Task Abstention for Large Language Models in Code Generation

In three linesMethod enabling LLMs to abstain from code generation tasks prone to hallucination. Uses calibrated abstention rule grounded in multiple hypothesis testing, assesses consistency through code execution outcomes. Provides distribution-free theoretical guarantee. Evaluated on open-source code LLMs.

Summary generated by Claude — human-verified