Reproduce “Hallucination by Code Generation LLMs: Taxonomy, Benchmarks, Mitigation, and Challenges” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yunseo Lee, John Youngeun Song, Dongsun Kim, Jindae Kim, Mijung Kim, Jaechang Nam
This survey provides a structured overview of hallucinations in code-generating Large Language Models (LLMs), proposing a detailed, objective taxonomy for various error types. Researchers from UNIST, Handong Global University, Korea University, and Seoul National University of Science and Technology collaborated to review existing benchmarks, identify root causes, and analyze mitigation strategies, addressing a critical gap in CodeLLM reliability research.
Key insights
- Existing CodeLLM benchmarks often lack language diversity and fail to reflect real-world software engineering workloads, limiting their ability to fully capture hallucination types.
- Hallucinations in code generation stem from a combination of training data issues, inherent limitations in trained model architecture, and ambiguous prompt design.
- A standardized, objective taxonomy for code hallucinations is essential for fostering coherent research, enabling clear communication, and facilitating consistent evaluation across studies.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.