Latent Forcing: Modeling, Inference, Applications

• Latent forcing is a modeling paradigm that incorporates Gaussian process priors with differential equations to infer system dynamics and unobserved driving forces.
• It employs state-space reformulations and advanced inference techniques like Kalman filtering and variational methods to efficiently estimate states and latent processes.
• The approach underpins diverse applications—from sequential generative modeling and causal attribution in climate extremes to diffusion-based image generation with notable performance improvements.

Latent forcing is a modeling paradigm in which unknown, typically unobservable, driving functions—termed latent forces—are introduced into mechanistic systems, such as differential equations, and modeled with flexible nonparametric priors, most commonly Gaussian processes (GPs). This hybrid approach augments interpretable, physics- or engineering-inspired operators with data-driven components. Latent force models (LFMs) and their extensions enable the joint inference of system parameters, states, and the unmeasured inputs themselves. The concept of latent forcing has become a cornerstone of contemporary approaches to machine learning-driven spatiotemporal modeling, sequential generative modeling, causal attribution, and efficient diffusion-based image generation.

1. Fundamental Formulation of Latent Force Models

The classical LFM couples known mechanistic dynamics with Gaussian process priors over unobserved forcing terms. For a linear, time-invariant system, the most general continuous-time form is: $D[f](t) = \sum_{k=1}^K S_k u_k(t) + \epsilon(t)$ where $D$ is a known linear differential operator, $S_k$ are sensitivity constants, $u_k(t)$ are independent latent forcing functions $u_k(\cdot) \sim \mathcal{GP}(0, k_k(\cdot,\cdot))$, and $\epsilon(t)$ is observational noise (Reece et al., 2013, Hartikainen et al., 2012).

In multiplicative LFMs, the state-dependent drift becomes: $$\frac{d\mathbf{x}(t)}{dt} = \left(\mathbf{A}_0 + \sum_{r=1}^R \mathbf{A}_r g_r(t) \right) \mathbf{x}(t)$$ where each GP-driven $g_r(t)$ modulates a matrix-valued, possibly Lie algebra–constrained drift, yielding non-Gaussian state geometry and enabling group-structured stochastic flows (Tait et al., 2018).

This paradigm generalizes naturally to non-linear ODEs and parabolic PDEs with additive or multiplicative latent forcing, leading to models of considerable expressivity and scientific relevance (Moss et al., 2021).

2. State-Space Reformulation and Inference Algorithms

LFMs admit an exact or approximate reduction to linear state-space models, provided the GP prior's kernel admits an SDE representation (e.g., Matérn, approximated squared-exponential). This leads to the following augmented stochastic system: $dX_a(t) = F_a X_a(t) \, dt + L_a \, dW_a(t)$ where $X_a(t)$ aggregates the physical state and all latent-forcing states, and the matrices $D$0 encode system and GP dynamics (Hartikainen et al., 2012). For discrete observations, inference proceeds via Kalman filtering and Rauch–Tung–Striebel (RTS) smoothing. The forward-backward recursion yields full posteriors for both the state and latent force at every timestep, with per-step cubic cost in the total state dimension (Reece et al., 2013, Wolff et al., 23 Dec 2025).

For nonlinear systems, approximations such as the extended or unscented Kalman filters, or variational inference with expressive multimodal posteriors (e.g., local inverse autoregressive flows), maintain tractable updates while enabling joint inference over parameters, latent trajectories, and observations (Ward et al., 2019, Wolff et al., 23 Dec 2025).

3. Latent Forcing in Contemporary Machine Learning Architectures

Latent forcing principles have been adapted and generalized outside classical physical modeling.

• Stochastic recurrent networks and sequence modeling: Z-Forcing augments each step in an RNN with a stochastic latent variable, which is explicitly forced via an auxiliary cost to carry information about future states (backward RNN), thus driving temporal variability and preventing posterior collapse. The auxiliary forcing cost ensures that latent variables meaningfully encode both past and future information, yielding improved generative and predictive performance (Goyal et al., 2017).
• Multimodal LLMs (MLLMs): In video understanding, latent self-forcing compels MLLMs to internalize a sequence of continuous latent vectors (visual "thoughts") aligned with both video and language context, in the absence of explicit supervision. This inductive signal is introduced via latent-alignment (contrastive) and diversity (decorrelation) losses applied to the model’s internal latent sequence, ensuring grounding and promoting coverage without reliance on annotated reasoning traces (Hu et al., 22 Jun 2026).
• Diffusion models and generative image modeling: Latent Forcing introduces a two-stream denoising trajectory whereby latent embeddings, acting as a scratchpad, are generated and denoised before the model attends to high-frequency pixel features. The order of latent and pixel denoising is treated as a critical architectural degree of freedom, providing both fast convergence and state-of-the-art fidelity in pixel-space image generation (Baade et al., 11 Feb 2026).

4. Latent Forcing for Causal and Spatiotemporal Phenomena

Latent forcing admits extension to high-dimensional, structured domains such as climate science and SPDEs:

• Causal inference in climate extremes: Anthropogenic forcing is modeled as a latent spatially-varying effect within a hierarchical Bayesian framework. The approach uses a latent Gaussian spatial model, specifically a multivariate intrinsic conditional autoregressive prior, to encode spatial dependencies in extreme event statistics. A highly efficient "max-and-smooth" Laplace–Gibbs Bayesian inference reconstructs both factual and counterfactual statistics, enabling the attribution of causal effects to latent anthropogenic forcing and the delineation of credible spatial hotspots of impact (Giri et al., 25 Apr 2026).
• Stochastic partial differential equations: In learning SPDEs solely from solution realizations, the latent forcing (additive Gaussian noise) is projected using a spectral Galerkin scheme and expanded with a truncated Wiener chaos basis. The resulting parameterized latent system admits variational Bayes inference on both deterministic propagators and stochastic forcing without observing the driving Wiener process, yielding sharp recovery of both law-level and realization-level statistics (Zeng et al., 12 Feb 2026).

5. Basis Expansions and Computational Efficiency

When the GP kernel of the latent force is periodic or quasi-periodic, a finite trigonometric or eigenfunction basis efficiently approximates the prior. Basis coefficients are then treated as latent states, drastically reducing dimensionality and enabling O(T) inference scaling, as opposed to naive cubic scaling in standard GP regression: $D$1 with $D$2 sinusoidal or eigenfunction basis and $D$3 Gaussian coefficients (Reece et al., 2013).

Spectral approaches—Nyström/KPCA for determining basis functions—outperform random Fourier features for periodic processes. Extensions further allow for time-adaptive bases, quasi-periodic latents (basis weights evolving as GPs), and multi-periodic expansions in block-diagonalized state spaces (Reece et al., 2013).

6. Optimal State Estimation and Constraints

Optimal state estimation frameworks—full-information estimation (FIE) and delayed moving-horizon estimation (dMHE)—enable LFMs to integrate system-inherent constraints, such as non-negativity or boundedness, directly within the inference loop. These recast the latent-force inference problem as a constrained nonlinear program over both state and latent trajectories, solved via direct transcription and NLP solvers (Wolff et al., 23 Dec 2025). This ensures physical plausibility and superior empirical accuracy in domains with strict operative constraints (e.g., endocrinology, aerospace trajectories).

7. Empirical Benchmarks, Applications, and Extensions

LFMs outperform non-periodic, non-forced, and resonator-based models across a range of applications, such as:

Application	LFM Type	RMSE / Metric	Relative Gain
Queueing (M/M/1)	Q-Periodic SQM	RMSE = 1.8 ± 0.2	17% RMSE of non-periodic baseline
Thermal dynamics (building)	Periodic, SQM	RMSE = 0.52–0.59	45% RMSE of resonator model
Speech, MNIST (sequence gen.)	Z-Forcing aux	ELBO > baseline, lower PPL	Non-zero KL, posterior usage
Video-Language (MLLM)	Self-Forcing	2–6 pts > prior MLLMs	68× faster inference
ImageNet diffusion	Latent Forcing	FID₅₀ₖ = 4.18	SOTA at this compute scale

Strictly periodic, quasi-periodic, and composite latent components admit efficient Kalman-based and operator-based inference via finite-basis representation and state-space augmentation. Neural and Fourier operators amortize latent inference across thousands of instances, providing orders of magnitude speedup in high-throughput regimes (Moss et al., 2021).

Possible extensions include adaptive bases (updating with posterior data), composite/multimodal latent forcing, general control/MPG applications, and continuous-time infinite-dimensional latent spaces.

---

References

• (Reece et al., 2013) Efficient State-Space Inference of Periodic Latent Force Models
• (Hartikainen et al., 2012) Sequential Inference for Latent Force Models
• (Wolff et al., 23 Dec 2025) Inference in Latent Force Models Using Optimal State Estimation
• (Tait et al., 2018) Approximate Inference for Multiplicative Latent Force Models
• (Ward et al., 2019) Black-Box Inference for Non-Linear Latent Force Models
• (Moss et al., 2021) Approximate Latent Force Model Inference
• (Goyal et al., 2017) Z-Forcing: Training Stochastic Recurrent Networks
• (Giri et al., 25 Apr 2026) Estimating Causal Attribution of Anthropogenic Forcing on High-Temperature Extremes Using a Latent Gaussian Spatial Model
• (Zeng et al., 12 Feb 2026) Latent-Variable Learning of SPDEs via Wiener Chaos
• (Hu et al., 22 Jun 2026) VideoLatent: Video-Language Learning via Latent Self-Forcing
• (Baade et al., 11 Feb 2026) Latent Forcing: Reordering the Diffusion Trajectory for Pixel-Space Image Generation