ExeCRE: New Framework Improves Code Generation Reliability via Execution Consistency
A new framework named ExeCRE (Execution-Consistency guided code Reliability Estimation) has been developed by researchers to improve the dependability of self-correcting code generated by large language models (LLMs). This framework tackles a significant issue in current self-correction methods: the reliance on supervision signals that may be unreliable, which can result in misleading feedback, unnecessary modifications, and incorrect outcomes. Instead of depending on tests or LLM feedback, ExeCRE assesses code reliability through statistical analysis of execution output consistency across numerous randomly generated inputs. It gathers execution results, translates them into consistency signals, and utilizes these signals to direct the self-correction process. The research paper can be found on arXiv with the identifier 2608.04439, marking an important advancement in AI and code generation.
Key facts
- ExeCRE is a framework for estimating code reliability in self-correcting code generation.
- It uses execution-consistency patterns over randomly generated inputs.
- It addresses the problem of unreliable supervision signals in self-correction pipelines.
- The method collects execution outputs and projects them into consistency signals.
- The paper is available on arXiv with ID 2608.04439.
- The research targets large language models (LLMs) in code generation.
- The framework aims to reduce misleading feedback and incorrect final answers.
- It was announced as a cross submission on arXiv.
Entities
Institutions
- arXiv