Beyond Outlining: Heterogeneous Recursive Planning for Adaptive Long-form Writing with Language Models

Abstract: Long-form writing agents require flexible integration and interaction across information retrieval, reasoning, and composition. Current approaches rely on predefined workflows and rigid thinking patterns to generate outlines before writing, resulting in constrained adaptability during writing. In this paper we propose WriteHERE, a general agent framework that achieves human-like adaptive writing through recursive task decomposition and dynamic integration of three fundamental task types: retrieval, reasoning, and composition. Our methodology features: 1) a planning mechanism that interleaves recursive task decomposition and execution, eliminating artificial restrictions on writing workflow; and 2) integration of task types that facilitates heterogeneous task decomposition. Evaluations on both fiction writing and technical report generation show that our method consistently outperforms state-of-the-art approaches across all automatic evaluation metrics, demonstrating the effectiveness and broad applicability of our proposed framework. We have publicly released our code and prompts to facilitate further research.

• The paper introduces WriteHERE, a framework that uses heterogeneous recursive planning to adaptively decompose long-form writing tasks.
• It employs a state-based hierarchical task scheduling algorithm to seamlessly integrate information retrieval, reasoning, and text composition.
• Evaluation on technical and narrative datasets demonstrates that WriteHERE outperforms existing methods in enhancing content adaptability and quality.

Adaptive Long-form Writing with WriteHERE

Introduction

The paper "Beyond Outlining: Heterogeneous Recursive Planning for Adaptive Long-form Writing with LLMs" (2503.08275) introduces WriteHERE, a framework designed to facilitate adaptive long-form writing by integrating retrieval, reasoning, and composition tasks. This approach transcends conventional outlining techniques, which often impose rigid writing workflows. WriteHERE instead employs heterogeneous recursive planning to manage task decomposition dynamically, enabling writing agents to adapt fluently to evolving narrative needs.

Figure 1: Illustration of the WriteHERE framework for long-form writing. The core of the framework is a heterogeneous recursive planning mechanism that breaks down complex writing goals into primitive subtasks across three cognitive categories. The process is represented as a Directed Acyclic Graph, where a State-based Hierarchical Task Scheduling algorithm manages the adaptive interleaving of task planning and execution.

Heterogeneous Recursive Planning

The framework incorporates a recursive planning mechanism inspired by Hierarchical Task Network (HTN) planning principles. This method allows for flexible integration of specialized agents and type-aware task decomposition. WriteHERE identifies three cognitive task types essential for writing: retrieval of information, reasoning about content, and composition of text. Tasks are dynamically decomposed into subtasks categorized under these types, ensuring a comprehensive approach to long-form writing adaptation.

Figure 2: The abstract flow of tasks. The arrow indicates the information flow of a task: the system state at the arrowhead is modified by the labeled task, while the hollow circle end signifies that the associated system state remains unchanged.

Implementation and Evaluation

WriteHERE's ability to handle diverse long-form writing tasks was tested by implementing it for both technical report generation and narrative writing. The framework was evaluated using the Tell me a story dataset for fiction writing and the Wildseed dataset for structured document generation. The experiments demonstrated significant improvements in content quality and adaptability over existing methods like Agent's Room and STORM, showcasing the efficacy of WriteHERE in producing higher-quality and more adaptable narrative outputs.

Figure 3: The evaluation results of WriteHERE v.s. Agent's Room at different generation lengths.

Contributions and Implications

The key contributions of the paper include the novel integration of heterogeneous recursive planning for task decomposition and the state-based hierarchical task scheduling algorithm that efficiently manages adaptive execution. By addressing the limitations of fixed task sequences in long-form writing, WriteHERE enhances the ability of writing agents to dynamically adapt their approach in response to new information and evolving narrative requirements.

The implications of this research are substantial for the field of AI-driven content generation. Practically, WriteHERE can be applied to areas requiring complex document creation, such as technical writing, storytelling, and academic research. Theoretically, it suggests new pathways for developing multi-agent systems capable of deeper integration of cognitive processes, potentially informing future advances in AI systems that require nuanced decision-making and adaptability.

Conclusion

WriteHERE advances the capabilities of LLMs in long-form writing through its innovative heterogeneous recursive planning and dynamic task scheduling approach. The framework's demonstrated effectiveness in various writing domains suggests promising directions for future research, particularly in enhancing the adaptability and contextual integration of AI writing agents. As content demands continue to grow in complexity, frameworks like WriteHERE will be critical in meeting these challenges, paving the way for more intelligent and adaptable LLM applications.