MO-CAPO: Multi-Objective Prompt Optimization for LLMs
A new algorithm, MO-CAPO, jointly optimizes LLM performance and inference cost, addressing limitations of existing single-objective and NSGA-II-based methods. It introduces a deployment-oriented cost objective and budget allocation for efficient optimization, evaluated across four tasks and three LLMs.
Key facts
- MO-CAPO is a multi-objective prompt optimization algorithm.
- It jointly optimizes performance and inference cost.
- It uses budget allocation for cost-efficient optimization.
- It proposes a deployment-oriented cost objective capturing full computational profile.
- Evaluated across four tasks and three LLMs.
- Compared to NSGA-II-based and single-objective optimizers.
- Published on arXiv with ID 2605.18869.
- Addresses sensitivity of LLMs to prompt design.
Entities
Institutions
- arXiv