Score Stability in Dynamic Systems

• Score Stability is a family of concepts assessing the robustness of score-driven updates and decisions under perturbations such as noise, misspecification, or weight changes.
• It leverages techniques like contraction mappings, Lipschitz continuity, and error bounds to ensure stable behavior in deep networks, filtering, and generative models.
• Applications span deep learning, econometrics, ranking, forecasting, and editing, providing actionable insights for designing robust and reliable systems.

Score stability is not a single universally standardized construct; it is a family of domain-specific notions concerned with whether a score, score-driven update, or score-induced decision remains well behaved under perturbation, iteration, misspecification, discretization, or weight changes. In recent literature, the term appears in at least six technically distinct senses: stability of score-based recurrences in deep networks, stability of score-driven filters under misspecification, stability of rankings induced by weighted scoring functions, stability of diffusion-based score distillation for image and 3D editing, stability of learned score fields in score-based generative models and filtering, and stability of score- or threshold-based decisions in psychometrics and credit risk. Across these settings, a common motif is that stability is expressed through contraction, non-expansiveness, bounded sensitivity, overlap or distance under perturbation, or explicit error bounds that propagate local score errors to global discrepancies (Godin, 11 Mar 2026, Heel et al., 7 Feb 2025, Chen et al., 1 Jul 2026, Zhu et al., 12 Jul 2025, Song et al., 2020, Ma et al., 15 Jan 2026).

1. Domain-specific meanings of score stability

In deep learning, score stability can refer to the behavior of recurrent depth updates built from a shared block. In "SCORE: Replacing Layer Stacking with Contractive Recurrent Depth" (Godin, 11 Mar 2026), SCORE uses

$h_{t+1} = h_t + \Delta t \, (F(h_t) - h_t)$

or equivalently

$$h_{t+1} = (1 - \Delta t)\, h_t + \Delta t\, F(h_t),$$

and stability is tied to the role of $\Delta t$ as a “stability knob,” the contractivity of the map $G(x) = (1-\Delta t)x + \Delta t F(x)$, and the resulting gradient behavior across shared recurrent depth.

In econometrics, score stability is the stability of score-driven filters when the postulated observation model may differ from the true data-generating process. "Stability and performance guarantees for misspecified multivariate score-driven filters" (Heel et al., 7 Feb 2025) studies implicit score-driven (ISD) and explicit score-driven (ESD) filters, where stability means invertibility of the filtered parameter path and existence of finite MSE bounds relative to the pseudo-true parameter path.

In ranking, the term concerns robustness of a reported ranking to changes in scoring weights. "The General Stability of Ranking" (Chen et al., 1 Jul 2026) defines exact stability as

$S_{\mathrm{exact}}(\pi^*) = \mu(R_{\pi^*})/\mu(W)$

and general stability as

$$S_{\mathrm{gen}}(\pi^*, d, h) = \frac{1}{\mu(W)} \int_{w \in W} h(d(\pi(w),\pi^*))\,dw,$$

with the paper’s primary instantiation using $h(t)=e^{-t}$. Here stability is geometric and volume-based, but generalized by ranking distance.

In diffusion-based editing, score stability is an optimization property. "Stable Score Distillation" (Zhu et al., 12 Jul 2025) describes instability in text-guided editing as drift, conflicting guidance, weak spatial control, and variance in the distilled update. Its solution is an anchored single-classifier formulation with a constant null-text branch intended to dampen drift and preserve structure.

In score-based generative modeling, the term refers to the regularity of the score field and the robustness of reverse-time sampling. "Improved Techniques for Training Score-Based Generative Models" (Song et al., 2020), "Regularity of the score function in generative models" (Stéphanovitch, 24 Jun 2025), "Generalization bounds for score-based generative models: a synthetic proof" (Stéphanovitch et al., 7 Jul 2025), and "On Forgetting and Stability of Score-based Generative models" (Strasman et al., 29 Jan 2026) all connect stability to Lipschitz or one-sided Lipschitz control, dissipativity, forgetting, and error propagation through reverse-time dynamics.

Other uses are operational rather than dynamical. In "Beyond Accuracy: A Stability-Aware Metric for Multi-Horizon Forecasting" (Ma et al., 15 Jan 2026), stability is called coherence and measures whether forecasts for the same target remain consistent as the forecast origin changes. In "NeSS-ST: Detecting Good and Stable Keypoints with a Neural Stability Score and the Shi-Tomasi Detector" (Pakulev et al., 2023), stability is the maximum positional variation of a keypoint under local perspective perturbations. In "Reliability of decisions based on tests" (Waldorp et al., 2020), stability is decision reliability under coordinate noise. In "A critical review of existing and new population stability testing procedures in credit risk scoring" (Pisanie et al., 2023), population stability means that the distributions of applicant attributes at review remain representative of those at development.

A plausible implication is that “score stability” functions as an umbrella term whose formal content is set by the object being propagated: embeddings, latent parameters, rankings, forecast distributions, keypoints, or reverse-time probability laws.

2. Contractive updates, score fields, and dynamical stability

The SCORE architecture makes stability explicit at the level of the update map. If $F$ is $L$-Lipschitz, then for

$G(x) = (1 - \Delta t)x + \Delta t F(x),$

the paper gives

$$h_{t+1} = (1 - \Delta t)\, h_t + \Delta t\, F(h_t),$$0

Hence, if $$h_{t+1} = (1 - \Delta t)\, h_t + \Delta t\, F(h_t),$$1 and $$h_{t+1} = (1 - \Delta t)\, h_t + \Delta t\, F(h_t),$$2, $$h_{t+1} = (1 - \Delta t)\, h_t + \Delta t\, F(h_t),$$3 is non-expansive; if $$h_{t+1} = (1 - \Delta t)\, h_t + \Delta t\, F(h_t),$$4 and $$h_{t+1} = (1 - \Delta t)\, h_t + \Delta t\, F(h_t),$$5, $$h_{t+1} = (1 - \Delta t)\, h_t + \Delta t\, F(h_t),$$6 is a contraction, and the fixed-point error recursion

$$h_{t+1} = (1 - \Delta t)\, h_t + \Delta t\, F(h_t),$$7

implies geometric convergence to a fixed point (Godin, 11 Mar 2026). The same paper states that $$h_{t+1} = (1 - \Delta t)\, h_t + \Delta t\, F(h_t),$$8 and $$h_{t+1} = (1 - \Delta t)\, h_t + \Delta t\, F(h_t),$$9 both exhibit stable behavior across architectures, with $\Delta t$0 used by default.

This contractive viewpoint has direct analogues in score-based generative models. In "Improved Techniques for Training Score-Based Generative Models" (Song et al., 2020), stability is approached through the geometry of annealed Langevin dynamics, the scaling of the learned score with noise level, and the use of exponential moving average (EMA) of model weights. The paper proposes parameterizing

$\Delta t$1

choosing $\Delta t$2 comparable to the largest pairwise training distance, using a geometric noise schedule, and tuning Langevin step sizes so that $\Delta t$3. The paper explicitly describes EMA with decay $\Delta t$4 as a stabilizer that suppresses training noise and drift in the learned score field.

Theoretical work makes these heuristics more structural. "Regularity of the score function in generative models" (Stéphanovitch, 24 Jun 2025) defines the score by

$\Delta t$5

and proves one-sided Lipschitz regularity:

$\Delta t$6

It also proves

$\Delta t$7

which directly supports stability analyses for reverse-time SDEs and probability flow ODEs (Stéphanovitch, 24 Jun 2025). "Generalization bounds for score-based generative models: a synthetic proof" (Stéphanovitch et al., 7 Jul 2025) uses a related one-sided Lipschitz property,

$\Delta t$8

to ensure the exponential factor in a diffusion stability bound remains finite.

"On Forgetting and Stability of Score-based Generative models" (Strasman et al., 29 Jan 2026) reframes this as Harris-type contraction for the reverse-time Markov kernels. It states that under a Lyapunov drift condition and a Doeblin-type minorization condition, the reverse chain forgets its initialization geometrically fast and propagates one-step discretization and score-approximation errors with geometric discounting. Its global bound has the form

$\Delta t$9

making stability an explicit decomposition into initialization, discretization, and score-approximation terms (Strasman et al., 29 Jan 2026).

A related but distinct use appears in physics inversion. "Solving Inverse Physics Problems with Score Matching" (Holzschuh et al., 2023) shows that a single-step training loss for a learned correction to an approximate inverse simulator is equivalent, as $G(x) = (1-\Delta t)x + \Delta t F(x)$0, to score matching for a scaled score. The paper argues that recursively predicting longer parts of the trajectory during training relates to maximum likelihood training of the corresponding probability flow, and its empirical account of “excellent accuracy and temporal stability” is tied to this multi-step alignment between training and inference.

3. Score-driven filters and forecast coherence

In score-driven time-series filtering, stability is a property of the filtering recursion itself. "Stability and performance guarantees for misspecified multivariate score-driven filters" (Heel et al., 7 Feb 2025) defines the ISD update

$G(x) = (1-\Delta t)x + \Delta t F(x)$1

and the ESD update

$G(x) = (1-\Delta t)x + \Delta t F(x)$2

with prediction

$G(x) = (1-\Delta t)x + \Delta t F(x)$3

Invertibility is defined by exponential forgetting of initialization:

$G(x) = (1-\Delta t)x + \Delta t F(x)$4

The paper’s central distinction is that ISD stability depends on concavity and proximal curvature, whereas ESD additionally requires Lipschitz continuity of the score and a step-size restriction $G(x) = (1-\Delta t)x + \Delta t F(x)$5 (Heel et al., 7 Feb 2025). It further gives explicit finite-sample and asymptotic MSE bounds relative to the pseudo-true parameter path under mild moment conditions.

In forecasting, stability appears as cross-origin consistency rather than contraction of a hidden state. "Beyond Accuracy: A Stability-Aware Metric for Multi-Horizon Forecasting" (Ma et al., 15 Jan 2026) defines a probabilistic multi-horizon accuracy term via the Energy Score and a stability term via the squared weighted Energy Distance between successive-overlap marginals. With

$G(x) = (1-\Delta t)x + \Delta t F(x)$6

the full score is

$G(x) = (1-\Delta t)x + \Delta t F(x)$7

where $G(x) = (1-\Delta t)x + \Delta t F(x)$8 averages

$G(x) = (1-\Delta t)x + \Delta t F(x)$9

across origins (Ma et al., 15 Jan 2026). For point forecasts, the stability term reduces to the weighted Euclidean norm of overlapping forecast differences.

The paper reports that AC-optimized SARIMA models on the M4 Hourly benchmark achieve a $S_{\mathrm{exact}}(\pi^*) = \mu(R_{\pi^*})/\mu(W)$0 reduction in vertical forecast volatility, with average standard deviation decreasing from $S_{\mathrm{exact}}(\pi^*) = \mu(R_{\pi^*})/\mu(W)$1 to $S_{\mathrm{exact}}(\pi^*) = \mu(R_{\pi^*})/\mu(W)$2, a median $S_{\mathrm{exact}}(\pi^*) = \mu(R_{\pi^*})/\mu(W)$3 improvement in the stability score, and a median $S_{\mathrm{exact}}(\pi^*) = \mu(R_{\pi^*})/\mu(W)$4 improvement in the weighted accuracy score (Ma et al., 15 Jan 2026). Here score stability is explicitly a trade-off parameterized by $S_{\mathrm{exact}}(\pi^*) = \mu(R_{\pi^*})/\mu(W)$5, with $S_{\mathrm{exact}}(\pi^*) = \mu(R_{\pi^*})/\mu(W)$6 giving accuracy-only optimization and $S_{\mathrm{exact}}(\pi^*) = \mu(R_{\pi^*})/\mu(W)$7 favoring minimal forecast revision.

A plausible implication is that score-driven filters and forecast-coherence metrics formalize two complementary temporal notions of stability: the first is stability of a latent-state update map under misspecification, and the second is stability of externally visible predictive revisions.

4. Rankings, decisions, and population drift

In ranking problems, stability is geometric sensitivity to the scoring weights. "The General Stability of Ranking" (Chen et al., 1 Jul 2026) considers items with feature vectors $S_{\mathrm{exact}}(\pi^*) = \mu(R_{\pi^*})/\mu(W)$8, linear scores $S_{\mathrm{exact}}(\pi^*) = \mu(R_{\pi^*})/\mu(W)$9, and a weight domain

$$S_{\mathrm{gen}}(\pi^*, d, h) = \frac{1}{\mu(W)} \int_{w \in W} h(d(\pi(w),\pi^*))\,dw,$$0

For a target ranking $$S_{\mathrm{gen}}(\pi^*, d, h) = \frac{1}{\mu(W)} \int_{w \in W} h(d(\pi(w),\pi^*))\,dw,$$1, exact stability is

$$S_{\mathrm{gen}}(\pi^*, d, h) = \frac{1}{\mu(W)} \int_{w \in W} h(d(\pi(w),\pi^*))\,dw,$$2

while general stability is

$$S_{\mathrm{gen}}(\pi^*, d, h) = \frac{1}{\mu(W)} \int_{w \in W} h(d(\pi(w),\pi^*))\,dw,$$3

The paper motivates this generalization by arguing that exact stability gives no credit to rankings that differ only by minor lower-order swaps. It provides a two-dimensional sweep algorithm, an unbiased multidimensional sampler, and Conv-SC for quasiconvex distance functions. On eight real datasets, it reports that exact versus $$S_{\mathrm{gen}}(\pi^*, d, h) = \frac{1}{\mu(W)} \int_{w \in W} h(d(\pi(w),\pi^*))\,dw,$$4 stability has correlation $$S_{\mathrm{gen}}(\pi^*, d, h) = \frac{1}{\mu(W)} \int_{w \in W} h(d(\pi(w),\pi^*))\,dw,$$5, while $$S_{\mathrm{gen}}(\pi^*, d, h) = \frac{1}{\mu(W)} \int_{w \in W} h(d(\pi(w),\pi^*))\,dw,$$6 and $$S_{\mathrm{gen}}(\pi^*, d, h) = \frac{1}{\mu(W)} \int_{w \in W} h(d(\pi(w),\pi^*))\,dw,$$7 are nearly collinear at approximately $$S_{\mathrm{gen}}(\pi^*, d, h) = \frac{1}{\mu(W)} \int_{w \in W} h(d(\pi(w),\pi^*))\,dw,$$8, and it gives the CWUR patents top-30 example where exact stability is approximately $$S_{\mathrm{gen}}(\pi^*, d, h) = \frac{1}{\mu(W)} \int_{w \in W} h(d(\pi(w),\pi^*))\,dw,$$9 but top-sensitive Kendall stability is $h(t)=e^{-t}$0 (Chen et al., 1 Jul 2026).

In psychometrics, decision stability is robustness of a Boolean decision rule to perturbations of item outcomes. "Reliability of decisions based on tests" (Waldorp et al., 2020) studies Boolean functions $h(t)=e^{-t}$1, including linear threshold functions

$h(t)=e^{-t}$2

and defines noise stability

$h(t)=e^{-t}$3

where $h(t)=e^{-t}$4 is obtained by flipping each coordinate of $h(t)=e^{-t}$5 with probability $h(t)=e^{-t}$6. Decision-flip probability is

$h(t)=e^{-t}$7

For the majority function, the paper cites the asymptotic relation

$h(t)=e^{-t}$8

and argues that weighted sum scores have desirable properties connected to monotonicity, stability, and Rousseau’s criterion (Waldorp et al., 2020). Stability is thus sensitivity of a threshold decision under measurement error.

In credit risk scoring, population stability is a review-time comparison between development and current distributions. "A critical review of existing and new population stability testing procedures in credit risk scoring" (Pisanie et al., 2023) states the null hypothesis as $h(t)=e^{-t}$9 for all bins $F$0, and reviews the Population Stability Index

$F$1

with the widely used heuristic thresholds $F$2, $F$3, and $F$4, while criticizing these as insensitive to sample size and the number of bins. The same paper proposes the effect-size statistic

$F$5

and the overlapping statistic

$F$6

It argues that $F$7 indicates practically meaningful distributional change and emphasizes the intuitive interpretation of overlap (Pisanie et al., 2023).

These three literatures share a structural idea: stability is a property of the mapping from scores to decisions under admissible perturbations. The perturbation may be a weight change, a flip of item outcomes, or a shift in the input distribution.

5. Vision, editing, and localization stability

In text-guided image and 3D editing, score stability refers to the stability of the distilled optimization direction. "Stable Score Distillation" (Zhu et al., 12 Jul 2025) begins from the score-distillation setting in which a differentiable renderer $F$8 produces an image or latent $F$9, and the update uses denoising predictions from a pretrained diffusion model. Its core anchored cross-prompt direction is

$L$0

and its central stabilization is the constant-term null-text branch

$L$1

The paper also decomposes the update into cross-prompt and cross-trajectory terms, adds a prompt enhancement branch

$L$2

and optionally uses source latent regularization for 3D editing (Zhu et al., 12 Jul 2025). Its stated intuition is that the cross-trajectory term reduces drift and oscillation while preserving source structure. Empirically, it reports CLIP Similarity $L$3 and CLIP Directional Similarity $L$4 in 3D editing, with user-study preference $L$5, and on PIE-Bench reports background preservation metrics including LPIPS $L$6, MSE $L$7, and Structure Distance $L$8 for SSD+CDS (Zhu et al., 12 Jul 2025).

In local feature detection, stability is positional repeatability under perturbation rather than optimization coherence. "NeSS-ST" (Pakulev et al., 2023) defines the Stability Score (SS) by perturbing local neighborhoods with random homographies, redetecting the Shi-Tomasi maximum in each perturbed patch, computing the sample covariance $L$9 of displacement vectors, and setting

$G(x) = (1 - \Delta t)x + \Delta t F(x),$0

Smaller $G(x) = (1 - \Delta t)x + \Delta t F(x),$1 means higher stability. The Neural Stability Score (NeSS) is trained to regress this quantity from images, using the normalized $G(x) = (1 - \Delta t)x + \Delta t F(x),$2 loss

$G(x) = (1 - \Delta t)x + \Delta t F(x),$3

The pipeline uses $G(x) = (1 - \Delta t)x + \Delta t F(x),$4 homographies, $G(x) = (1 - \Delta t)x + \Delta t F(x),$5 with $G(x) = (1 - \Delta t)x + \Delta t F(x),$6 support for Shi responses, and a U-Net with 4 down-sampling layers (Pakulev et al., 2023). Reported downstream performance includes HPatches homography mAA at $G(x) = (1 - \Delta t)x + \Delta t F(x),$7 px of $G(x) = (1 - \Delta t)x + \Delta t F(x),$8 overall and $G(x) = (1 - \Delta t)x + \Delta t F(x),$9 on viewpoint, IMC-PT relative pose mAA of rotation $$h_{t+1} = (1 - \Delta t)\, h_t + \Delta t\, F(h_t),$$00 and translation $$h_{t+1} = (1 - \Delta t)\, h_t + \Delta t\, F(h_t),$$01, MegaDepth rotation $$h_{t+1} = (1 - \Delta t)\, h_t + \Delta t\, F(h_t),$$02 and translation $$h_{t+1} = (1 - \Delta t)\, h_t + \Delta t\, F(h_t),$$03, and ScanNet rotation $$h_{t+1} = (1 - \Delta t)\, h_t + \Delta t\, F(h_t),$$04 and translation $$h_{t+1} = (1 - \Delta t)\, h_t + \Delta t\, F(h_t),$$05 (Pakulev et al., 2023).

These papers show two sharply different but formally related uses of the term. In editing, score stability is variance control of a learned gradient direction in a high-dimensional latent optimization problem. In keypoint detection, it is maximal directional variance of a local estimator under controlled geometric perturbations.

6. Theoretical error control, regularity, and recurring principles

Several papers make score stability explicit through quantitative bounds that relate local score errors to global discrepancies. In score-based nonlinear filtering, "A Score-based Nonlinear Filter for Data Assimilation" (Bao et al., 2023) models the filtering density through a learned conditional score $$h_{t+1} = (1 - \Delta t)\, h_t + \Delta t\, F(h_t),$$06 and updates the score analytically by

$$h_{t+1} = (1 - \Delta t)\, h_t + \Delta t\, F(h_t),$$07

with $$h_{t+1} = (1 - \Delta t)\, h_t + \Delta t\, F(h_t),$$08 in experiments. The paper argues that storing the filtering density in the score model, rather than in finite particle weights, improves robustness to degeneracy and dimensionality, and it reports stable performance on examples up to 100-dimensional Lorenz-96 (Bao et al., 2023).

In flow-based reinforcement learning, "ScoRe-Flow" (Qiu et al., 13 Apr 2026) defines the intermediate-marginal score

$$h_{t+1} = (1 - \Delta t)\, h_t + \Delta t\, F(h_t),$$09

derives the closed-form approximation

$$h_{t+1} = (1 - \Delta t)\, h_t + \Delta t\, F(h_t),$$10

and uses the score-modulated SDE

$$h_{t+1} = (1 - \Delta t)\, h_t + \Delta t\, F(h_t),$$11

Because the score magnitude diverges like $$h_{t+1} = (1 - \Delta t)\, h_t + \Delta t\, F(h_t),$$12, the paper enforces

$$h_{t+1} = (1 - \Delta t)\, h_t + \Delta t\, F(h_t),$$13

to keep the score-modulated drift bounded (Qiu et al., 13 Apr 2026). It reports $$h_{t+1} = (1 - \Delta t)\, h_t + \Delta t\, F(h_t),$$14 faster convergence than flow-based SOTA on D4RL locomotion tasks and up to $$h_{t+1} = (1 - \Delta t)\, h_t + \Delta t\, F(h_t),$$15 higher success rates on Robomimic and Franka Kitchen manipulation tasks.

In the Landau equation, stability is measured in relative entropy. "Stability of the spatially homogeneous Landau equation in relative entropy and applications to score-based numerical methods" (Ilin, 16 Oct 2025) proves

$$h_{t+1} = (1 - \Delta t)\, h_t + \Delta t\, F(h_t),$$16

under stated moment and regularity assumptions, giving

$$h_{t+1} = (1 - \Delta t)\, h_t + \Delta t\, F(h_t),$$17

and derives the a posteriori score-based bound

$$h_{t+1} = (1 - \Delta t)\, h_t + \Delta t\, F(h_t),$$18

where

$$h_{t+1} = (1 - \Delta t)\, h_t + \Delta t\, F(h_t),$$19

This makes score stability an a posteriori guarantee: small weighted score-matching loss over time implies small KL error to the true Landau solution (Ilin, 16 Oct 2025).

Across the literature, the recurring principles are remarkably consistent. Stability is improved by contraction or one-sided Lipschitz control, by damping or trust-region structure, by weight sharing or proximal regularization, by geometric discounting of early errors, and by replacing brittle exact-match criteria with graded distance-sensitive notions. Instability is repeatedly associated with overly aggressive step sizes, non-Lipschitz curvature, weak overlap between successive states or weight regions, non-anchored auxiliary branches, and rare-event geometry in high dimensions (Godin, 11 Mar 2026, Heel et al., 7 Feb 2025, Chen et al., 1 Jul 2026, Zhu et al., 12 Jul 2025, Stéphanovitch, 24 Jun 2025, Strasman et al., 29 Jan 2026).

A plausible implication is that the modern study of score stability has converged on a common mathematical vocabulary—Lipschitz bounds, contraction coefficients, Energy or Wasserstein distances, entropy inequalities, and overlap measures—even when the underlying objects differ radically. In that sense, the term now denotes not merely robustness of a scalar score, but robustness of score-governed dynamical systems, estimators, and decisions under the perturbations that matter in their respective applications.