← Research overview

Latent Laplace Diffusion for Irregular Multivariate Time Series

Zinuo You, Jin Zheng, John Cartlidge
ICML 2026 · Spotlight

Generate a whole forecast window at arbitrary timestamps, without stepping through physical time.

First author and lead implementation author · 2023–2025

Irregular time is part of the problem

Sensor and event streams rarely arrive on a perfect grid. Regridding can distort timing, while sequential continuous-time solvers add work as the requested horizon grows.

One trajectory, many query times

LLapDiff learns conditional diffusion in a compact latent trajectory space. Stable complex-conjugate modes and gap-aware context support direct evaluation at requested timestamps. The same continuous-time formulation supports forecasting and missing-value imputation.

Reading the demonstration

A damped oscillatory signal makes the query mechanism visible. Move the query timestamp and remove observations: the known synthetic trajectory can still be evaluated directly. The signal is supplied analytically; this is not a prediction from LLapDiff and observation removal does not fit a new model.

Interactive toy experiment

One trajectory, arbitrary query times

Remove irregular observations, move the query timestamp, or change damping. The known signal is evaluated directly at the query.

One trajectory, arbitrary query times-1.00.00.16.01.312.0TimeObservationsQuery region

Gray: known signal · Blue dots: observations · Blue line and hollow dot: query

14 of 24 observations retained · y(8.0) = -0.2693. Missingness does not refit this known signal.

How this toy works

y(t) = exp(−dt)[sin(1.5t) + 0.35 cos(3.4t)]. Twenty-four irregular observation times span 0 to 6; queries extend to 12.

The signal is supplied analytically. Removing observations does not fit a model or produce a posterior. This illustrates stable modes and timestamp queries, not forecasting accuracy.

Synthetic illustration · Not a trained-model output or a published benchmark.