A recent study introduces CGTime, a 4B-parameter computation-grounded time-series-language model that effectively aligns multivariate time series with natural language. The authors argue that traditional methods fall short due to a self-supervision trap, where the labeling capability of language models (LLMs) is inherently limited by their perceptual skills. Commonly used data sets focus on single variables, failing to capture essential patterns like cross-channel correlation and lead-lag structure, which are critical for accurate representation in multiple time series.
To address these limitations, the authors propose decoupling perception from description, allowing deterministic code to compute relevant statistical facts from real, open-source multivariate time series data. The LLM's role is then narrowed to verbalizing these precomputed facts, effectively delegating the perception task to the computation layer. This approach enables CGTime to outperform larger, more general-purpose models on tasks specifically related to multivariate understanding. In their tests, CGTime achieved a multivariate fact score of 0.283 on a held-out benchmark, surpassing scores of 0.173 for the GPT-4o-mini and 0.203 for the GPT-5.4-nano, with results validated through Holm-corrected paired significance tests. Furthermore, the model exhibits enhanced accuracy in stating verifiable numerical facts in generated captions and encompasses a wider variety of statistical properties.
The study advocates for this new methodology as a solution to the long-standing trilemma in multimodal learning, where methods could be reliable, realistic, or scalable, but not all three simultaneously.