The field of software engineering is increasingly leveraging Large Language Models (LLMs) to address complex challenges, with a notable shift towards enhancing software security, testing, and repository-level tasks. Recent developments highlight the application of LLMs in vulnerability detection across various programming languages, demonstrating varied effectiveness and uncovering the potential for improved detection capabilities beyond traditional methods. Hybrid fuzzing techniques have been advanced by integrating LLMs to bypass the limitations of symbolic execution, significantly improving coverage and efficiency in exposing program behaviors. In the realm of unit testing, LLMs have been systematically evaluated, revealing their superior performance over existing methods and the promising potential of fine-tuning and prompt engineering approaches. The integration of visual data into issue resolving processes marks a novel direction, enhancing the capabilities of LLMs in handling complex repository-level tasks. Furthermore, the creation of real-world benchmarks for repository-level code translation underscores the challenges and current limitations of LLMs in accurately translating entire codebases, pointing towards the need for further advancements in this area.

Noteworthy Papers

• Vulnerability Detection in Popular Programming Languages with Language Models: Demonstrates the varied effectiveness of LLMs in detecting vulnerabilities across different programming languages, with JavaScript showing superior performance.
• Large Language Model assisted Hybrid Fuzzing: Introduces an LLM-based hybrid fuzzer that significantly outperforms existing tools in coverage and efficiency.
• A Large-scale Empirical Study on Fine-tuning Large Language Models for Unit Testing: Reveals that LLMs outperform state-of-the-art methods in unit testing tasks, highlighting the effectiveness of fine-tuning.
• CodeV: Issue Resolving with Visual Data: Proposes a novel approach leveraging visual data to enhance LLMs' capabilities in resolving GitHub issues.
• RepoTransBench: A Real-World Benchmark for Repository-Level Code Translation: Highlights the current limitations of LLMs in repository-level code translation, despite improvements with iterative debugging.