Researchers have found that some AI models can produce coherent text even when given no useful input. They studied this phenomenon in speech recognition and machine translation systems by analyzing the models' reserved 'null tokens', which are used to indicate the end of a message. The study found that these null tokens often contain a signal that can help the model decide whether or not to generate more text, but current decoding methods don't reliably use this information. By adjusting the null token score, researchers were able to reduce the amount of 'hallucinated' text generated by the models, which is text that doesn't correspond to any actual input. --- Why it matters: This research matters because it highlights a limitation in current AI systems and suggests new ways to evaluate their performance. By developing methods to effectively use the null token signal, researchers can improve the accuracy of speech recognition and machine translation systems, reducing the amount of 'hallucinated' text they produce. Source: https://arxiv.org/abs/2608.15940