TLDR: Meow is a novel, metadata-driven framework that uses a two-stage training approach (supervised fine-tuning and reinforcement learning) with Large Language Models to generate high-quality, hierarchical academic survey outlines end-to-end. It outperforms existing methods in structural fidelity and stylistic coherence, offering an efficient solution for automated literature reviews.
In the rapidly expanding world of academic research, keeping up with the sheer volume of new publications has become a monumental task. Researchers often spend countless hours conducting literature reviews and creating survey papers to summarize existing knowledge. Recognizing this challenge, a new framework called Meow has been introduced to automate the crucial step of outline writing for academic surveys.
Meow, which stands for “metadata-driven outline writing,” is the first of its kind to approach outline generation as an end-to-end task. Unlike previous methods that treated outline writing as a minor step in a larger workflow, Meow focuses on creating organized and accurate hierarchical outlines directly from paper metadata. This means it takes information like titles, abstracts, and authors of relevant papers and, in a single step, generates a comprehensive outline for a survey paper.
The core idea behind Meow is to leverage the advanced capabilities of Large Language Models (LLMs) to understand and organize complex academic information. Existing automated survey tools often produce rigid outlines that lack a deep understanding of the topic or fine-grained stylistic nuances. Meow aims to overcome these limitations by formulating outline writing as a direct generation task, making the process more efficient and the output more sophisticated.
How Meow Works
The framework involves two main stages. First, a high-quality dataset was meticulously curated from millions of academic papers across platforms like arXiv, bioRxiv, and medRxiv. This dataset includes not only paper metadata but also human-written hierarchical outlines, which are essential for training the model. The data curation process involved rigorous filtering to ensure only well-organized survey articles with complete citation information were included. A unique aspect of this stage is the use of Chain of Thought (CoT) distillation, where an LLM (DeepSeek-R1) was prompted to derive taxonomies from references, essentially teaching the model to reason step-by-step.
The second stage involves a two-stage training approach. It begins with supervised fine-tuning (SFT) on the CoT dataset, where the model learns to generate outlines based on the provided examples and reasoning chains. Following this “cold-start” initialization, the model undergoes reinforcement learning (RL) optimization using Group Relative Policy Optimization (GRPO). This stage refines the outline quality by incorporating specific reward functions. These rewards include a “Structural Similarity Reward,” which measures how closely the generated outline’s tree structure matches a human-written reference, and a “Format Compliance Reward,” which ensures the output adheres to the required schema.
Also Read:
- Enhancing Speech Summarization in Multi-modal AI Models Through Advanced Training
- Enhancing Language Models with Specialized Knowledge through Iterative Reinforcement Learning
Performance and Impact
Meow’s performance has been rigorously evaluated against several leading LLMs, including DeepSeek-R1, DeepSeek-V3, GPT-5 Nano, and Gemini 2.5 Flash-Lite, as well as the academic community’s SurveyX tool. The results show that Meow-8B-SFT-GRPO, the fully trained version of the model, consistently achieves superior structural fidelity and stylistic coherence. It scored highly on metrics like “Structure Locate,” “Content Exclusion,” and “Pragmatics Concise,” indicating its ability to produce outlines that are well-organized, avoid redundancy, and use clear, descriptive titles.
Notably, Meow’s approach significantly reduces the “Structural Distance” from human-written outlines, meaning its generated outlines are structurally very similar to those crafted by experts. This demonstrates that by focusing on structured reasoning and human-aligned reward modeling, Meow can generate outlines that better conform to academic writing conventions and research logic.
This innovative framework offers a scalable pathway toward fully automated survey generation, potentially saving researchers immense time and effort. By providing a robust foundation for survey papers, Meow helps ensure that the structural and logical framework of academic reviews is sound and comprehensive. You can learn more about this research by reading the full paper here: MEOW: END-TO-END OUTLINE WRITING FOR AUTOMATIC ACADEMIC SURVEY.
