CB-Orchestrator preprint reports lower token costs for AI workflow training

Researchers report that CB-Orchestrator, a framework for choosing how AI agents carry out tasks, improves performance while reducing training token use. On the MATH dataset, their preprint reports better performance than strong recent baselines with training-phase API token overhead reduced by no less than 40.55%. The framework first generates candidate workflows through evolutionary search, then learns which workflow to select for each query. Its selection method balances trying alternatives with reusing successful workflows to reduce the feedback needed for training. Including workflow pool generation, total end-to-end training token overhead fell by no less than 29.73% on MATH. The reported reductions concern training, rather than inference costs.