<ul><li>Large language models (LLMs) and code generation models have shown human-level performance in tasks like mathematical reasoning and software development.</li><li>Hallucination in LLMs raises concerns regarding their applicability to systems with high safety standards, drawing attention in the AI community.</li><li>The hallucination of code generation models is a less explored issue, with the challenge lying in verifying the functionality of generated code due to its non-natural form.</li><li>To address this, a selective code generator is proposed, which automatically generates unit tests using dynamic code analysis tools to ensure functional correctness of generated code, controlling the rate of code hallucination for trustworthiness.</li></ul>

Selective Code Generation for Functional Guarantees

Discover more