AI TrackerFrontier desk · AEST
Back to news
normalaitech

StepBack: Step-Level Error-Localized Resampling for Efficient Test-Time Reasoning

arXiv cs.CL · 6 Nov 2026, 04:12 AEST

StepBack is a new method for efficient test-time reasoning that improves upon traditional self-consistency strategies. It reduces token usage by up to 41.26% while maintaining accuracy across various benchmarks.

Key facts

Read aloud

StepBack is a new method for efficient test-time reasoning that improves upon traditional self-consistency strategies. It reduces token usage by up to forty one point two six percent while maintaining accuracy across various benchmarks. Self-consistency SC is a widely used decoding strategy for improving chain-of-thought reasoning, but it requires sampling many complete reasoning trajectories with a preset budget. In this paper, we propose StepBack, a step-level error-localized resampling method for efficient test-time reasoning. Empirically, StepBack matches the accuracy of SC on average across seven mathematical reasoning and multi-hop QA benchmarks, while reducing token usage by up to forty one point two six percent. By nature, StepBack can be combined with adaptive inference methods such as Adaptive Consistency, further improving their inference-time efficiency.

Sources in this cluster