MOCHA: Multi-Objective Chebyshev Annealing for Agent Skill Optimization

Researchers introduced MOCHA, a multi-objective method for optimizing LLM agent skills with Chebyshev scalarization and exponential annealing.

The approach balances task accuracy with deployment constraints instead of tuning for a single objective.

Across six tasks, MOCHA outperformed baseline prompt optimizers, including gains on FEVER and TheoremQA.

It also found more Pareto-optimal skill variants, suggesting better trade-offs across competing goals.