December 8, 2025CybersecurityOpen Access

Beyond prompting: the role of phrasing tasks in vulnerability prediction for Java

View Full Paper

Ask AI

Bookmark

Authors

Discussion

Empirical investigation shows varying F1 scores for LLMs predicting vulnerability in Java, suggesting better prompt strategies are essential.

Vulnerability is predicted using machine learning and deep learning approaches, particularly through large language models.
F1 scores indicated GPT-3.5 varied from 0.53 to 0.69 while Llama-2 showed a broader score range from 0.35 to 0.72.
Assessment utilized zero-shot techniques and rephrased tasks, revealing inconsistencies in predictions.
Results highlight the importance of tailored response analysis and prompt engineering for improved accuracy in predictions.

Hinrichs et al. (2025) studied this question.

View Full Paper

Ask AI

Bookmark

Empirical investigation shows varying F1 scores for LLMs predicting vulnerability in Java, suggesting better prompt strategies are essential.