Researchers have developed a new type of large language model called Credal Large Language Models (CLLMs) that can better represent uncertainty in its predictions. Unlike standard models, CLLMs use an ensemble of adapters to create a 'credal set' that shows the range of possible outcomes rather than just one confident answer. This allows for two commitment scores: Credal Token Commitment and Semantic Commitment Consistency. The authors tested CLLMs on various tasks such as question-answering, hallucination detection, and selective prediction, and found it to be competitive with other models while providing better uncertainty representation. --- Why it matters: This matters because current language models often produce confident but incorrect answers, which can lead to problems in applications like decision-making or critical thinking. CLLMs aim to address this issue by providing a more nuanced representation of uncertainty, making them potentially useful for tasks that require careful consideration and evaluation of options. Source: https://arxiv.org/abs/2608.23244