Title: Enhancing Self-Consistency Through Ranking

Original: arXiv:2606.05054v1 Announce Type: new

Abstract: While self-consistency enhances large language models by generating multiple reasoning paths and choosing the most common response, standard majority voting frequently misses the correct answer even when it is included in the sample set. To overcome this drawback, we introduce Ranking-Improved Self-Consistency (RISC), a method that reframes the selection of answers within self-consistency as a ranking task. Rather than depending on a single measure of uncertainty or confidence, RISC employs a lightweight LambdaRank model to evaluate candidate answers using five specifically crafted features. These features are designed to capture answer frequency, semantic centrality, and consistency within reasoning traces. Our evaluation of RISC across three datasets, under various test-time constraints, demonstrates that it consistently offers a superior accuracy-efficiency balance compared to both standard self-consistency and robust baseline methods, with significant improvements observed on question answering benchmarks. Additional analysis reveals that the introduced features are not only effective on their own but also complementary, underscoring the benefit of learning to integrate multiple informative signals for answer selection at test time.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC