A Tale of Two Graphs: Separating Knowledge Exploration from Outline Structure for Open-Ended Deep Research
Zhuofan Shi, Dongmei Zhang, Qingwei Lin, Saravan Rajmohan, Dongge Han, Jue Zhang, Fangkai Yang, Zekun Yao, Ming Ma, Victor Rühle
Open MIND
Problems Identified (5)
OEDR evidence accumulation failures: Linear search-then-generate OEDR agents suffer lost-in-the-middle failures as accumulated evidence grows.
Outline-only knowledge gap detection: Outline-centric OEDR planning provides weak supervision for identifying missing relations and triggering targeted exploration.
Long-horizon evidence synthesis: Open-Ended Deep Research requires agents to iteratively search, connect, and synthesize evidence into structured reports.
OEDR evidence accumulation failures: Linear search-then-generate OEDR agents suffer lost-in-the-middle failures as accumulated evidence grows.
Outline-only knowledge gap detection: Outline-centric OEDR planning provides weak supervision for identifying missing relations and triggering targeted exploration.
Proposed Solutions (5)
DualGraph memory: DualGraph memory separates what the agent knows from how it writes by maintaining co-evolving outline and knowledge graphs.
Outline and knowledge graph separation: The method maintains an Outline Graph for structure and a Knowledge Graph as semantic memory storing fine-grained entities, concepts, and relations.
Topology-guided targeted search: DualGraph analyzes knowledge graph topology and outline graph structural signals to generate targeted search queries for iterative exploration and refinement.
DualGraph memory: DualGraph memory separates what the agent knows from how it writes by maintaining co-evolving outline and knowledge graphs.
Outline and knowledge graph separation: The method maintains an Outline Graph for structure and a Knowledge Graph as semantic memory storing fine-grained entities, concepts, and relations.
Results (3)
State-of-the-art benchmark gains:
DeepResearch Bench RACE score:
Dual-graph ablation support:
Research Domain
LLM agents for open-ended deep research