MOCHA: Multi-Objective Chebyshev Annealing for LLM Agent Skill Optimization

A novel technique known as MOCHA (Multi-Objective Chebyshev Annealing) has been developed to enhance the skills of LLM agents while adhering to platform limitations. LLM agents utilize skills, which are structured natural-language specifications that consist of multi-field artifacts constrained by factors like truncated description fields, compact instruction bodies, and restricted context windows. These limitations create a multi-objective optimization challenge, necessitating the simultaneous improvement of task performance and compliance with platform restrictions. Current prompt optimizers either overlook these trade-offs or simplify them into a weighted sum, failing to capture Pareto-optimal solutions in non-convex objective areas. MOCHA employs Chebyshev scalarization for a comprehensive exploration of the Pareto front, including non-convex regions, paired with exponential annealing. This methodology is outlined in a paper available on arXiv (2605.19330v1).