ScaffoldAgent: Utility-Guided Dynamic Outline Optimization for Open-Ended Deep Research

arXiv·medium signal

Treats research-report outline evolution as a structured decision process with three operations (Expansion, Contraction, Revision), guided by a utility signal combining retrieval gains, structural coherence, and generation quality. On DeepResearch Bench and DeepResearch Gym it improves long-form report generation and factual grounding over existing deep-research agents. Relevant to builders of deep-research and report-writing agent pipelines.