New method uses geometric consensus to select better LLM responses without training or voting

Radial Consensus Score (RCS) addresses the limitation of majority voting by identifying the most reliable LLM response among multiple candidates

RCS computes a weighted Fréchet mean of answer embeddings to find a semantic center, then ranks responses by their distance to this center

The training-free approach captures relationships between candidate answers and avoids underweighting high-quality but less frequent responses

RCS provides a flexible framework supporting multiple weighting schemes for improved best-of-N selection in language models