Latent Self-Consistency for Reliable Majority-Set Selection in Short- and Long-Answer Reasoning
This addresses reliability issues in LLM reasoning for both short- and long-form questions, offering a practical solution with broad applicability.
The paper tackled the problem of inconsistent outputs from Large Language Models in reasoning tasks by introducing Latent Self-Consistency, which selects semantically consistent responses using learnable token embeddings, achieving superior performance over existing methods across 11 benchmarks with less than 1% inference overhead.
Probabilistic decoding in Large Language Models (LLMs) often yields inconsistent outputs, particularly on complex or long-form questions. Self-Consistency (SC) mitigates this for short-form QA by majority voting over exact strings, whereas Universal Self-Consistency (USC) and Weighted Unigram Consistency Score (WUCS) extend to long-form responses but lose accuracy on short-form benchmarks. We introduce Latent Self-Consistency (LSC), which selects the most semantically consistent response using learnable token embeddings. A lightweight forward generation of summary tokens increases inference time by less than 1% and requires no changes to the model architecture. Across 6 short-form and 5 long-form reasoning benchmarks (e.g., MATH, MMLU, TruthfulQA), LSC surpasses SC, USC and WUCS on all short-form and long-form ones on average, while maintaining negligible computational overhead. These results position LSC as a practical consistency-selection method that works reliably across answer formats. Additionally, LSC provides well-calibrated confidence estimates, maintaining low Expected Calibration Error across both answer formats.