ParEVO

Writing efficient parallel code for irregular data structures, such as graphs and sparse matrices, is notoriously difficult. Programmers must carefully balance load distribution, minimize synchronization overhead, and avoid subtle race conditions. While Large Language Models (LLMs) excel at generating sequential logic, they frequently fail in these high-performance parallel domains where correct abstractions, precise synchronization, and optimized memory access patterns are paramount. Current synthetic benchmarks primarily evaluate correctness on naive loops, ignoring the critical metric of runtime scaling (work-span) on complex data.

We bridge this gap with ParEVO, a framework designed to synthesize high-performance parallel algorithms for irregular data. Our contributions include: (1) The Parlay-Instruct Corpus, a curated dataset of 13,820 tasks synthesized via a "Critic-Refine" pipeline that explicitly filters for empirically performant algorithms that effectively utilize Work-Span DeepSeek (C++), Qwen (C++ & Rust), and Gemini (C++) models fine-tuned to align probabilistic generation with the rigorous semantics of the ParlayLib parallel data structures and algorithms library (and safe parallel Rust); and (3) an Evolutionary Coding Agent (ECA) that significantly improves the "last mile" of correctness by iteratively repairing code using feedback from compilers, dynamic race detectors, and performance profilers.