Reproduce “StepBack: Step-Level Error-Localized Resampling for Efficient Test-Time Reasoning” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yang Ouyang, Jung-Eun Kim
StepBack uses signal-based rollback to reuse reasoning prefixes while generating continuations under a fixed answer budget. It scores draft steps to estimate where a path may become unreliable, retains earlier steps, and resamples the suffix. On GSM8K, Qwen-3B reaches 91.59% accuracy versus 91.21% for SC, with 38.37% fewer generated tokens. The comparison fixes 32 answers and does not establish lower latency.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.