All publications

Language models for code optimization: survey, challenges and future directions

2025·arXiv·Read the paper
LM code optimization50+ studiespromptingfine-tuningsearch hybridsagentsgeneration + measurementwhere results survivethe surveyed field, and its load-bearing corner

Figure 1: A field in taxonomy. LM-based code optimization approaches branch by strategy; the surveyed evidence concentrates where generation meets measurement.

A systematic review of more than fifty studies applying language models to code optimization, organized into a taxonomy spanning prompting strategies, fine-tuned optimizers, search-based hybrids, and agentic pipelines.

The survey identifies the field's recurring gap: generation is abundant, validation is scarce. Approaches that couple LM generation with real measurement and correctness checking consistently produce the results that survive scrutiny — the direction the field, and Artemis, are converging on.

Key results

  • 50+ studies systematically reviewed and classified
  • Identifies measurement-validated generation as the load-bearing pattern
  • Roadmap of open challenges for LM-based optimization