StepBack
A step-level error localization and resampling method for efficient test-time reasoning. It uses a small-scale draft trajectory to locate the first erroneous step and resamples only that step and its suffix; combined with majority voting, it maintains accuracy while substantially reducing token consumption.
Timeline (last 90 days)
The paper "StepBack: Step-Level Error-Localized Resampling for Efficient Test-Time Reasoning" proposes a step-level error-localized resampling method that matches the average accuracy of self-consistency (SC) on multiple benchmarks while reducing token usage by up to 41.26%.
StepBack was evaluated on seven mathematical reasoning and multi-hop QA benchmarks, and the results show that average accuracy is comparable to the self-consistency method while token usage is reduced by up to 41.26%.
StepBack first samples a small set of draft trajectories and uses lightweight signals (such as step-level log probabilities, answer consistency, semantic coherence scores, and rule/small-model verification) to locate the most likely "first error position" in each trajectory.
After locating the erroneous step, StepBack resamples only that step and its subsequent suffix rather than generating the entire trajectory from scratch, thereby preserving the correct reasoning in the prefix.
StepBack takes a majority vote over candidate answers, where the trajectories participating in the vote are a mixture of the original draft trajectories and the repaired trajectories produced by resampling the suffix after the erroneous step.
StepBack can in nature be combined with adaptive reasoning methods (such as Adaptive Consistency) to further improve reasoning efficiency.
The effectiveness of StepBack depends on the accuracy of the lightweight signals in locating the first error position; when localization is inaccurate, resampling may land on an incorrect prefix.
The paper's experiments cover only two types of tasks, mathematical reasoning and multi-hop QA, and StepBack's performance on other forms of reasoning still requires more verification.
All Facts (8)
The paper's experiments cover only two types of tasks, mathematical reasoning and multi-hop QA, and StepBack's performance on other forms of reasoning still requires more verification.
50%UnverifiedStepBack was evaluated on seven mathematical reasoning and multi-hop QA benchmarks, and the results show that average accuracy is comparable to the self-consistency method while token usage is reduced by up to 41.26%.
50%UnverifiedStepBack first samples a small set of draft trajectories and uses lightweight signals (such as step-level log probabilities, answer consistency, semantic coherence scores, and rule/small-model verification) to locate the most likely "first error position" in each trajectory.
50%UnverifiedAfter locating the erroneous step, StepBack resamples only that step and its subsequent suffix rather than generating the entire trajectory from scratch, thereby preserving the correct reasoning in the prefix.
50%UnverifiedStepBack takes a majority vote over candidate answers, where the trajectories participating in the vote are a mixture of the original draft trajectories and the repaired trajectories produced by resampling the suffix after the erroneous step.
50%UnverifiedThe paper "StepBack: Step-Level Error-Localized Resampling for Efficient Test-Time Reasoning" proposes a step-level error-localized resampling method that matches the average accuracy of self-consistency (SC) on multiple benchmarks while reducing token usage by up to 41.26%.
50%UnverifiedThe effectiveness of StepBack depends on the accuracy of the lightweight signals in locating the first error position; when localization is inaccurate, resampling may land on an incorrect prefix.
50%UnverifiedStepBack can in nature be combined with adaptive reasoning methods (such as Adaptive Consistency) to further improve reasoning efficiency.
50%