frontier models
Reasoning Restarts at the First Suspected Error
StepBack localizes a likely first mistake, then resamples only that step and the suffix rather than whole reasoning traces.
Summary
StepBack localizes a likely first mistake, then resamples only that step and the suffix rather than whole reasoning traces.
The method begins with a small set of draft trajectories, uses lightweight signals to find likely first-error positions and votes across corrected suffixes. Across seven math-reasoning and multi-hop question-answering benchmarks, the authors report average accuracy matching self-consistency while using up to 41.26 percent fewer tokens. It can also be combined with adaptive-consistency methods. The savings depend on the tested tasks and localization signals.
Why it matters
StepBack localizes a likely first mistake, then resamples only that step and the suffix rather than whole reasoning traces.
Limits and context
No additional limitation was separately recorded.
Key claims
StepBack localizes a likely first mistake, then resamples only that step and the suffix rather than whole reasoning traces.
Evidence: source-2026-11-06-011
Sources
- arXiv preprint 2611.08403, original v1arXiv · primary research
Corrections
No corrections have been recorded for this story.