Logic: Long-form Outline Generation via Imitative and Critical Self-refinement

H. Liu, Yongliang Shen, Zhe Zheng, Hongjun Ma, Xingyu Wu, Yin Zhang⋆, Weiming Lü2025 · DOI

Introduction

Generating high-quality, detailed outlines is a critical precursor to successful long-form expository writing, such as Wikipedia entries. While Large Language Models (LLMs) excel at short-form content, they often struggle with the structural complexity and information density required for longer documents. Existing approaches frequently suffer from a "knowledge density gap," where outputs lack fine-grained detail, or a "retrieval-coherence dilemma," where facts are retrieved but lack logical flow.

To address these challenges, we introduce LOGIC, a novel framework that mimics the human writing process through imitative structural learning and iterative critical self-refinement.

Methodology: The LOGIC Pipeline

The LOGIC system employs a four-stage pipeline designed to ensure both logical coherence and information depth:

  1. Serial Planning: Rather than generating an entire outline in one pass, LOGIC drafts sections serially. This ensures each section maintains a continuous narrative thread by being aware of preceding content.
  2. Modular Premise Generation: The system establishes a "premise" across four dimensions—definition, background, events, and significance—serving as a structural "North Star."
  3. Imitative Retrieval: A key innovation where the model retrieves the outlines of related entities. By analyzing successful structures of similar topics, the model "imitates" professional organization.
  4. Critical Self-Refinement: A dedicated Critic Model provides specific feedback on drafts, which a Generation Model then uses to iteratively polish and deepen the content.
The LOGIC framework combines imitative retrieval of related entity structures with a critic-refiner loop for iterative improvement.

Experimental Evaluation

LOGIC was evaluated using two primary benchmarks:

  • FreshWiki: A dataset of recent, high-quality Wikipedia articles.
  • WikiOutline: A new, diverse dataset curated by the authors containing 80 "Good Article" status pages across categories like Characters, Events, Films, Disasters, and Places.

Key Results

LOGIC demonstrates significant improvements over state-of-the-art baselines such as STORM and OmniThink. Quantitative analysis reveals that LOGIC enhances organization scores by 22.85% and logicality by 16.19% on the FreshWiki dataset.

Comparison of outline structures and metrics across different generation methods. LOGIC (E) achieves the highest scores in organization, relevance, and depth.

The "Depth" metric saw a substantial increase of 54.29% over direct generation, indicating that the refinement process effectively adds necessary detail. Human evaluators consistently preferred LOGIC’s outlines, citing superior knowledge density and a more expert-like tone.

Conclusion and Future Work

LOGIC successfully bridges the gap between raw fact retrieval and coherent long-form planning by prioritizing structural imitation and iterative refinement. While the process is computationally more intensive than single-pass generation, it allows smaller models (like Llama-3-8B) to rival the performance of much larger models like GPT-4o. Future work will explore Reinforcement Learning (RL) to further optimize efficiency and address challenges in maintaining reader interest and objectivity.

Summary updated

BibTeX
@misc{liu2025logic,
  author = {H. Liu and Yongliang Shen and Zhe Zheng and Hongjun Ma and Xingyu Wu and Yin Zhang⋆ and Weiming Lü},
  title = {{Logic: Long-form Outline Generation via Imitative and Critical Self-refinement}},
  year = {2025},
  doi = {10.18653/v1/2025.findings-emnlp.983},
}
Figure