Researchers have found that injecting new subjects into a language model's stream of text at regular intervals can increase judged surprise and connection by 1.2 to 1.4 points and 0.8 points, respectively, compared to simply repeating the same text. This effect is attributed to a single operation: introducing a new subject every few hundred tokens. The study also found that other interventions, such as reset context or pre-registered replication, had little impact on judged surprise and connection. However, the researchers note that these findings do not necessarily imply a mechanism of creativity. --- Why it matters: This research matters to engineers and researchers in AI because it provides insights into how language models generate novel text and how they can be improved. Understanding the effects of periodic subject changes on judged surprise and connection can inform the development of more effective language generation techniques. Source: https://arxiv.org/abs/2608.19893