LW2S Reduces Token Costs in Multi-Agent Workflows via Component Omission

Learning What to Skip (LW2S) uses counterfactual credit assignment to train safety models that decide when to omit workflow steps like execution or verification. This approach reduces token costs while matching or improving performance in mathematical reasoning, QA, and code generation.