Self-Evolving Agents via Likelihood-Guided Tool-Space Optimization

LOTS uses observed tool outputs to learn reusable, task-specific tool selections that improve agent performance while reducing tool context.