Instruct Large Language Models to Generate Scientific Literature Survey Step by Step
TLDR
A prompt-based method using LLMs to generate scientific literature surveys step-by-step, achieving low cost and competitive performance.
Reasoning
The paper presents a practical, cost-effective approach for automated survey generation, with strong empirical results from a competition. However, it is narrowly focused on survey writing rather than broader scientific discovery or experimentation, limiting its scope.
Read-first score
Read-first score 42.2, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 46.
Field roles
Rank sensitivity
Stability: volatile; rank range: 71.
Keyword Scores
Deep Analysis
Innovations
- Step-by-step prompt design to sequentially generate title, abstract, hierarchical headings, and main content of a literature survey
- Cost-efficient generation by restricting input and output length in LLM queries
- High-level heading generation from a top-down perspective
Methodology
The authors design a series of prompts to guide large language models (specifically Qwen-long) to generate scientific literature surveys in a step-by-step manner: first the title, then abstract, then hierarchical headings, and finally the main content. Input and output lengths are restricted to minimize API costs. The approach was evaluated in the NLPCC 2024 Scientific Literature Survey Generation task.
Key Results
The method achieved third place with an overall score only 0.03% lower than second place, a soft heading recall of 95.84% (second best), and a generation cost of 0.1 RMB per survey.