ScaffoldAgent: Dynamic Outline Optimization for Deep Research

ScaffoldAgent models outline evolution as a structured decision process and applies a utility-guided feedback signal to steer outline updates for open-ended deep research. The paper, submitted on 18 Jun 2026 (arXiv:2606.20122) by Zhibang Yang and 10 other authors, proposes three explicit operations on a report scaffold and demonstrates gains on DeepResearch Bench and DeepResearch Gym.

What is ScaffoldAgent?

ScaffoldAgent is a framework that treats an outline as a mutable scaffold and optimizes it through three operations: Expansion, Contraction, and Revision. The system pairs those operations with a utility-guided feedback mechanism that estimates the downstream value of each operation from retrieval gain, structural coherence, and trial-generation quality.

The paper frames outline evolution as a structured decision process, rather than fixing the outline up front or relying on local heuristics. The authors argue this avoids scaffold drift under continuous information accumulation and addresses delayed feedback when evaluating outline edits.

How does ScaffoldAgent work?

ScaffoldAgent applies three controlled update operations to outline nodes and uses a utility signal to choose which nodes to change, when to schedule operations, and when to stop. The utility signal aggregates three components: retrieval gain, structural coherence, and trial-generation quality, and it guides node selection, operation scheduling, and termination during inference.

Concretely, the framework interleaves retrieval, outline modification, and trial generation. Retrieval supplies new evidence; Expansion, Contraction, and Revision alter the scaffold; trial-generation produces sample text used to estimate the quality impact of each change. The estimated downstream value becomes the utility that directs further updates.

What evidence supports its claims?

The authors report experiments on two evaluation suites, DeepResearch Bench and DeepResearch Gym, and state that ScaffoldAgent "consistently improves long-form report generation and factual grounding over existing deep research agents." The paper is nine pages long and includes six figures illustrating the method and results. The manuscript is archived as arXiv:2606.20122 and includes a DOI link: https://doi.org/10.48550/arXiv.2606.20122.

The submission lists Zhibang Yang as first author and ten coauthors, and was submitted on 18 Jun 2026. The experimental claim centers on stronger factual grounding and improved long-form coherence compared with prior agents designed for open-ended deep research.

Why it matters

ScaffoldAgent changes how an agent treats structure during long-form research. By converting outline edits into a decision process and scoring edits with a utility estimate tied to retrieval and generation outcomes, the system explicitly connects information acquisition to document structure. That makes outline updates measurable rather than heuristic and reduces the risk of drift when agents keep ingesting new sources.

For practitioners building long-form research assistants, the approach offers a clear lever: optimize the utility estimator and the operation scheduler to prefer outline changes that demonstrably improve retrieval payoff and generated text quality. For evaluators, the emphasis on benchmarked improvements on DeepResearch Bench and DeepResearch Gym gives a concrete test bed to compare future methods.

What to watch

Look for the code, data, and replication assets the authors link from the arXiv entry and for follow-up work testing utility estimation choices. The next confirmatory signals will be public code releases or independent reproductions of the reported improvements on DeepResearch Bench and DeepResearch Gym.