Adapting LLM to Generate Vulnerability Descriptions and Repair Suggestions

Software vulnerabilities pose serious threats to the security of modern software systems. Existing detection tools can flag or pinpoint vulnerability but offer little guidance on how to fix them. In this paper, we leverage the generative and reasoning strengths of LLMs to bridge this gap, combining prompt tuning, retrieval-augmented generation, and chain-of-thought techniques to produce both clear vulnerability descriptions and concrete repair suggestions. We evaluated five mainstream LLMs on the BigVul dataset, and the combination of RAG and CoT methods achieves leading performance across multiple metrics in both the vulnerability description generation and repair suggestion generation tasks.