Research map / Time series / Causal and generative time series

TIDES: Implicit Time-Awareness in Selective State Space Models

NeurIPS 2026 · Time series › Causal and generative time series · approach: Dynamical system reconstruction · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes TIDES, a selective SSM that places input-dependence on the diagonal state matrix's real part (plus B and C) while keeping the discretization step equal to the physical time gap, enabling native irregular time handling with extrapolation.

Problem

Selective SSMs like Mamba make the discretization step a learned input-dependent gate, so physical irregular timestamps no longer enter the model meaningfully and fail to extrapolate; continuous-time SSMs like S5 handle irregular time but are LTI and lack per-token expressivity.

Approach

Move input-dependence off the step size and onto the continuous-time state matrix (Re(lambda), B, C), keeping the step equal to the physical sampling interval so time is handled implicitly by discretization while retaining selectivity.

Key results

• On UEA, TIDES achieves average accuracy 66.1 and average rank 3.00, ahead of Mamba-3 (64.7, rank 4.17).
• On Physiome-ODE, TIDES attains average rank 2.40 versus LinODEnet's 2.62, with a bootstrap CI on the rank difference including zero.
• On Fading Flash at Delta=0.1, TIDES relative error is 6.56 versus 141.20 for Mamba-3 and 154.08 for Mamba-2.

Surprising findings

• Input-dependent Im(Lambda) slightly hurts performance on the random-drop EigenWorms experiment.
• Mamba-family models match TIDES at moderate test irregularity but collapse at the most extreme test sampling ratio (r_test=0.9).

Limitations

• Evaluated only on irregular-time and regular-grid classification/forecasting, not language modeling or long-range regular benchmarks.
• Additional hyperparameters (rank, projector depths, reparameterization) increase tuning burden.
• Uses a generic parallel scan rather than a hardware-aware kernel, so wall-clock comparisons are not fully optimized.

Experimental setup

Datasets: UEA multivariate archive (SCP1, SCP2, MI, EW, ETC, HB) Physiome-ODE MACHO ASAS-SN TimeSen2Crop EuroCropsML SHD SSC DVS128 Gesture USHCN EigenWorms

Benchmarks: UEA Physiome-ODE EigenWorms random drop MACHO ASAS-SN TimeSen2Crop EuroCropsML SHD SSC DVS128 Gesture USHCN Fading Flash

Metrics: accuracy average rank MSE MAE relative error overall accuracy

Baselines: S5 Mamba Mamba-2 Mamba-3 Rough Transformer LRU NCDE NRDE LogNCDE LinODEnet GraFITi CRU

Builds on: S5 Mamba HiPPO Event-SSM

Closest papers by meaning

• Demystifying the Token Dynamics of Deep Selective State Space Models (ICLR 2025 · similarity 0.88)
• Theoretical Foundations of Deep Selective State-Space Models (NeurIPS 2024 · similarity 0.87)
• Structured Sparse Transition Matrices to Enable State Tracking in State-Space Models (NeurIPS 2025 · similarity 0.85)
• LogSig-SSM: Time-Series Modelling with Multi-Scale Log-Signature Compression for State-Space Models (NeurIPS 2026 · similarity 0.85)
• Structural Inference of Dynamical Systems with Conjoined State Space Models (NeurIPS 2024 · similarity 0.85)
• ABC: Any-Subset Autoregression via Non-Markovian Diffusion Bridges in Continuous Time and Space (NeurIPS 2026 · similarity 0.85)