Selective Code Generation for Functional Guarantees (2505.13553v1)

Abstract: LLMs show human-level performance and their specialized descendants, code generation models, play core roles in solving complex tasks, including mathematical reasoning and software development. On the downside, the hallucination of LLMs mainly hinders their applicability to systems requiring higher safety standards, thus drawing the attention of the AI community. However, the hallucination of code generation models is rarely considered. One critical bottleneck in considering code hallucination is the intricate property of code to identify whether generated code has the intended functionality due to its un-natural form, different to natural languages. Handful of unit tests have been considered to address this issue, but scaling-up its size is extremely expensive. We address this core bottleneck by automatically generating unit tests using dynamic code analysis tools, which leverages the \emph{executable nature} of code. Given generated unit tests from true code for measuring functional correctness of generated code, we propose to learn a \emph{selective code generator}, which abstains from answering for unsure generation, to control the rate of code hallucination among non-abstaining answers in terms of a false discovery rate. This learning algorithm provides a controllability guarantee, providing trustworthiness of code generation. Finally, we propose to use generated unit tests in evaluation as well as in learning for precise code evaluation, calling this evaluation paradigm \emph{FuzzEval}. We demonstrate the efficacy of our selective code generator over open and closed code generators, showing clear benefit of leveraging generated unit tests along with the controllability of code hallucination and reasonable selection efficiency via our selective code generator.