ProtoFlow Uses VQ Codebooks for Non-Autoregressive Time Series Forecasting
October 2, 2026
ProtoFlow replaces autoregressive token generation in multivariate time series forecasting with prototype-guided flow matching. By initializing transport from learned vector quantization codebooks instead of generic Gaussian priors, the method avoids exposure bias and reduces inference sampling steps compared to diffusion models.
HOW THIS AFFECTS YOU
●
builderThis offers a more efficient inference path for high-dimensional time series forecasting tasks.
●
researcherYou can explore non-autoregressive alternatives to AR-based VQ forecasting.