instead of re-sampling whole reasoning trajectories, StepBack localizes the likely first error step from a small draft set and resamples that step and its suffix; it matched self-consistency accuracy across seven math and multi-hop QA benchmarks while cutting token use by up to 41.26%.