Researchers have developed a new framework called Perseus for segmenting time series data. Unlike previous methods that use dense state labels for training, Perseus uses sparse expert prompts to provide inference-time corrections. This allows the model to retain user interaction history and improve accuracy in multi-granularity settings. The authors claim that their method achieves up to 85% accuracy improvement compared to stateless prompting baselines. However, these claims are based on experiments conducted on six datasets, which may not be representative of all real-world applications. --- Why it matters: This matters because many real-world systems generate multivariate time series data with hierarchical states, and accurate segmentation is crucial for tasks such as anomaly detection and event prediction. Perseus's ability to retain user interaction history and improve accuracy in multi-granularity settings makes it a promising solution for these applications. Source: https://arxiv.org/abs/2510.09930