Redundancy-Aware Self-Consistency for Efficient LLM Reasoning

Abstract

Large language models increasingly solve multi-step reasoning tasks, where a single sampled chain of reasoning is frequently unreliable. Reaching a correct answer therefore requires aggregating many sampled reasoning paths into one trustworthy decision. The dominant family, self-consistency and its sampling-and-voting variants, suits such reasoning tasks, but its raw-count vote treats sampled paths as separate units of evidence, even when their reasoning is highly redundant. Within this regime, adaptive-budget methods curb the cost by deciding when to stop, yet their stopping signals are commonly based on the same redundancy-blind counts. We therefore propose Redundancy-Aware Consistency (RAC), which strengthens self-consistency by aggregating reasoning paths according to their effective support rather than their raw count. We establish the boundedness of this effective support and show that RAC recovers self-consistency when within-answer paths are mutually dissimilar. Experiments on four benchmarks with the primary backbone and a three-benchmark subset with a second backbone show that RAC achieves higher observed accuracy than the evaluated self-consistency and adaptive-budget baselines. On the primary backbone, it also uses substantially fewer samples.