A new arXiv paper (2610.24609) investigates two decision points in LLM code generation pipelines. Admission rules determine which generated programs are kept, while selection policies decide which single candidate is returned for each task. The study audits how these decisions interact with execution contracts.
The authors use two separately frozen, disjoint HumanEvalPlus cohorts, likely to ensure isolation and avoid contamination between evaluation sets. The abstract does not report specific results, but the described methodology points to a systematic audit of these decision layers.
Understanding the interplay between admission and selection is important for building reliable code-generation systems. By tying the audit to execution contracts, the work adds a practical dimension to evaluating LLM output quality.