Skip to content

Universal Time-Series Tokenization Bottleneck

Home / Open Questions / Universal Time-Series Tokenization Bottleneck

Background: Time series generation commonly employs vector quantization (VQ) with autoregressive models, which are susceptible to exposure bias and sequential error accumulation. Current frameworks typically rely on dataset-specific VQ tokenizers, limiting the potential for cross-domain generalization and unified latent modeling.

Question / Future Work: The development of a universal, cross-domain time-series tokenizer with a shared, robust codebook is required to enable non-autoregressive generative models to operate in a unified latent space across diverse, heterogeneous time-series sources.

Why It Matters: A universal tokenizer is critical for transitioning from domain-specific generation models to general-purpose time-series foundation models that can leverage cross-domain data scaling.

Metadata & Links