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

ScaffoldAgent introduces a dynamic outline optimization framework for open-ended deep research that evolves report structures through expansion, contraction, and revision operations. The system uses utility-guided feedback mechanisms to evaluate outline modifications based on retrieval gains and coherence, demonstrating improved performance on deep research benchmarks compared to existing approaches.

ScaffoldAgent addresses a fundamental challenge in automated research systems: maintaining coherent report structures while continuously acquiring new information. Traditional approaches either lock outlines before writing or apply ad-hoc refinements, creating structural inconsistencies as evidence accumulates. This research proposes treating outline evolution as a structured decision problem with three discrete operations, enabling systematic scaffold updates that prevent drift and maintain report coherence.

The framework's innovation lies in its utility-guided feedback mechanism, which evaluates potential outline modifications by estimating downstream value across multiple dimensions: retrieval gain indicates how well new information supports existing claims, structural coherence ensures logical flow, and trial-generation quality measures practical writing effectiveness. This multi-dimensional evaluation prevents the myopic local optimizations that plague earlier systems.

For the broader AI research community, ScaffoldAgent demonstrates how explicit structural scaffolding improves complex reasoning systems. Rather than treating outline development as incidental to report writing, this work elevates it to a first-class optimization problem. The framework's applicability extends beyond academic research to technical documentation, investigative journalism, and enterprise knowledge synthesis—domains requiring coherent long-form synthesis from diverse sources.

The experimental validation on DeepResearch Bench and DeepResearch Gym shows measurable improvements in both generation quality and factual grounding, suggesting the approach addresses real constraints in current systems. Future development should explore how utility-guided optimization scales to papers requiring hundreds of sources and whether the framework adapts to different domain-specific coherence requirements.

• →ScaffoldAgent treats outline evolution as a structured optimization problem with three operations: expansion, contraction, and revision.
• →Utility-guided feedback mechanisms evaluate outline modifications across retrieval gain, structural coherence, and generation quality dimensions.
• →The framework prevents scaffold drift that occurs when information accumulates during multi-round retrieval processes.
• →Experimental results demonstrate consistent improvements in long-form report quality and factual accuracy over existing deep research agents.
• →The approach has applications beyond academic research to technical documentation, journalism, and enterprise knowledge synthesis.