Manifold Trajectory Kinetics (MTK) Overview

• Manifold Trajectory Kinetics (MTK) is a framework that analyzes system dynamics by optimizing trajectories on structured manifolds to capture task-relevant features.
• It is applied across diverse domains such as chemical kinetics, low-rank kinetic simulations, robotics, and large-model safety, each adapting the manifold concept to its specific challenges.
• MTK leverages geometric, variational, and learning-based formulations to reduce high-dimensional dynamics, offering improved detection, simulation, and control performance in complex systems.

Searching arXiv for the cited MTK-related papers to ground the article in the relevant literature. arxiv.search({"query":"id:(Dietrich et al., 2022) OR id:(Lee, 2024) OR id:(Pei et al., 27 May 2026) OR id:(Einkemmer et al., 2024) OR id:(Huguet et al., 2022) OR id:(Lebiedz et al., 2014) OR id:(Szczapa et al., 2022) OR id:(Zhang et al., 5 Jun 2026)","max_results":10,"sort_by":"submittedDate","sort_order":"descending"}) I found the following relevant arXiv records:

• "Defending Jailbreak Attacks on LLMs via Manifold Trajectory Kinetics" (Zhang et al., 5 Jun 2026)
• "Learning a Kinodynamic Trajectory Manifold for Impact-Aware Compliant Catching of Fast-Moving Objects" (Pei et al., 27 May 2026)
• "A review of low-rank methods for time-dependent kinetic simulations" (Einkemmer et al., 2024)
• "Trajectory Manifold Optimization for Fast and Adaptive Kinodynamic Motion Planning" (Lee, 2024)
• "Approximating normally attracting invariant manifolds using trajectory-based optimization" (Dietrich et al., 2022)
• "Automatic Estimation of Self-Reported Pain by Trajectory Analysis in the Manifold of Fixed Rank Positive Semi-Definite Matrices" (Szczapa et al., 2022)
• "Manifold Interpolating Optimal-Transport Flows for Trajectory Inference" (Huguet et al., 2022)
• "On fundamental unifying concepts for trajectory-based slow invariant attracting manifold computation in multiscale models of chemical kinetics" (Lebiedz et al., 2014) Manifold Trajectory Kinetics (MTK) denotes, across the works assembled here, a trajectory-based and manifold-centric treatment of dynamics in which the primary object is not an isolated state but the evolution of states on a structured manifold. Depending on the domain, that manifold is an invariant slow manifold in chemical kinetics, a fixed-rank or tensor low-rank manifold in kinetic PDEs, a learned kinodynamic trajectory manifold in robotics, a latent data manifold in trajectory inference, a quotient manifold of fixed-rank positive semidefinite matrices in motion analysis, or a layerwise neighborhood manifold in large-model safety. The designation is explicit in LLM/VLM jailbreak detection (Zhang et al., 5 Jun 2026); in several other sources it is a contextual synthesis rather than the paper’s own term, but the recurring construction is the same: reduce high-dimensional dynamics by optimizing, evolving, or comparing trajectories that are constrained by manifold geometry (Dietrich et al., 2022, Einkemmer et al., 2024, Lee, 2024, Huguet et al., 2022).

1. Terminological status and domain coverage

MTK is not a uniformly standardized label across the cited literature. In "Defending Jailbreak Attacks on LLMs via Manifold Trajectory Kinetics" (Zhang et al., 5 Jun 2026), it is the paper’s explicit method name. In several other works, the term does not appear, but the underlying methodology is described as a trajectory-based slow manifold approximation, a low-rank manifold evolution, a learned trajectory manifold for kinodynamic planning, a latent Neural ODE flow on a data manifold, or a manifold-valued trajectory analysis framework (Dietrich et al., 2022, Lebiedz et al., 2014, Lee, 2024, Einkemmer et al., 2024, Huguet et al., 2022, Szczapa et al., 2022, Pei et al., 27 May 2026).

A concise way to organize these usages is as follows.

Domain	Manifold object	Trajectory mechanism
Chemical kinetics	Slow invariant manifold or NAIM	Trajectory optimization, reverse-mode BVPs
Time-dependent kinetic simulation	Rank-$r$ matrix/tensor manifold	DLR tangent projection, SAT step-and-truncate
Robotics	Learned kinodynamic trajectory manifold	Offline manifold learning, TMO, latent decoding
Trajectory inference and motion analysis	Latent manifold or fixed-rank PSD manifold	Neural ODE interpolation, manifold alignment
LLM/VLM safety	Layerwise neighborhood structure of hidden states	Rank-trajectory anomaly detection

This suggests a useful unifying description: MTK studies dynamics by restricting attention to trajectories that either lie on, approximate, or are interpreted through a manifold whose geometry captures the task-relevant degrees of freedom.

2. Invariant-manifold formulations in chemical kinetics

In chemical kinetics, MTK is most directly associated with trajectory-based slow manifold computation. A coordinate-free formulation models a smooth autonomous ODE as a smooth injective semiflow $\{\phi^t\}_{t\ge 0}$ on a complete Riemannian manifold $(Q,g)$ with infinitesimal generator $X$, thereby allowing nonlinear constraints such as adiabatic conditions to be handled intrinsically (Dietrich et al., 2022). The basic coordinate-free comparison set is the velocity-level constraint

$$K_\varepsilon := \{\, p \in Q : \| X_p \| = \varepsilon \,\}, \qquad \varepsilon>0,$$

which avoids a prescribed slow–fast coordinate split and excludes equilibria (Dietrich et al., 2022).

The Lyapunov-based formulation is built from finite-time Lyapunov exponents and their adjoint counterparts,

$$\lambda^t(v_p)=\frac{1}{|t|}\log\frac{\|d^t(v_p)\|}{\|v_p\|}, \qquad \tilde\lambda^t(\omega_p)=\frac{1}{|t|}\log\frac{\|d(\phi^{-t})^*\omega_p\|}{\|\omega_p\|},$$

and from finite-horizon objectives such as

$$F_T^+(p) := \sup_{t \in (0,T)\cap \mathrm{Def}(p)} f_t(p).$$

Within this framework, minimizers of the finite-time objectives approximate nonuniformly normally attracting orbits and, more generally, nonuniformly normally attracting invariant manifolds (NAIMs) (Dietrich et al., 2022).

The relevant invariant-manifold structure is a splitting

$TQ|_M = TM \oplus E,$

together with the normal-attraction inequality

$$\|d^t(w_p)\|\,\|v_p\| \le C(p)\,e^{-t\nu}\,\|d^t(v_p)\|\,\|w_p\|, \qquad t\ge 0.$$

This formalizes the geometric balance that trajectory optimization is intended to recover: tangential dynamics remain comparatively slow while transverse directions contract more strongly (Dietrich et al., 2022).

A complementary unifying account identifies two basic concepts behind trajectory-based slow invariant manifold computation: derivative-of-the-state-vector criteria that eliminate fast modes, and boundary-value formulations that exploit manifold attractivity (Lebiedz et al., 2014). The first concept includes the Zero-Derivative Principle, Flow Curvature Method, ILDM/FET-type constructions, stretching-based diagnostics, QSSA/PEA, and trajectory optimization with $\Phi(z)=\|J_S(z)S(z)\|_2^2$. The second concept constructs manifold points as endpoints of trajectories satisfying mixed endpoint constraints, typically in reverse mode with reaction progress variables fixed at $\{\phi^t\}_{t\ge 0}$0 and upstream conditions moved to $\{\phi^t\}_{t\ge 0}$1 when feasible (Lebiedz et al., 2014).

The variational form is

$\{\phi^t\}_{t\ge 0}$2

with choices such as $\{\phi^t\}_{t\ge 0}$3 or, more generally, $\{\phi^t\}_{t\ge 0}$4 (Lebiedz et al., 2014). The associated Hamiltonian formulation,

$\{\phi^t\}_{t\ge 0}$5

produces adjoint equations and mixed transversality conditions that make the method an optimal-control BVP rather than a local algebraic approximation (Lebiedz et al., 2014).

This line of work is exemplified by the Davis–Skodje benchmark, Michaelis–Menten enzymatic kinetics, and an isothermal hydrogen combustion mechanism. In these cases, increasing the horizon sharpens manifold approximation, and the trajectory bundles cluster along the attracting slow set (Dietrich et al., 2022). A plausible implication is that MTK in this setting is best understood as an invariance-oriented replacement for local timescale diagnostics: it seeks actual trajectories whose growth-rate balances satisfy NAIM-type inequalities, rather than merely identifying slow subspaces.

3. Low-rank kinetic simulation as manifold-constrained evolution

For time-dependent kinetic PDEs, MTK takes the form of evolving the solution on a low-rank manifold rather than on the full six-dimensional phase-space grid. The unknown phase-space density $\{\phi^t\}_{t\ge 0}$6 is governed by equations of the form

$\{\phi^t\}_{t\ge 0}$7

with representative cases including Vlasov–Poisson, Vlasov–Maxwell, radiative transfer, BGK/Fokker–Planck models, and charged-particle transport (Einkemmer et al., 2024).

The manifold viewpoint begins from the matrix rank-$\{\phi^t\}_{t\ge 0}$8 set

$\{\phi^t\}_{t\ge 0}$9

or its continuous analogue in separable $(Q,g)$0 variables. In dynamical low-rank (DLR) form, one writes

$(Q,g)$1

with orthonormal gauge conditions $(Q,g)$2 and $(Q,g)$3, and evolves the approximation by the tangent-projected equation

$(Q,g)$4

The corresponding factor ODEs are

$(Q,g)$5

although robust integrators avoid explicit $(Q,g)$6 when singular values are small (Einkemmer et al., 2024).

The review distinguishes two principal MTK realizations. DLR is a continuous-time manifold flow that projects the kinetic RHS onto the tangent bundle of the rank-$(Q,g)$7 manifold. SAT is a discrete step-and-project method that first advances the full solution by a conventional solver and then truncates back to the low-rank manifold by SVD, HOSVD, TT, or HT projection (Einkemmer et al., 2024). In matrix form, the best rank-$(Q,g)$8 truncation is controlled by the Eckart–Young–Mirsky relation

$(Q,g)$9

The same review emphasizes projector-splitting integrators, BUG and augmented BUG schemes, tensor Tucker/HT/TT extensions, conservative truncation strategies, and macro–micro formulations for preserving mass, momentum, and energy in appropriate settings (Einkemmer et al., 2024). Applications include collisionless advection, kinetic shear Alfvén waves, bump-on-tail and two-stream instabilities, diffusion-limit radiative transfer, and electron radiation therapy.

Within this PDE literature, MTK is not a named algorithm but a geometric principle: the physically relevant time history is a trajectory on a low-rank manifold. This suggests a strong conceptual continuity with slow-manifold kinetics. In both cases, the ambient dynamics are high-dimensional and stiff, and the computational strategy is to evolve only the manifold-constrained degrees of freedom while preserving as much invariance or structure as possible.

4. Learned kinodynamic trajectory manifolds in robotics

In robotics, MTK appears as the modeling and optimization of kinodynamic behavior on a learned low-dimensional trajectory manifold. In "Trajectory Manifold Optimization for Fast and Adaptive Kinodynamic Motion Planning" (Lee, 2024), the core problem is the reduction of the high-dimensional trajectory search space by learning, offline, a manifold of continuous-time, differentiable trajectories satisfying kinodynamic constraints. The state and control are

$X$0

with inverse dynamics

$X$1

The manifold is encoded by Differentiable Motion Manifold Primitives, whose decoder has the basis-function form

$X$2

so that $X$3, $X$4, $X$5, and torques can be obtained by differentiation and inverse dynamics (Lee, 2024).

The planning objective on the manifold is

$X$6

with task-specific penalties plus constraint terms on joint limits, velocities, accelerations, jerks, Cartesian velocities, torques, and self-collision. The offline pipeline consists of data collection via trajectory optimization, manifold learning, latent flow learning, and Trajectory Manifold Optimization (TMO) for decoder fine-tuning (Lee, 2024). In the reported 7-DoF dynamic throwing task, 3,523 trajectories are collected; the latent dimension is $X$7; the decoder uses $X$8 basis functions; fixed $X$9 s is used for manifold training; and the approach achieves a sampling time of $$K_\varepsilon := \{\, p \in Q : \| X_p \| = \varepsilon \,\}, \qquad \varepsilon>0,$$0 s for 100 trajectories and $$K_\varepsilon := \{\, p \in Q : \| X_p \| = \varepsilon \,\}, \qquad \varepsilon>0,$$1 s total with rejection sampling and constraint checking. After TMO and rejection sampling, the reported success rate is 100% on seen and unseen tasks, with error $$K_\varepsilon := \{\, p \in Q : \| X_p \| = \varepsilon \,\}, \qquad \varepsilon>0,$$2 m and 100% constraint satisfaction (Lee, 2024).

A second robotic realization shifts from online optimization to direct state-to-trajectory synthesis. "Learning a Kinodynamic Trajectory Manifold for Impact-Aware Compliant Catching of Fast-Moving Objects" (Pei et al., 27 May 2026) constructs a low-dimensional conditional manifold

$$K_\varepsilon := \{\, p \in Q : \| X_p \| = \varepsilon \,\}, \qquad \varepsilon>0,$$3

where each catching trajectory $$K_\varepsilon := \{\, p \in Q : \| X_p \| = \varepsilon \,\}, \qquad \varepsilon>0,$$4 is a fixed-horizon sequence of joint positions and velocities aligned at interception. The local manifold geometry is given by the pullback metric

$$K_\varepsilon := \{\, p \in Q : \| X_p \| = \varepsilon \,\}, \qquad \varepsilon>0,$$5

and a learned state-to-latent map

$$K_\varepsilon := \{\, p \in Q : \| X_p \| = \varepsilon \,\}, \qquad \varepsilon>0,$$6

enables direct synthesis

$$K_\varepsilon := \{\, p \in Q : \| X_p \| = \varepsilon \,\}, \qquad \varepsilon>0,$$7

without online nonlinear optimization (Pei et al., 27 May 2026).

That system is coupled with impact-aware compliant control,

$$K_\varepsilon := \{\, p \in Q : \| X_p \| = \varepsilon \,\}, \qquad \varepsilon>0,$$8

with gains scheduled by hand–object distance to soften impact (Pei et al., 27 May 2026). In MuJoCo simulation, the reported results are a success rate of 85.1% versus 54.7% for the baseline, peak contact force of 32.1 N versus 120.7 N, and net force of 4.3 N versus 9.85 N. Under uniform perturbations in the estimated object state, success remains above 70% for position errors up to 0.05 m and velocity errors up to 0.3 m/s (Pei et al., 27 May 2026).

In both robotic cases, the term MTK is a contextual mapping rather than the original paper’s terminology. Even so, the common principle is explicit: feasible dynamics are not searched in the ambient trajectory space but on a learned manifold whose geometry is made compatible with kinodynamic or impact-aware constraints.

5. Data-manifold trajectories and manifold-valued observation analysis

A different MTK line treats the manifold as an inferred geometric substrate for population dynamics or measured motion. "Manifold Interpolating Optimal-Transport Flows for Trajectory Inference" (Huguet et al., 2022) learns continuous-time stochastic population dynamics from sporadic snapshot samples by combining manifold learning, optimal transport, and Neural ODEs. A Geodesic Autoencoder learns a latent space in which Euclidean distances approximate a multiscale geodesic distance on the data manifold, and the latent dynamics are given by

$$K_\varepsilon := \{\, p \in Q : \| X_p \| = \varepsilon \,\}, \qquad \varepsilon>0,$$9

The full loss is

$$\lambda^t(v_p)=\frac{1}{|t|}\log\frac{\|d^t(v_p)\|}{\|v_p\|}, \qquad \tilde\lambda^t(\omega_p)=\frac{1}{|t|}\log\frac{\|d(\phi^{-t})^*\omega_p\|}{\|\omega_p\|},$$0

with optional stochasticity through

$$\lambda^t(v_p)=\frac{1}{|t|}\log\frac{\|d^t(v_p)\|}{\|v_p\|}, \qquad \tilde\lambda^t(\omega_p)=\frac{1}{|t|}\log\frac{\|d(\phi^{-t})^*\omega_p\|}{\|\omega_p\|},$$1

The method is evaluated on synthetic branching data, Dyngen differentiation, embryoid body scRNA-seq, and AML chemotherapy response, with, for example, held-out Petal interpolation achieving $$\lambda^t(v_p)=\frac{1}{|t|}\log\frac{\|d^t(v_p)\|}{\|v_p\|}, \qquad \tilde\lambda^t(\omega_p)=\frac{1}{|t|}\log\frac{\|d(\phi^{-t})^*\omega_p\|}{\|\omega_p\|},$$2, $$\lambda^t(v_p)=\frac{1}{|t|}\log\frac{\|d^t(v_p)\|}{\|v_p\|}, \qquad \tilde\lambda^t(\omega_p)=\frac{1}{|t|}\log\frac{\|d(\phi^{-t})^*\omega_p\|}{\|\omega_p\|},$$3, and runtime $$\lambda^t(v_p)=\frac{1}{|t|}\log\frac{\|d^t(v_p)\|}{\|v_p\|}, \qquad \tilde\lambda^t(\omega_p)=\frac{1}{|t|}\log\frac{\|d(\phi^{-t})^*\omega_p\|}{\|\omega_p\|},$$4 s, and Dyngen yielding $$\lambda^t(v_p)=\frac{1}{|t|}\log\frac{\|d^t(v_p)\|}{\|v_p\|}, \qquad \tilde\lambda^t(\omega_p)=\frac{1}{|t|}\log\frac{\|d(\phi^{-t})^*\omega_p\|}{\|\omega_p\|},$$5 and $$\lambda^t(v_p)=\frac{1}{|t|}\log\frac{\|d^t(v_p)\|}{\|v_p\|}, \qquad \tilde\lambda^t(\omega_p)=\frac{1}{|t|}\log\frac{\|d(\phi^{-t})^*\omega_p\|}{\|\omega_p\|},$$6 in the main table (Huguet et al., 2022). Here MTK is a manifold-ground transport flow whose kinetics are encoded by a latent vector field.

"Automatic Estimation of Self-Reported Pain by Trajectory Analysis in the Manifold of Fixed Rank Positive Semi-Definite Matrices" (Szczapa et al., 2022) realizes MTK in a quotient Riemannian geometry for facial landmarks. For each frame, mean-centered positions $$\lambda^t(v_p)=\frac{1}{|t|}\log\frac{\|d^t(v_p)\|}{\|v_p\|}, \qquad \tilde\lambda^t(\omega_p)=\frac{1}{|t|}\log\frac{\|d(\phi^{-t})^*\omega_p\|}{\|\omega_p\|},$$7 and instantaneous velocities $$\lambda^t(v_p)=\frac{1}{|t|}\log\frac{\|d^t(v_p)\|}{\|v_p\|}, \qquad \tilde\lambda^t(\omega_p)=\frac{1}{|t|}\log\frac{\|d(\phi^{-t})^*\omega_p\|}{\|\omega_p\|},$$8 are stacked into

$$\lambda^t(v_p)=\frac{1}{|t|}\log\frac{\|d^t(v_p)\|}{\|v_p\|}, \qquad \tilde\lambda^t(\omega_p)=\frac{1}{|t|}\log\frac{\|d(\phi^{-t})^*\omega_p\|}{\|\omega_p\|},$$9

and represented by the Gram matrix

$$F_T^+(p) := \sup_{t \in (0,T)\cap \mathrm{Def}(p)} f_t(p).$$0

where

$$F_T^+(p) := \sup_{t \in (0,T)\cap \mathrm{Def}(p)} f_t(p).$$1

Trajectory smoothing is performed by blended cubic Bézier fitting on the manifold, temporal alignment is handled by the Global Alignment Kernel, and regression is done by SVR with late fusion across facial regions (Szczapa et al., 2022).

The geometry is encoded through the Bures/Wasserstein-type distance

$$F_T^+(p) := \sup_{t \in (0,T)\cap \mathrm{Def}(p)} f_t(p).$$2

or equivalently

$$F_T^+(p) := \sup_{t \in (0,T)\cap \mathrm{Def}(p)} f_t(p).$$3

On the UNBC-McMaster dataset, late fusion with augmentation achieves MAE $$F_T^+(p) := \sup_{t \in (0,T)\cap \mathrm{Def}(p)} f_t(p).$$4 and RMSE $$F_T^+(p) := \sup_{t \in (0,T)\cap \mathrm{Def}(p)} f_t(p).$$5 under LOSO, while on BioVid Part A the reported late-fusion result is MAE $$F_T^+(p) := \sup_{t \in (0,T)\cap \mathrm{Def}(p)} f_t(p).$$6 and RMSE $$F_T^+(p) := \sup_{t \in (0,T)\cap \mathrm{Def}(p)} f_t(p).$$7 under subject-wise 3-fold evaluation (Szczapa et al., 2022).

These two cases widen the scope of MTK substantially. The manifold no longer needs to be an invariant set of an ODE in the chemical sense; it may instead be a learned latent geometry for population transport or a quotient manifold of observation-derived tensors. What remains constant is that dynamics are inferred or compared through manifold-respecting trajectories rather than framewise Euclidean features.

6. Explicit MTK in LLM and VLM jailbreak detection

The only source in which MTK is named directly is "Defending Jailbreak Attacks on LLMs via Manifold Trajectory Kinetics" (Zhang et al., 5 Jun 2026). There the large model is treated as a kinetic system that transforms a prompt across transformer layers, and detection is based on how the prompt’s neighborhood structure evolves on the hidden-state manifold. For a prompt $$F_T^+(p) := \sup_{t \in (0,T)\cap \mathrm{Def}(p)} f_t(p).$$8, the layerwise representation is the last-token hidden state

$$F_T^+(p) := \sup_{t \in (0,T)\cap \mathrm{Def}(p)} f_t(p).$$9

Given benign anchor bank $TQ|_M = TM \oplus E,$0 and malicious anchor bank $TQ|_M = TM \oplus E,$1, MTK ranks distances from $TQ|_M = TM \oplus E,$2 to all anchor representations and computes the mean benign-neighbor rank

$TQ|_M = TM \oplus E,$3

yielding the rank trajectory

$TQ|_M = TM \oplus E,$4

A benign-only Isolation Forest is then trained on these trajectories, and a threshold $TQ|_M = TM \oplus E,$5 is chosen from the $TQ|_M = TM \oplus E,$6-quantile of benign scores (Zhang et al., 5 Jun 2026).

The explanatory signal often visualized is

$TQ|_M = TM \oplus E,$7

with jailbreak prompts exhibiting a depth-dependent “bumping” pattern: they begin near malicious neighborhoods and later shift toward benign neighborhoods to evade refusal, sometimes with multiple sign changes in $TQ|_M = TM \oplus E,$8 (Zhang et al., 5 Jun 2026). This is the paper’s core departure from fixed-slice detectors based on raw inputs, gradients, or single hidden layers.

The empirical results are unusually explicit. On pseudo-malicious prompts, the method attains a jailbreak true positive rate of 95% at a false positive rate of 5% on benign prompts and 2% on pseudo-malicious prompts (Zhang et al., 5 Jun 2026). Under adaptive attacks, it maintains a true positive rate of 85%. Across four LLMs and ten jailbreak attacks, it achieves mean AUROC values of 0.940 on LLaMA2-7B, 0.944 on LLaMA3-8B, 0.953 on Mistral-7B, and 0.923 on Vicuna-7B, with the best AUROC in 31 of 40 model-attack settings (Zhang et al., 5 Jun 2026). Against the strongest MTK-targeted surrogate, reported evasion attack success rates are 0.17 on LLaMA2 and 0.13 on Vicuna. The same pipeline extends to VLMs, with average AUROC 0.940 for LLaVA-1.6-7B and 0.924 for Qwen-VL-Chat, and multimodal pseudo-malicious false positive rates of 0.014 and 0.073, respectively (Zhang et al., 5 Jun 2026).

Here MTK is neither model reduction nor latent manifold learning in the classical sense. It is a manifold-of-neighborhoods view of transformer inference: the kinetic trajectory is the sequence of local geometric relations induced by the forward pass.

7. Unifying principles, limitations, and open questions

A plausible unifying description is that MTK treats a manifold as the correct geometric state space for trajectory analysis when ambient coordinates are either too high-dimensional, too stiff, too noisy, or too fragile under adversarial perturbation. The manifold may be invariant and attracting, as in NAIM/SIM theory; algebraic and low-rank, as in DLR and SAT; learned and conditional, as in DMMP and catching manifolds; quotient-Riemannian, as in fixed-rank PSD trajectory analysis; or neighborhood-induced, as in LLM hidden-state rank trajectories (Dietrich et al., 2022, Einkemmer et al., 2024, Szczapa et al., 2022, Zhang et al., 5 Jun 2026).

A second unifying point is that MTK is trajectory-first. In the chemical-kinetic setting, this means optimizing over actual ODE trajectories or their endpoint-constrained bundles (Lebiedz et al., 2014). In low-rank kinetic simulation, it means following tangent-projected manifold flows or repeated retractions after each time step (Einkemmer et al., 2024). In robotics, it means decoding or refining whole continuous-time trajectories rather than solving the full ambient-space problem online (Lee, 2024, Pei et al., 27 May 2026). In MIOFlow, it means learning a vector field whose pushforwards match observed marginals (Huguet et al., 2022). In LLM safety, it means classifying the depthwise evolution of local neighborhoods rather than any single feature slice (Zhang et al., 5 Jun 2026).

The principal limitations are similarly recurrent. Chemical-kinetic MTK requires repeated ODE integrations and constrained optimization, and higher-dimensional manifold assembly introduces sensitivity to metric and slice selection (Dietrich et al., 2022). The variational SIM literature still lacks a generally applicable finite-horizon Lagrangian that yields exact slow manifolds without limit arguments (Lebiedz et al., 2014). Low-rank kinetic methods face open questions in robust high-order error analysis for stiff problems, positivity preservation, nonlinear collision operators, and distributed-memory tensor HPC (Einkemmer et al., 2024). The kinodynamic planning papers report empirical rather than formal guarantees; one uses rejection sampling to eliminate the remaining constraint violations, while the catching system remains subject to sim-to-real gap and manifold-coverage limitations (Lee, 2024, Pei et al., 27 May 2026). MIOFlow does not provide calibrated posterior uncertainty, and its performance depends on manifold metric quality and temporal sampling (Huguet et al., 2022). Fixed-rank PSD trajectory analysis scales quadratically in trajectory length for kernel construction and depends on reliable landmark extraction (Szczapa et al., 2022). MTK for jailbreak detection assumes that adversarial prompts induce anomalous neighborhood trajectories; the paper notes that a future stronger adversary might try to remain within benign neighborhoods at all depths while still eliciting unsafe outputs (Zhang et al., 5 Jun 2026).

Taken together, these works do not define a single canonical MTK theory. They do, however, define a coherent research pattern: represent dynamics by trajectories on manifolds whose geometry is chosen to encode invariance, feasibility, low complexity, or semantic neighborhood structure, and then perform optimization, simulation, inference, or detection in that manifold-constrained space. This suggests that MTK is best viewed not as one algorithm but as a geometric program for analyzing dynamical systems across chemical kinetics, kinetic simulation, robotics, observational trajectory analysis, and large-model safety.