New AI Model Excels at Multivariate Time Series Understanding.

Training multimodal models to align time series data with natural language descriptions often falls into a "self-supervision trap." This occurs because the quality of labels, typically generated by an LLM reading a series and writing a description, is limited by the very perceptual skill the model is supposed to learn. Furthermore, most datasets focus on single variables, failing to capture crucial multivariate patterns like cross-channel correlations or co-occurring anomalies, where LLM labeling limitations are most pronounced. To overcome this, researchers propose a novel approach that decouples perception from description. Instead of relying on an LLM for perception, deterministic code is used to compute a comprehensive set of statistics from real, open-source multivariate time series. The LLM's role is then limited to verbalizing these precomputed facts. This strategy assigns perception, which LLMs perform poorly, to computation, while leveraging the LLM's strength in expression. This methodology led to the development of CGTime, a 4-billion-parameter computation-grounded time-series-language model. CGTime significantly outperforms much larger general-purpose models on multivariate understanding tasks. It achieved the best multivariate fact score on a held-out benchmark, demonstrating a substantial improvement over models like GPT-4o-mini and GPT-5.4-nano. The model also generates captions with more accurate verifiable numerical facts and covers a broader range of statistical properties, resolving the trilemma of reliability, realism, and scalability in time series-language alignment.