Interpretable Neural Dynamics Modeling
- The paper introduces interpretable models that embed explicit dynamical laws to recover intrinsic latent state dynamics.
- It leverages structured state-space formulations—such as neural ODEs and symbolic abstractions—to test scientific hypotheses.
- The approach enables analysis of velocity fields, fixed points, and regime transitions to extract causal and control insights.
Searching arXiv for recent and foundational papers on interpretable neural dynamics and state-space modeling. Interpretable-by-Design Neural System Dynamics Modeling denotes a family of approaches in which interpretability is embedded in the model class itself rather than added through post-hoc explanation. Across this literature, the common design principle is to represent temporal phenomena with explicit dynamical systems—typically low-dimensional latent state-space models, structured differential equations, decomposed dynamical operators, or symbolic transition abstractions—so that the learned model can be analyzed in the language of dynamics: velocity fields, fixed points, attractors, bifurcations, regimes, causal structure, and control objectives. In neuroscience, this framing is particularly prominent because neural population activity is often treated as a partial, noisy observation of an underlying dynamical process whose structure is scientifically meaningful (Zhao et al., 2017). More recent work extends the same principle to multiscale systems, cybernetic closed-loop modeling, hybrid physical-neural models, and symbolic abstraction, but the core commitment remains the same: the scientific hypothesis is expressed directly as a dynamical law rather than hidden inside an opaque predictor (Park et al., 26 Apr 2026).
1. Conceptual basis and scope
The central distinction of interpretable-by-design approaches is that they treat the model’s internal structure as the object of scientific reasoning. In nonlinear state-space formulations, observed signals are generated from latent states evolving under explicit dynamics, so interpretability arises from the latent law , not from a retrospective attribution on black-box features. In one canonical formulation for neural data, the latent state $\vx_t \in \mathbb{R}^d$ evolves according to
$\dot{\vx} = F_\theta(\vx(t), \vu(t)), \qquad \vy(t) \sim P(\vy(t)\mid G_\theta(\vx(t), \vu(t))),$
or, in discrete time,
$\vx_{t+1} = f_\theta(\vx_t, \vu_t) + \epsilon_t.$
Because the latent dynamics are a genuine dynamical system, they can be visualized and analyzed in terms of velocity fields, fixed points, limit cycles, attractors, and bifurcations (Zhao et al., 2017).
This same interpretability principle appears in several neighboring formulations. In integrative neurocybernetic modeling, nonlinear state-space models are treated as the core generative language, with the latent dynamics explicitly identified as the object that captures “neural computation and dynamics” (Park et al., 26 Apr 2026). In decomposed linear dynamical systems, complex non-stationary and nonlinear dynamics are represented as a sparse combination of simpler, more interpretable components, so that both the dictionary elements and the time-varying coefficients are directly inspectable (Mudrik et al., 2022). In symbolic abstraction approaches, the learned neural model is converted into a transition system whose partitions and edges expose region-to-region behavior in a graph-like representation (Yang et al., 2024).
A recurring misconception is that interpretability in this area is synonymous with low predictive complexity or linearity. The literature instead treats interpretability as a structural property of the hypothesis class. Some models remain highly expressive—using Gaussian processes, Neural ODEs, variational inference, or hybrid neural-mechanistic components—while preserving interpretable primitives such as local linear regimes, explicit vector fields, latent control objectives, or named concept variables (Hu et al., 2024). This suggests that the relevant contrast is not simply “simple versus powerful,” but “mechanistically legible versus post-hoc explained.”
2. Explicit latent dynamics as the primary interpretability mechanism
A major branch of the field models neural or physical observations as outputs of explicit latent dynamics. In early nonlinear neural trajectory modeling, the emphasis was on learning the continuous-time velocity field directly rather than learning an arbitrary nonlinear map from state to next state. One formulation writes
$\vx_{t+1} = \vx_t + \vg(\vx_t) + \vB(\vx_t)\vu_t + \epsilon_t,$
with
$\vg(\vx_t)=\vW_g \bm{\phi}(\vx_t) - e^{-\tau^2}\vx_t, \qquad \vecOp(\vB(\vx_t)) = \vW_B \bm{\phi}(\vx_t),$
where $\bm{\phi}(\vx)$ is a basis-function expansion. The explicit velocity-field parameterization makes phase-portrait analysis natural, and the built-in global contraction term discourages runaway extrapolation outside the support of observed trajectories (Zhao et al., 2016).
In a related development, variational joint filtering formulates an online nonlinear state-space model in which the latent dynamics are parameterized as
$\vx_{t+1} = \vx_t + f(\vx_t) + \vB_t \vu_t + \epsilon_{t+1}, \qquad f(\vx_t)=\vW\phi(\vx_t),$
$\phi_i(\vx)=\exp\!\left(-\gamma_i \|\vx-\vc_i\|_2^2\right).$
This basis choice is explicitly motivated by interpretability: it directly learns a smooth vector field in latent space, and the regularized RBF dynamics make fixed points and flow structure easier to identify (Zhao et al., 2017).
Sequential autoencoders provide another route to explicit latent dynamics. In one comparison between recurrent latent dynamics and NODE-based latent dynamics, the NODE formulation is
$\vx_t \in \mathbb{R}^d$0
with decoder
$\vx_t \in \mathbb{R}^d$1
The reported conclusion is not merely that NODEs reconstruct data well, but that they recover accurate low-dimensional latent dynamics, latent trajectories, and fixed point structure at the true latent dimensionality, whereas widely used RNN-based dynamics can rely on superfluous latent dimensions and spurious dynamical features (Sedler et al., 2022).
Interpretable reduced-order modeling generalizes the same logic beyond neuroscience. In iLED, the latent dynamics are decomposed into a linear operator and an explicit memory closure: $\vx_t \in \mathbb{R}^d$2 with memory state
$\vx_t \in \mathbb{R}^d$3
Interpretability here is tied to the explicit separation between a stable linear backbone and a nonlinear memory correction motivated by Mori–Zwanzig and Koopman operator theory (Menier et al., 2023).
3. State-space inference, online learning, and probabilistic observation models
A defining property of many interpretable-by-design models is the separation between latent dynamics and probabilistic observation. For neural data, one influential observation model is
$\vx_t \in \mathbb{R}^d$4
where $\vx_t \in \mathbb{R}^d$5 may be Poisson for spike counts or Gaussian for continuous signals, and $\vx_t \in \mathbb{R}^d$6 is typically $\vx_t \in \mathbb{R}^d$7 for Poisson or identity for Gaussian observations (Zhao et al., 2017). This preserves the mechanistic interpretation of the latent state while accommodating non-trivial observation noise rather than assuming a simple Gaussian readout.
Variational joint filtering derives a recursive evidence lower bound for online filtering. Starting from
$\vx_t \in \mathbb{R}^d$8
it introduces a Gaussian variational family
$\vx_t \in \mathbb{R}^d$9
with diagonal covariance, parameterized by a recognition model
$\dot{\vx} = F_\theta(\vx(t), \vu(t)), \qquad \vy(t) \sim P(\vy(t)\mid G_\theta(\vx(t), \vu(t))),$0
The ELBO contains a reconstruction log-likelihood term, an entropy term, and a dynamics log-likelihood term, and optimization uses the reparameterization trick together with stochastic gradient variational Bayes (Zhao et al., 2017). The interpretability claim is therefore compatible with an opaque inference engine: the recognition network may be black-box, while the generative dynamics remain structured and analyzable.
This distinction is elevated to a design principle in later perspective work. Integrative neurocybernetic modeling explicitly recommends keeping the generative model interpretable while allowing the inference machinery to be highly scalable and more opaque. In this view, one may use Transformers or other high-capacity networks as inference tools, but the scientific model itself should remain dynamical and mechanistic (Park et al., 26 Apr 2026). A plausible implication is that interpretability-by-design in this literature is primarily a property of the generative hypothesis class, not of every component in the computational stack.
Continuous-time probabilistic formulations extend the same idea. In latent SDE models, the hidden process evolves as
$\dot{\vx} = F_\theta(\vx(t), \vu(t)), \qquad \vy(t) \sim P(\vy(t)\mid G_\theta(\vx(t), \vu(t))),$1
while observations are generated through modality-specific decoders,
$\dot{\vx} = F_\theta(\vx(t), \vu(t)), \qquad \vy(t) \sim P(\vy(t)\mid G_\theta(\vx(t), \vu(t))),$2
The associated path-space variational inference uses Girsanov’s theorem and yields an ELBO with observation likelihood terms, initial-state KL, and path-space KL (ElGazzar et al., 2024). Because drift, diffusion, input modulation, and uncertainty are all explicit, the resulting model provides a natural language for interpretation while remaining fully probabilistic.
4. Structural design patterns for interpretability
Although explicit state-space dynamics form a common core, the literature develops several distinct structural patterns for making system dynamics legible.
One pattern is locally linear or regime-based decomposition. The Gaussian Process Switching Linear Dynamical System constructs a GP-SDE whose kernel is engineered to behave like a smoothly switching linear system. Its “Smoothly Switching Linear” kernel is
$\dot{\vx} = F_\theta(\vx(t), \vu(t)), \qquad \vy(t) \sim P(\vy(t)\mid G_\theta(\vx(t), \vu(t))),$3
with soft regime probabilities $\dot{\vx} = F_\theta(\vx(t), \vu(t)), \qquad \vy(t) \sim P(\vy(t)\mid G_\theta(\vx(t), \vu(t))),$4 given by multiclass logistic regression (Hu et al., 2024). The result is linear dynamics within each region and smooth interpolation across boundaries, preserving local linearity, identifiable regimes, and state-dependent dynamics while avoiding hard switching artifacts.
A second pattern is sparse decomposition into reusable operators. In dLDS, the latent transition is written as a sparse combination of dictionary elements. In continuous time,
$\dot{\vx} = F_\theta(\vx(t), \vu(t)), \qquad \vy(t) \sim P(\vy(t)\mid G_\theta(\vx(t), \vu(t))),$5
and in discrete time,
$\dot{\vx} = F_\theta(\vx(t), \vu(t)), \qquad \vy(t) \sim P(\vy(t)\mid G_\theta(\vx(t), \vu(t))),$6
Interpretability comes from the tractability of each operator and the time-varying sparse coefficient vector, which indicates which dynamical motifs are active and how strongly they contribute (Mudrik et al., 2022).
A third pattern is structured second-order state-space modeling with virtual sensing. Structured Kolmogorov–Arnold Neural ODEs impose
$\dot{\vx} = F_\theta(\vx(t), \vu(t)), \qquad \vy(t) \sim P(\vy(t)\mid G_\theta(\vx(t), \vu(t))),$7
The same learned function governs both latent acceleration dynamics and measured acceleration, forcing physical consistency between latent state and observation (Liu et al., 23 Jun 2025). The KAN component is then used both as a universal approximator and as a symbolic learner, so explicit governing expressions can be extracted and substituted back into the Neural ODE.
A fourth pattern is mechanistic hybridization. In latent stochastic differential equation modeling, the framework can incorporate existing mathematical models, neural networks, or a hybrid of both in the drift and diffusion functions. The coupled-oscillator hybrid is given in complex form by
$\dot{\vx} = F_\theta(\vx(t), \vu(t)), \qquad \vy(t) \sim P(\vy(t)\mid G_\theta(\vx(t), \vu(t))),$8
and in real coordinates by coupled SDEs whose parameters retain mechanistic meaning such as natural frequency and coupling strength (ElGazzar et al., 2024). In a physical systems case study for cathodic electrophoretic deposition, the neural component is restricted to simple scalar augmentations such as a coverage fraction model $\dot{\vx} = F_\theta(\vx(t), \vu(t)), \qquad \vy(t) \sim P(\vy(t)\mid G_\theta(\vx(t), \vu(t))),$9 and a structured current augmentation $\vx_{t+1} = f_\theta(\vx_t, \vu_t) + \epsilon_t.$0, both embedded inside governing equations rather than replacing them (Jacobsen et al., 2024).
5. Causality, control, and closed-loop interpretation
Interpretable-by-design modeling increasingly extends beyond open-loop latent dynamics toward causal and control-oriented formulations. In integrative neurocybernetic modeling, the brain is framed as a controller with an internal state, sensory feedback, actions on the environment, and a latent control objective (Park et al., 26 Apr 2026). The proposed meta-dynamical state-space model augments a standard SSM with a higher-level embedding: $\vx_{t+1} = f_\theta(\vx_t, \vu_t) + \epsilon_t.$1 where $\vx_{t+1} = f_\theta(\vx_t, \vu_t) + \epsilon_t.$2 indexes a family of dynamical systems and is intended to represent structured variation across datasets, individuals, tasks, or conditions (Park et al., 26 Apr 2026). Interpretability is therefore extended from state trajectories to cross-context organization and control hypotheses.
This control-oriented perspective is closely related to concept- and causal-layer decompositions proposed in hybrid neural system dynamics frameworks. One such pipeline proceeds in three stages: concept learning, causal learning, and equation learning. Concept-based interpretability maps raw inputs to semantically interpretable variables; causal machine learning and causal discovery infer a causal directed graph over concepts; mechanistic interpretability then derives interpretable structural dynamic equations informed by that graph (D'Elia, 20 May 2025). The same three-stage structure is proposed for logistic terminal automation, where raw operational data are transformed into concepts such as yard congestion, crane idleness, crane handling efficiency, terminal workload, cross-modal wait times, gate throughput, and access road bottlenecks, then into a causal graph, then into interpretable dynamic equations (D'Elia et al., 9 Sep 2025).
A related but more specific causal formulation appears in neural Granger-causal modeling. GVAR preserves the autoregressive form while replacing fixed coefficients with neural-network-produced, time-dependent generalized coefficients: $\vx_{t+1} = f_\theta(\vx_t, \vu_t) + \epsilon_t.$3 Because each $\vx_{t+1} = f_\theta(\vx_t, \vu_t) + \epsilon_t.$4 is a directly inspectable coefficient matrix, the model is interpretable by construction with respect to interaction strength, sign, and temporal variability (Marcinkevičs et al., 2021). This supports Granger-causal analysis of nonlinear multivariate dynamics without relegating causal structure to indirect feature importance.
These developments clarify an important point of debate. Some work uses “causal reliability” to distinguish intervention-supporting models from purely predictive ones (D'Elia, 20 May 2025). Others, such as GVAR, remain explicitly within the Granger-causal framework and do not claim intervention-level causality (Marcinkevičs et al., 2021). The literature therefore does not treat all causal language as equivalent. Rather, it distinguishes predictive temporal dependence, structural causal organization, and latent control objectives as different levels of mechanistic commitment.
6. Symbolic abstraction, verification, and architecture-level interpretability
Another major branch of interpretable-by-design dynamics modeling focuses not on latent scientific variables alone but on formal symbolic representations derived from learned dynamics. A transition-system abstraction framework starts from a neural dynamical model
$\vx_{t+1} = f_\theta(\vx_t, \vu_t) + \epsilon_t.$5
defines a localized working zone $\vx_{t+1} = f_\theta(\vx_t, \vu_t) + \epsilon_t.$6, partitions it into $\vx_{t+1} = f_\theta(\vx_t, \vu_t) + \epsilon_t.$7, and constructs a transition matrix $\vx_{t+1} = f_\theta(\vx_t, \vu_t) + \epsilon_t.$8 where
$\vx_{t+1} = f_\theta(\vx_t, \vu_t) + \epsilon_t.$9
Transitions are obtained by set-valued reachability analysis, with
0
The resulting artifact is a graph-like abstraction of the neural model that supports verification of CTL specifications such as 1, 2, and 3 (Yang et al., 2024).
A related dual-level neural hybrid framework combines data-driven Maximum Entropy partitioning, multiple local neural networks, and a higher-level transition abstraction. The low-level model is a neural hybrid system
4
while the high-level abstraction is
5
with transition relation computed from reachable sets (Yang et al., 2024). Here interpretability is achieved through explicit partitions, local approximations, and symbolic cell-to-cell transitions rather than through direct inspection of a monolithic function approximator.
Architecture-level interpretability is also addressed in model-combination work. A common interface, HUDA-ODE, expresses mixed hybrid universal discrete algebraic models as
6
and model composition is parameterized by interpretable connection matrices 7 between subsystems (Thummerer et al., 2024). The resulting wildcard architecture is considered easy to interpret because each block in 8 has a clear semantic role: input-to-subsystem coupling, subsystem-to-subsystem feedthrough, or subsystem-to-output contribution. This suggests that interpretability-by-design may apply not only to state representations but also to how model components are interconnected.
7. Evaluation criteria, applications, and open tensions
The evaluation logic of this literature differs from that of purely predictive sequence modeling. Output reconstruction or forecast error remains important, but many papers argue that interpretability requires additional tests: recovery of fixed points and slow points, correctness of regime boundaries, meaningful low-dimensional latent states, posterior uncertainty over vector fields, causal sign detection, or symbolic verification over abstracted dynamics (Hu et al., 2024). In the SAE comparison, state 9 is introduced precisely because firing-rate reconstruction alone does not determine whether the true latent dynamics have been recovered (Sedler et al., 2022). In gpSLDS, posterior uncertainty over $\vx_{t+1} = \vx_t + \vg(\vx_t) + \vB(\vx_t)\vu_t + \epsilon_t,$0 is treated as part of the interpretability story because it indicates where fixed points or boundaries are statistically supported (Hu et al., 2024).
Applications span several domains. In neuroscience, the learned structures are linked to head-direction coding, decision making, rhythmic activity, working memory, aggression-related hypothalamic dynamics, monkey LIP decision boundaries, V1 oscillatory latent dynamics, delayed reach preparation, and full-brain C. elegans activity (Zhao et al., 2017). In engineering and physical systems, interpretable-by-design dynamics modeling appears in multiscale PDE reduction, e-coating dynamics, human handwriting verification, bouncing-ball hybrid modeling, audio system identification, acoustic echo cancellation, autonomous multimodal transportation, and logistic terminal automation (Menier et al., 2023).
Several tensions recur across the literature. One is the trade-off between expressive flexibility and structural constraint. gpSLDS explicitly aims to sit between generic Gaussian-process latent models and discrete switching linear models (Hu et al., 2024). SKANODE argues that standard NODEs, ANODEs, and SONODEs remain black-box because their latent variables are not guaranteed to correspond to physical quantities (Liu et al., 23 Jun 2025). iLED seeks comparable accuracy to state-of-the-art recurrent neural network-based approaches while replacing black-box recurrence with a theory-grounded latent structure (Menier et al., 2023). These examples suggest that the field treats interpretability as a constraint on admissible representations rather than a simple sparsity preference.
Another tension concerns where opacity is permissible. Some work allows black-box recognition models, neural re-parametrization, or teacher models, provided the executed scientific model remains transparent (Zhao et al., 2017). “Neural generation and interpretable execution” formalizes this principle by separating a neural generation stage from a transparent execution stage: $\vx_{t+1} = \vx_t + \vg(\vx_t) + \vB(\vx_t)\vu_t + \epsilon_t,$1 In that framework, interpretability is defined as inference equivariance, and scalable verification is achieved by treating interpretability as a Markovian property over local transitions rather than verifying a full global mapping (Barbiero et al., 17 Feb 2025). This suggests a broader theoretical view in which interpretable-by-design modeling is compatible with expressive neural machinery, provided the mechanism of inference or execution remains locally checkable and human-legible.
A final open issue is that many frameworks remain proposals or partial roadmaps rather than fully standardized algorithms. The INSD pipeline in transportation explicitly does not yet report completed empirical validation, benchmark metrics, or an end-to-end implemented training procedure (D'Elia, 20 May 2025). Integrative neurocybernetic modeling is largely a perspective articulating nonlinear SSMs, meta-dynamical extensions, scalable inference, knowledge distillation, mixed open/closed-loop training, and connectomics-informed priors as building blocks rather than a single algorithm (Park et al., 26 Apr 2026). This suggests that the field is still consolidating around design principles rather than converging on one canonical architecture.
Taken together, interpretable-by-design neural system dynamics modeling is best understood as a modeling philosophy with several mature technical instantiations. Its defining move is to encode dynamical, causal, structural, or symbolic hypotheses directly into the model class, so that learned representations are inspectable in scientifically meaningful coordinates. Whether the chosen primitives are vector fields, locally linear regimes, sparse operator dictionaries, explicit control objectives, concept-level causal graphs, symbolic equations, or transition systems, the unifying claim is that explanation should emerge from the form of the model itself rather than from post-hoc narration of an opaque computation (Zhao et al., 2017, Park et al., 26 Apr 2026).