DMAIC-Bench: From Self-Reflection to Self-Restraint for Agents in Multi-Stage Analytical Tasks

DMAIC-Bench evaluates an agent's ability to use intermediate evidence to decide whether to continue, redirect, or stop a multi-stage workflow. Using 97 tasks across 15 industrial cases, the benchmark distinguishes warranted self-restraint from indiscriminate stopping via paired counterfactual variants. Tested models averaged 92% routing accuracy on matched GO decisions but only 19% on NO-GO decisions.