All publications

Tuning LLM-based Code Optimization via Meta-Prompting

2025·IEEE/ACM ASE 2025
code contextAST · profile · domainmeta-prompterwrites the promptsprompt → GPTprompt → Claudeprompt → localbenchone barone meta-prompt layer, a fleet of tailored prompts, a single benchmark

Figure 1: Meta-prompted optimization. A meta-prompter conditions on code context (a) and emits per-model prompts to a diverse generator fleet (b); candidates meet the same benchmark regardless of their origin.

Different LLMs want different prompts, and hand-tuning a prompt per model per codebase does not scale. MPCO introduces a meta-prompting layer: a model that writes the optimization prompts, conditioned on the target code's context, for each generator model in a multi-LLM fleet.

In Artemis's industrial setting this turned prompt engineering from a manual bottleneck into part of the pipeline itself — each generator receives a prompt shaped for its strengths, and the benchmark still arbitrates the results.

Key results

  • Context-aware prompts improve optimization performance across diverse LLMs
  • Scales across real-world industrial codebases
  • Published at IEEE/ACM ASE 2025