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

Abstract: Open-ended deep research (OEDR) requires systems to acquire knowledge through multi-round retrieval and generate coherent long-form reports. The outline plays a central role as a structural scaffold that coordinates retrieval, evidence organization, and generation. However, existing methods either fix the outline before writing or refine it with local heuristics, leading to scaffold drift under continuous information accumulation and delayed feedback for evaluating outline modifications. We propose ScaffoldAgent, a utility-guided dynamic outline optimization framework for OEDR. ScaffoldAgent models outline evolution as a structured decision process with three operations: Expansion, Contraction, and Revision, enabling controlled updates to the report scaffold. It further introduces a utility-guided feedback mechanism that estimates the downstream value of each outline operation from retrieval gain, structural coherence, and trial-generation quality. The resulting utility signal guides node selection, operation scheduling, and termination during inference. Experiments on DeepResearch Bench and DeepResearch Gym show that ScaffoldAgent consistently improves long-form report generation and factual grounding over existing deep research agents.

• The paper introduces a utility-guided structured decision process for dynamically optimizing research outlines, enhancing evidence integration and report coherence.
• It employs three specialized agents (Outline, Search, Reporter) to coordinate expansion, contraction, and revision, ensuring robust and adaptive report synthesis.
• Experimental results show significantly improved report quality and factual grounding with reduced retrieval calls compared to existing baselines.

Utility-Guided Dynamic Outline Optimization for OEDR: Summary and Analysis

Motivation and Problem Formulation

The paper "ScaffoldAgent: Utility-Guided Dynamic Outline Optimization for Open-Ended Deep Research" (2606.20122) introduces a framework designed for Open-Ended Deep Research (OEDR), a task demanding iterative evidence retrieval, hierarchical organizational adaptation, and coherent report synthesis. Existing agentic systems for OEDR either fix the report structure a priori or update it via local heuristics, resulting in structural drift and delayed evaluation of modifications. This work identifies two acute challenges: (1) Uncontrolled scaffold drift—redundancy, uneven granularity, and misaligned sections due to outline evolution that is not globally regulated; (2) Delayed feedback— the value of outline updates is latent, manifesting only after downstream retrieval and writing.

ScaffoldAgent Framework

ScaffoldAgent conceptualizes report outline evolution as a structured decision process. The outline acts as an active scaffold, mediating retrieval, evidence organization, and text generation throughout iterative research. It is manipulated via three explicit operations inspired by AGM belief revision theory: Expansion (topic decomposition and fine-grained subtyping), Contraction (merging redundant branches), and Revision (refreshing weakly grounded nodes while maintaining structural stability). The outline tree evolves from the input question, with nodes representing sections/subsections and accumulating evidence and utility statistics.

Three specialized agents orchestrate the process:

• Outline Agent: central controller, maintains outline tree, selects target node, decides operation.
• Search Agent: retrieves external evidence per update.
• Reporter Agent: performs local trial writing, generates final report after convergence.

Node selection follows a UCB-style rule prioritizing both remediation and exploration. Each update operation is evaluated via a composite utility function: retrieval utility (relevance and novelty), structural utility (coherence, balance, redundancy), and generation utility (citation support, intent coverage, redundancy).

Utility-Driven Feedback and Optimization

After every action, ScaffoldAgent computes local utility as a weighted sum:

$$U_t = \lambda_1 U_{ret} + \lambda_2 U_{str} + \lambda_3 U_{gen}$$

where $U_{ret}$ measures retrieval quality, $U_{str}$ gauges structural soundness, and $U_{gen}$ scores writing potential (citation entailment, coverage, redundancy penalization). Utility stats are updated for the affected node, guiding subsequent action selection and enabling controlled, non-myopic outline evolution. Iterative refinement continues until marginal utility gain falls below threshold, ensuring termination only when the scaffold stabilizes.

Experimental Results

ScaffoldAgent was evaluated on DeepResearch Bench and DeepResearch Gym, benchmarking long-form report quality (RACE: Overall, Comprehensiveness, Insight, Readability, etc.) and factuality (FACT: citation accuracy/rate). Results demonstrate consistently superior performance:

• On DeepResearch Bench with Qwen3-32B backbone, ScaffoldAgent achieved RACE Overall = 44.70 (vs. IRCoT 42.46; +2.24), highest Eff.c. (30.42) and C.acc. (54.32), outperforming both single-agent and multi-agent baselines.
• On DeepSeek-V3.2, ScaffoldAgent scaled to RACE Overall = 48.27, Eff.c. = 51.18, C.acc. = 62.20; monotonic improvement across metrics, robust to backbone choice.
• On DeepResearch Gym, attaining highest average score (75.83), leading in Depth, Breadth, Insight, and nearly matching Supportability (EDR 89.50 vs. ScaffoldAgent 79.80).

Component ablations confirm the criticality of all three actions; removing Contraction or Revision significantly degrades structural quality and factual grounding, and Expansion-only variants cause runaway outline growth. Similarly, utility term ablations reveal Uret as essential for evidence density, Ustr for global organization, and Ugen for downstream report quality.

Multi-turn refinement experiments show ScaffoldAgent's capacity for localized updates in response to user follow-ups, outperforming rewrite-oriented baselines in preserving content, logical consistency, and instruction satisfaction.

Notably, inference cost is reduced: ScaffoldAgent requires less retrieval and fewer search calls than WebWeaver and EDR while maintaining output length and evidence density, indicating efficient utility-guided search concentration.

Practical and Theoretical Implications

Practically, ScaffoldAgent establishes a principled approach for maintaining dynamically evolving outlines during OEDR. The utility-driven framework provides real-time signals that coordinate search, content integration, and structural adaptation, substantially improving report quality and grounding. The explicit decomposition of outline operations aligns with cognitive mechanisms in human research synthesis, but is efficiently executed by agentic LLMs.

Theoretically, modeling outline evolution as a structured decision process with utility-guided actions bridges heuristic refinement and formal sequential optimization. The utility signals unify retrieval, organization, and generation quality, enabling non-local credit assignment for outline modifications. Potential extensions include policy learning for outline evolution via reinforcement learning, hierarchical multi-turn adaptation, and integration with advanced search engines or more capable backbone LLMs.

Limitations and Future Directions

Current benchmarks are primarily single-turn; future work should enhance multi-turn protocols and develop more interactive OEDR scenarios. Evaluation is limited to open-source LLMs and search interfaces—scalability with frontier models and retrieval infrastructures warrants further investigation. ScaffoldAgent currently utilizes utility-based inference-time control; future research could directly learn outline-evolution policies. Longer report trajectories, multi-agent coordination, and task-experience memory offer further avenues for improvement.

Conclusion

ScaffoldAgent represents an authoritative advancement in structured OEDR, providing explicit, utility-guided outline optimization. The integration of Expansion, Contraction, and Revision operations, backed by real-time feedback from retrieval, structure, and generation dimensions, enables precise and adaptive report scaffold control. Experimental results substantiate significant gains in both report quality and factual grounding, positioning ScaffoldAgent as a robust foundation for future research agents in open-ended, complex synthesis tasks (2606.20122).