The paper opens with a metaphor: an ensemble of language models is a jury, not an electorate. The outputs are testimony from witnesses whose reliability is unknown, so the key question is not which answer gets the most votes but which witness is credible.
The abstract notes that most ensemble methods assign each model a role in producing the final answer. The paper argues instead that the real task is estimating credibility, and it proposes a label-free method for doing so—meaning the reliability of each model is assessed without relying on ground-truth labels.
This reframing could matter for practical applications where ensembles are used to improve accuracy but labeled validation data is scarce. By focusing on credibility rather than majority opinion, the approach aims to make ensemble decisions more robust.