer to the same entity, and create a source-grounded page for each one.
How it works
before and afterThe Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.