COM method boosts time series LLMs by preserving token continuity
New geometric constraints fix a key flaw in token-based time series LLMs.
Token-based large language models for time series (TS-LLMs) have gained traction for analysis and reasoning, but prior approaches largely ignored two fundamental properties of time series data: continuity (smooth transitions between consecutive points) and ordinality (the inherent order of timestamps). A new paper from researchers at arXiv introduces COM, a strategy that explicitly constrains token embeddings to respect these properties during both initialization and training.
By integrating geometric constraints into the embedding space, COM ensures that tokens representing nearby time points remain close together and that the order of tokens is encoded in a measurable way. Empirical tests on standard time series analysis benchmarks show that COM consistently improves the accuracy of token-based TS-LLMs, outperforming previous methods without architectural changes. The code is publicly available, making adoption straightforward for practitioners.
- COM adds geometric constraints during token embedding initialization and training to preserve continuity and ordinality.
- Achieves consistent improvements across multiple time series analysis benchmarks.
- Works with existing token-based TS-LLMs and requires no architecture changes.
Why It Matters
Better time series LLMs enable more accurate forecasting and anomaly detection in finance, IoT, and healthcare.