ScaffoldAgent: Utility-Guided Dynamic Outline Optimization for Open-Ended Deep Research
Open-ended deep research involves using AI to gather information through multiple rounds of searching and synthesizing that data into a long, coherent report. A major challenge in this process is maintaining a logical structure—or "scaffold"—for the report as new information is discovered. Existing systems often struggle because they either lock in an outline too early or update it using simple rules that don't account for the overall quality of the report. ScaffoldAgent is a new framework designed to solve this by treating the outline as a dynamic, evolving structure that is continuously optimized based on how well it supports the final research goals.
Managing the Evolving Outline
Instead of a static plan, ScaffoldAgent treats the report outline as a living tree that grows and changes. To keep this structure organized, the system uses three specific operations: Expansion, which breaks down broad topics into more detailed sub-sections; Contraction, which merges redundant or overlapping branches to keep the report concise; and Revision, which updates specific sections that are poorly supported by evidence. By using these three tools, the agent can adapt the report’s structure as it learns more, preventing the "scaffold drift" that often causes reports to become disorganized or repetitive.
Utility-Guided Decision Making
A key hurdle in deep research is knowing whether a change to an outline is actually an improvement, as the impact of a structural change might not be clear until the final report is written. ScaffoldAgent addresses this with a "utility-guided" feedback mechanism. After every change, the system calculates a utility score based on three factors: how much new, relevant information was retrieved; how coherent and balanced the new structure is; and how well the current outline supports the actual writing of the report. This score acts as a guide, helping the agent decide which parts of the outline need more work and when the research is complete and ready for final generation.
Improved Research Performance
Experiments conducted on established benchmarks, such as DeepResearch Bench and DeepResearch Gym, demonstrate that ScaffoldAgent consistently outperforms existing research agents. By unifying retrieval, structural organization, and trial-generation quality into a single decision-making process, the system produces reports that are not only more comprehensive and readable but also better grounded in factual evidence. This approach proves that treating the outline as an active, optimized component of the research process is more effective than relying on fixed plans or local, isolated updates.
Comments (0)
to join the discussion
No comments yet
Be the first to share your thoughts!