MOCHA: Multi-Objective Chebyshev Annealing for Agent Skill Optimization

In three linesMOCHA is a multi-objective optimization algorithm for refining LLM agent skills. It uses Chebyshev scalarization and exponential annealing to explore the complete Pareto front, including non-convex regions. On 6 tasks, MOCHA improves performance by 7.5% on average (up to 14.9% on FEVER) while discovering twice as many Pareto-optimal skill variants as baselines.

Summary generated by Claude — human-verified