Researchers have developed a new method to improve the consistency of multilingual language models (LLMs). They found that different languages can produce different rankings for the same model. To address this issue, they created a 'Consensus-Based Calibration' (CBC) estimator that can recover the interaction between language and model without human labels. The CBC method improves the ranking consistency of LLMs across tasks and languages, with significant improvements in panel agreement with human preferences. --- Why it matters: This matters to AI researchers because it provides a solution to a long-standing problem in multilingual LLM evaluation. By improving ranking consistency, this method can help developers create more reliable and accurate language models that can be used in various applications. Source: https://arxiv.org/abs/2608.22432