ParEVO: Synthesizing Code for Irregular Data: High-Performance Parallelism through Agentic Evolution

2026-03-04 · Source: cs.NE updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering · Depth: Expert, quick

Summary

ParEVO is a novel framework designed to synthesize high-performance parallel algorithms for irregular data structures, addressing the challenges of concurrent programming where static scheduling and predictable data dependencies are absent. Traditional Large Language Models often fail to generate correct and scalable code for these tasks, leading to issues like race conditions and deadlocks. ParEVO introduces three key contributions: the Parlay-Instruct Corpus, a dataset of 13,820 tasks filtered for empirically performant algorithms; specialized DeepSeek, Qwen, and Gemini models fine-tuned for ParlayLib semantics; and an Evolutionary Coding Agent (ECA) that iteratively repairs code using feedback from compilers, race detectors, and profilers. On the ParEval benchmark, ParEVO achieved an average 106x speedup, with a maximum of 1103x, and a 13.6x speedup on complex irregular graph problems, surpassing commercial models and matching expert human baselines with up to a 4.1x speedup on specific kernels.

Key takeaway

For AI Scientists and Machine Learning Engineers developing high-performance computing solutions for irregular data, ParEVO demonstrates that combining specialized LLMs with an evolutionary coding agent can overcome the limitations of traditional code generation. You should consider integrating similar iterative feedback loops and domain-specific fine-tuning into your code synthesis workflows to achieve significant performance gains and reduce concurrency errors in parallel applications.

Key insights

ParEVO synthesizes high-performance parallel code for irregular data using an evolutionary agent and specialized LLMs.

Principles

• Empirical performance filtering is crucial for parallel code generation.
• Iterative repair with dynamic feedback improves code correctness.
• Aligning LLM generation with library semantics enhances reliability.

Method

ParEVO uses a "Critic-Refine" pipeline to create a corpus, fine-tunes LLMs (DeepSeek, Qwen, Gemini) to ParlayLib, and employs an Evolutionary Coding Agent (ECA) for iterative code repair via compiler, race detector, and profiler feedback.

In practice

• Utilize Work-Span primitives for parallel algorithm design.
• Integrate dynamic analysis tools for code correctness.
• Fine-tune LLMs on domain-specific code semantics.

Topics

• Parallel Computing
• Irregular Data Structures
• Code Synthesis
• Evolutionary Coding Agent
• Large Language Models

Code references

• WildAlg/ParEVO

Best for: Machine Learning Engineer, AI Scientist, Research Scientist, AI Researcher, AI Engineer, Software Engineer

Related on AIssential

• Code2Math: Can Your Code Agent Effectively Evolve Math Problems Through Exploration? — Code agents can autonomously evolve math problems, creating more complex and solvable variations for LLM training.
• BenchEvolver: Frontier Task Synthesis via Solution-Centric Evolution — BenchEvolver automatically hardens coding benchmarks by evolving solutions, restoring model discrimination and providing training data.
• Building a C compiler with a team of parallel Claudes — Parallel LLM agent teams can autonomously develop complex software projects, expanding the scope of AI-driven development.
• HyEvo for Topological Reasoning Graph Optimization — A self-evolving system autonomously designs optimal hybrid LLM-code agent workflows, eliminating manual architecture design.
• MLEvolve: A Self-Evolving Framework for Automated Machine Learning Algorithm Discovery — MLEvolve enables LLM agents to self-evolve for long-horizon ML algorithm discovery via graph search, memory, and adaptive code generation.
• 4 Layer "AI Harness" For LLMs (+54%). Really? — The "AI Harness" improves LLM agent performance by addressing interface mismatches and syntax issues, not by modifying the frozen model.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.NE updates on arXiv.org.