ETH's TIDES moves input dependence off the step size so selective SSMs handle irregular time series

ethz · hf · 2026-10-08

ETH Zürich introduced TIDES, a selective state space model variant that natively handles irregularly sampled time series.

Problem: In Mamba-style selective SSMs, the discretization step Δ̃ is a learned function of the input, so it no longer equals the physical time gap Δ, hurting irregular-series handling. Continuous-time SSMs like S5 keep Δ̃≡Δ but remain linear time invariant, limiting per-token expressivity.
Method: TIDES moves input dependence off the step size and onto the diagonal state matrix, preserving both Δ̃≡Δ and selective SSM expressivity.
Diagnostic: A novel Fading Flash benchmark jointly tests input dependence and extrapolation to out-of-distribution Δ, isolating distinct failure modes of current architectures that TIDES avoids by construction.
Results: New best average rank on UEA time series classification and the Physiome ODE regression benchmark; matches or exceeds the reference baseline on 6 of 8 natively irregular datasets spanning astronomy, agriculture, neuromorphic sensing, and climate events.

Code: https://github.com/TaylanSoydan/TIDES

More from Research