Time-series language models (TSLMs) let models understand and reason about temporal data using language, but a core question remains: what information should their internal representations preserve, and what variation should be discarded? A new arXiv preprint, Twice Upon a Time: Complementary Representations for Time-Series Reasoning, directly addresses this issue.
The abstract, as available, poses this question but is truncated before detailing the proposed approach. The title hints that the answer lies in using complementary representations—multiple views that together capture different aspects of the data—rather than a single monolithic encoding.
If the paper delivers on this idea, it could help make time-series models more interpretable and robust, especially for tasks that require explaining trends, anomalies, or forecasts in plain language. The work is currently a preprint and has not yet been peer-reviewed.