COMPAS: Difficulty-Aware Joint Search for Optimizing Code Generation

A recent study published on arXiv (2608.04336) presents COMPAS, a novel technique aimed at enhancing code generation within large language models (LLMs). The research investigates the interplay between model selection, prompts, and decoding parameters, revealing that the interaction between prompts and decoding settings varies by model, with optimal configurations influenced by task complexity. COMPAS identifies group-specific quality-cost fronts by employing cost-effective model selection and a combined prompt-decoding search, subsequently directing each test task to its corresponding front in real-time. This method overcomes the shortcomings of current techniques that either optimize only certain aspects of the configuration or apply a uniform setup across all tasks.