Capability Degradation: Concepts & Models

Updated 9 February 2026

Capability Degradation is the measurable deterioration over time of a system’s ability to perform its intended function due to factors like internal wear, environmental influences, and adaptation processes.
Mathematical frameworks such as curvature-based segmentation, multi-mechanism coupled models, and physics-informed methods enable precise monitoring and forecasting of degradation across various systems.
Effective remediation strategies, including adapted training protocols and distributed system designs, mitigate degradation and enhance system safety, reliability, and performance.

Capability Degradation refers to the measurable, time-dependent deterioration or loss of a system’s ability to perform its intended functions, arising from internal mechanisms, environmental influences, or adaptation processes. The term is broadly applied across engineered systems, machine learning models, and biological or material systems, denoting both the macroscopic symptom (e.g., reduced capacity, increased error rates, loss of selectivity) and the underlying multiscale mechanisms. It is characterized by a shift in key performance metrics, often following complex, nonlinear, and sometimes regime-dependent trajectories, with significant implications for system safety, reliability, forecasting, and adaptive oversight.

1. Fundamental Definitions and Taxonomies

Capability Degradation encompasses a spectrum of phenomena across domains:

Physical systems: Progressive reduction in performance, such as energy storage or mechanical output, due to chemical/mechanical wear, corrosion, or fatigue.
Machine Learning and Model-Based Systems: Loss or drift in internal abilities (e.g., reasoning, generalization, language generation), especially when models are adapted to new modalities or domains, or as a result of poorly aligned fine-tuning (Kellogg et al., 5 Nov 2025, Wang et al., 28 Sep 2025, Yang et al., 2 Feb 2026).
Composite systems: Degradation that can arise from coupling between subsystems, regime transitions, or environmental stress thresholds (Chen et al., 13 Jan 2026, Li et al., 2023).

A general taxonomy, as formalized in LLM oversight (Kellogg et al., 5 Nov 2025), decomposes capability into distinct internal functions—e.g., summarization, translation, reasoning, safety guardrails—permitting capability-based monitoring. In engineered materials and devices, capability is tied to the preservation of key functional metrics (e.g., conductivity in AEMs, capacity/resistance in batteries).

2. Mathematical and Algorithmic Frameworks

Quantitative characterization of capability degradation involves rigorous mathematical models, tailored per domain and mechanism.

2.1. Battery Degradation

Curvature-based Regime Segmentation: In Li-ion batteries, capacity fade often exhibits distinct multi-regime kinetics. The work of Zhang et al. identifies “knee-onset” (curvature departure from baseline) and “knee” (onset of stable acceleration) in the capacity-vs-cycle trajectory, formalized via discrete curvature:

$y_{d,i} = y_{\text{sn},(i-\frac{w_s-1}{2})} + y_{\text{sn},(i+\frac{w_s-1}{2})} - 2y_{\text{sn},i}$

Regimes $s_1$ , $s_2$ , $s_3$ map to pre-knee, oscillatory, and accelerated degradation (Zhang et al., 2023).

Multi-Mechanism Coupled Models: The inclusion of multiple interacting degradation pathways is essential, as shown by Li et al., who model SEI growth, solvent dry-out, lithium plating, particle cracking, and mechanical loss (LAM), using differential equations for each mechanism and their couplings (Li et al., 2023).
Empirical/Physics-Informed Master Curves: Polymeric anion exchange membranes demonstrate universal degradation behavior, well-collapsed by a four-parameter sigmoidal form in log-conductivity:

$\log\sigma(t) = \log\sigma_\infty + \frac{\log\sigma_0 - \log\sigma_\infty}{1 + (t/t_0)^\alpha}$

Here, $t_0$ and $\alpha$ capture the time scale and sharpness of the regime transition; all physical parameters are linked to measurable quantities (Schertzer et al., 14 Oct 2025).

2.2. Machine Learning Models

Capability Drift in LLMs: Degradation is formalized as a drop in per-capability metrics $m_{c,t}$ over time or with respect to a baseline (Kellogg et al., 5 Nov 2025). Statistical process control techniques (z-scores, K-S tests) monitor drift across capabilities, enabling proactive remediation.
Parameter Importance Reallocation: In multimodal LLMs, textual capability degradation is mapped to a shift in the layer-wise parameter importance distribution $I_{\text{layer}}(k)$ , as determined by the gradient-based estimator:

$I_i(\theta) \approx \left| \frac{\partial L}{\partial \theta_i} \theta_i \right|$

Speech adaptation causes importance to migrate from deep layers toward shallow layers, reducing text-based reasoning performance (Wang et al., 28 Sep 2025).

Retrieval Paradigm Shift: When MLLMs are repurposed from generative to discriminative retrieval, their native compositional reasoning collapses. Capability degradation is quantified by the loss in Recall@k on a subset where the original model achieves perfect accuracy (Yang et al., 2 Feb 2026).

2.3. Environmental Regime Transitions

Piecewise Stochastic Models: For materials with mechanism transitions (e.g., temperature-driven in SAPF), degradation is modeled as a piecewise function of the stressors with a transition temperature $T_{\text{threshold}}$ :

$e(s) = \begin{cases} \exp(a_1\phi_1(T) + a_2\phi_2(\phi)) & T < T_{\text{threshold}} \ \exp(a_3\phi_3(T) + a_2\phi_2(\phi)) & T \ge T_{\text{threshold}} \end{cases}$

This structure allows accurate reliability and lifetime predictions across environmental profiles (Chen et al., 13 Jan 2026).

3. Mechanistic Origins and Regime Transitions

A core insight across domains is that capability degradation is rarely a monolithic, monotonic process. Physical systems often exhibit multi-regime degradation:

Knee Transitions: Lithium-ion batteries display a distinct shift from a slow to a rapid capacity fade at the “knee,” with a preceding oscillatory curvature regime marking the transition (Zhang et al., 2023).
Stress-Induced Mechanism Transitions: Ground storage of SAPF films leads to an abrupt regime change in degradation chemistry (Al₂O₃-only to Al₂O₃ + AlO(OH)) at $T_\mathrm{threshold}$ , reflected in both kinetics and reliability predictions (Chen et al., 13 Jan 2026).
Multimodal Model Adaptation: In LLM adaptation or fine-tuning, capability drift can arise from redistribution of internal resources (parameters) among modalities, or from the collapse of internal compositional reasoning after paradigm shifts (e.g., generative-to-discriminative retriever) (Wang et al., 28 Sep 2025, Yang et al., 2 Feb 2026).

A summary table of regime-driven degradation phenomena:

System Domain	Regime Transition Mechanism	Formal Definition or Indicator
Li-ion Batteries	Capacity knee/onset region (curvature method)	Minima in corrected arc-curve profile (Zhang et al., 2023)
SAPF (Aluminum films)	Temperature-thresholded corrosion pathway	Piecewise $e(s)$ at $T_\mathrm{threshold}$
AEM Conductivity	Chemically-driven sigmoidal transition	$\alpha$ , $t_0$ in four-parameter model (Schertzer et al., 14 Oct 2025)
ML Model Capabilities	Parameter/utility drift under adaptation/fine-tuning	Drift in $M_c$ , $I_{\text{layer}}$

4. Prediction, Detection, and Prognosis Frameworks

The detection and forecasting of capability degradation are accomplished by a combination of mechanistic, data-driven, and hybrid approaches:

Curvature-Based Early Warning: The curvature method enables online identification of incipient knees and rapid acceleration, providing robust, chemistry-agnostic early warnings and accurate predictions of end-of-life in batteries (Zhang et al., 2023).
Multi-Task Learning (MTL): For batteries, simultaneous forecasting of capacity and resistance degradation in a multi-task LSTM Seq2Seq network delivers superior accuracy and computational efficiency, robust to measurement noise and crossing aging regimes (Li et al., 2021).
Deep Koopman Operators: By encoding system trajectories into a latent space with approximately linear degradation, Koopman frameworks support transparent, trendable forecasting of remaining useful life (RUL) and the disentanglement of permanent degradation from control actions (Garmaev et al., 2023).
Physics-Enforced Neural Networks: In AEMs, physics-informed architectures enable high-accuracy, long-horizon degradation forecasts with minimal early-time experimental data, exploiting the universality of the four-parameter master curve (Schertzer et al., 14 Oct 2025).
LLM Capability Monitoring: Statistical process control combined with sentinel-task sampling allows for detection of per-capability degradation in generalist LLMs, enabling cross-task recovery and targeted remediation (Kellogg et al., 5 Nov 2025).

5. Remediation, Mitigation, and Robustness Strategies

Mitigating capability degradation—whether viewed as wearout, resource reallocation, or learning drift—requires domain-aligned interventions:

Adapted Training Protocols: Low-rank adaptation (LoRA) and layer-wise learning rate scheduling, by constraining parameter updates along pre-existing importance directions, have been shown to significantly reduce textual capability degradation in speech-enabled LLMs (Wang et al., 28 Sep 2025). Degradation-aware fine-tuning with additional trend penalties further improves zero-shot generalization and robustness in time-series foundation models for battery degradation (Chan et al., 13 May 2025).
Structured Data Mining and Recalibration: Model-agnostic frameworks such as ReCALL employ instance mining, chain-of-thought triplet generation, and grouped contrastive training to realign discriminative retrievers with the compositional reasoning of generative MLLMs, thereby recalibrating degraded capabilities (Yang et al., 2 Feb 2026).
Distributed System Design: In hardware-limited systems, such as LIS with hardware impairments, dividing the system into smaller, distributed units can significantly suppress the growth of degradation-induced noise and postpone utility collapse, as the impairment term scales down with the number of subsystems (Hu et al., 2018).
Physics-Linked Parameterization: Including all relevant degradation mechanisms in device models, and validating against independent degradation modes (e.g., LLI, LAM in batteries), removes non-uniqueness and supports robust extrapolation to new operating regimes (Li et al., 2023).

6. Evaluation, Validation, and Open Issues

Robust assessment of capability degradation and its models is contingent upon careful metric selection, comprehensive validation, and openness about uncertainties or domain-specific limitations.

Metric Suitability: Relying solely on macroscopic metrics (e.g., capacity, resistance) risks model non-uniqueness; validation must extend to independently measured degradation modes, as established in Li et al. (Li et al., 2023).
Cross-Condition Robustness: Models that generalize across protocols, chemistries, and stress profiles—such as the Battery-Timer foundation model with degradation-aware fine-tuning—demonstrate reduced error both in-domain and zero-shot (Chan et al., 13 May 2025).
Open Challenges: In LLMs, capability taxonomies are not yet exhaustive, automated drift metrics require ongoing human calibration, and vendor opacity remains an obstacle for workflow-wide robustness (Kellogg et al., 5 Nov 2025). For material systems, many published models assume fixed functional forms; extensions to multistage or regime-agnostic forms are warranted (Schertzer et al., 14 Oct 2025, Chen et al., 13 Jan 2026).

7. Broader Impact and Outlook

Capability degradation underpins the field of performance forecasting, safety assurance, and adaptive oversight in engineered systems and complex AI. Its rigorous characterization and monitoring are central to:

Asset Management and Warranty: Early prediction of onset times, knees, and EOL translates to efficient maintenance, second-life repurposing, and policy design for batteries and materials (Zhang et al., 2023, Li et al., 2023, Chen et al., 13 Jan 2026).
Model Safety and Adaptiveness: For LLMs and composite AI, capability-based monitoring offers a scalable route to oversight, surfacing emergent risks, and facilitating collaborative governance (Kellogg et al., 5 Nov 2025, Wang et al., 28 Sep 2025).
Cross-Domain Applicability: Piecewise regime modeling and physical-chemical-mechanistic coupling, as in SAPF or AEMs, generalize to materials experiencing environmental or operational regime switches.
Foundational Model Generalization: Integration of knowledge distillation and fine-tuning strategies that encode degradation-awareness into compact, efficient models supports real-time deployment across diverse operational environments (Chan et al., 13 May 2025).

Continued advancement in universal, regime-aware models, systematic capability-based monitoring, and physics-informed learning architectures is likely to further diminish latency between degradation onset and detection, enhance interpretability, and deepen cross-domain resilience and safety.