Introduction
In the modern academic landscape, the sheer volume of published research creates an "information explosion" that makes manual literature reviews increasingly difficult. While Large Language Models (LLMs) offer a potential solution for automated survey generation, creating a coherent, hierarchical "skeleton" or outline remains a significant challenge. Current automated tools often rely on rigid, one-size-fits-all templates or complex, multi-step agentic workflows that are slow, expensive, and prone to losing global logical coherence.
To address these limitations, we introduce Meow, the first end-to-end, metadata-driven framework specifically optimized for generating complex academic survey outlines. By treating outline generation as a single-step long-context task, Meow produces structures that rival human-written surveys in logical depth and stylistic precision.
The Meow Framework
Meow shifts away from iterative planning in favor of an end-to-end approach. The model takes a collection of paper metadata—including titles and abstracts—as input and generates a complete, nested outline in one pass. This process is refined through a two-stage training strategy:
- Stage 1: CoT Cold-Start (SFT): Using distillation from DeepSeek-R1, the model is taught "Chain-of-Thought" (CoT) reasoning. Rather than simply imitating final outlines, the model learns to categorize themes and synthesize taxonomies before drafting headings.
- Stage 2: Reinforcement Learning (GRPO): To optimize structural quality, we employ Group Relative Policy Optimization. The model is rewarded based on Structural Similarity (using Tree Edit Distance to compare against human benchmarks) and Format Compliance (ensuring strict adherence to academic Markdown hierarchies).
Data Curation and Benchmarking
To train Meow, we developed a high-quality pipeline utilizing 2.82 million papers from arXiv, bioRxiv, and medRxiv. From this pool, we extracted approximately 23,000 ground-truth survey articles. A critical component of our data preparation involved "Data Completion": since many bibliographies lack abstracts, we used vector similarity (all-MiniLM-L6-v2) to retrieve and attach missing abstracts, providing the model with a comprehensive long-context view of the research field.
To ensure rigorous evaluation, we established a new test set of 100 surveys published after 2025, preventing data leakage from the models' original pre-training sets.
Experimental Results
Our evaluations demonstrate that Meow-8B-SFT-GRPO significantly outperforms existing state-of-the-art methods and approaches human-level performance. On our primary benchmark, Meow achieved an LLM-as-a-Judge score of 36.79, remarkably close to the human-written baseline of 37.23.
Key findings from our experiments include:
- Structural Fidelity: Meow achieved the lowest "Structural Distance" (0.39), indicating its outlines are mathematically the most similar to those created by human experts.
- Efficiency of End-to-End Modeling: A single-step model with a long-context window proved superior to multi-step agentic workflows, suggesting that global visibility of the research field is essential for structural coherence.
- Superiority over Frontier Models: Meow-8B-SFT-GRPO outperformed significantly larger models, including DeepSeek-V3 and GPT-5 Nano, in specific metrics like "Structure Locate" and "Pragmatics Concise."
Conclusion and Future Work
Meow represents a significant step forward in automated academic writing by mastering the role of the "architect." While the current framework focuses exclusively on outline generation, it provides the necessary foundation for full-stack survey writing pipelines. Future research will explore integrating this structural expertise with content generation modules to produce complete, high-quality academic reviews automatically.
Reference: Ma, Z., et al. (2025). Meow: End-to-End Outline Writing for Automatic Academic Survey. arXiv:2509.19370v1 [cs.CL].
Summary updated