CCPS

A training-free inference-time sampling method for LLMs that combines a Chopthin resampler with a semantic majority voting mechanism. By preserving diversity among reasoning paths, the method improves large language model performance on reasoning tasks without requiring model fine-tuning.