CURE: Multidisciplinary Models and Methods

• CURE is a multifaceted concept that describes statistical models in survival analysis for long-term survivors, distinguishing between cure probability and latency periods.
• In medical AI and vision, CURE represents frameworks for counterfactual survival prediction, reliable question answering, and concept unlearning with quantifiable performance gains.
• Beyond health sciences, CURE is applied in composite manufacturing for optimizing thermoset crosslinking and in education as a course-based undergraduate research experience.

Searching arXiv for papers on “CURE” and “cure models” to ground the article in current literature. In the cited literature, “CURE” denotes several distinct research objects rather than a single unified concept. In survival analysis, it refers to models for populations containing a non-negligible fraction of individuals who will never experience the event of interest; in medical AI, vision, and machine learning, it appears as an acronym for multiple method families; in engineering, “cure” refers to thermoset crosslinking during composite manufacture; and in undergraduate STEM education, CURE denotes a Course-Based Undergraduate Research Experience (Delhelle et al., 2024, López-Cheda et al., 2024, Hariharan et al., 2024, Nguyen et al., 23 Feb 2026, Shivam et al., 23 Sep 2025, Elshaer et al., 16 Oct 2025, Messina et al., 21 Jan 2026, Kim et al., 3 Jul 2026, Biswas et al., 19 May 2025, Markert et al., 30 May 2026, Limaye et al., 30 May 2025).

1. Cure in survival analysis

In survival analysis, a cure model is used when a positive fraction of subjects will never experience the event of interest, even under indefinite follow-up. With event time $T$, censoring time $C$, observed time $Y=\min(T,C)$, and event indicator $\Delta=\mathbf{1}\{T\le C\}$, the defining feature is a defective event-time distribution with $\Pr(T=\infty)=p>0$. Equivalently, the overall conditional survival may be written in mixture form as

$S(t\mid x)=1-p(x)+p(x)S_0(t\mid x),$

where $1-p(x)$ is the cure probability and $S_0$ is the latency survival among uncured individuals (Delhelle et al., 2024, López-Cheda et al., 2024).

This framework is central in oncology and other long-horizon event settings because standard survival models implicitly assume that everyone would eventually fail if followed long enough. The cited work develops cure models for right-censored, current-status, left- and right-censored, and recurrent-event settings. In current-status data, only a single inspection time $U$ and status $\delta=\mathbf{1}(T\le U)$ are observed. For that case, a promotion time cure model specifies

$C$0

which gives cure a direct latent-cause interpretation through a Poisson number of carcinogenic foci (Hariharan et al., 2024).

The survival-analysis literature in the data block also distinguishes cure incidence from latency. The Cox proportional hazards cure model combines logistic regression for the probability of being uncured with Cox PH regression for the survival of uncured subjects. In that formulation,

$C$1

and

$C$2

so the population survival is

$C$3

This decomposition makes cure probability and failure-time dynamics separately estimable under appropriate assumptions (Mohammad et al., 2019).

2. Statistical developments in cure modeling

A major recent extension concerns dependent censoring. “Copula based dependent censoring in cure models” formulates a fully parametric joint model for $C$4 using parametric margins for $C$5 and $C$6 and a copula $C$7, so that

$C$8

The paper develops identifiability conditions for $C$9, proposes maximum-likelihood estimation, and shows that misspecifying independence when $Y=\min(T,C)$0 and $Y=\min(T,C)$1 are dependent can seriously bias survival and cure-fraction estimates. In the breast cancer application, Joe-copula models with Gamma margins yielded estimated Kendall’s $Y=\min(T,C)$2 around $Y=\min(T,C)$3–$Y=\min(T,C)$4 and cure fractions around $Y=\min(T,C)$5–$Y=\min(T,C)$6, whereas independence models produced different cure estimates and inferior AIC (Delhelle et al., 2024).

A second development is efficient semiparametric inference for Cox PH cure models. “Efficient Estimation For The Cox Proportional Hazards Cure Model” proves asymptotic normality of the profile likelihood estimator, shows that the efficient score obtained by projection theory equals the profile likelihood score, and expresses the efficient information matrix as the variance of that score. The paper also reports that standard errors based on the profile likelihood score function are similar to bootstrap standard errors from the SMCURE package, and illustrates the method on melanoma ECOG e1684 data (Mohammad et al., 2019).

A third line of work removes the need to model the susceptible latency distribution. “A likelihood-based approach for cure regression models” introduces an inverse-probability-of-censoring-weighted surrogate cure indicator,

$Y=\min(T,C)$7

and replaces the unobserved cure indicator in a Bernoulli-type likelihood. This yields a likelihood-based estimator for cure regression under random right-censoring without making assumptions on the distribution of survival times for susceptible subjects. The method extends naturally to $Y=\min(T,C)$8-type penalties, and the adaptive lasso is shown to satisfy an oracle property under the stated conditions (Burke et al., 2018).

Other papers broaden the observable structure of cure. “A Generalized Mixture Cure Model Incorporating Known Cured Individuals” allows some cured individuals to be observed directly and models the time to cure identification through $Y=\min(T,C)$9. It shows that explicitly incorporating known cured information can increase precision and decrease mean squared error, especially when cure identification time is stochastic; by contrast, traditional models that ignore known cured information perform well when curation occurs after a known cutoff point (Karakatsoulis, 2024). “A Multivariate Cure Model for Left- and Right-Censored Data with Application to Colorectal Cancer Screening Patterns” defines cure through the lifetime number of screenings $\Delta=\mathbf{1}\{T\le C\}$0, with $\Delta=\mathbf{1}\{T\le C\}$1 representing the never-screened population, and combines left- and right-censoring with within-subject dependence via a positive stable frailty model (Hagar et al., 2015). “npcure: An R Package for Nonparametric Inference in Mixture Cure Models” provides completely nonparametric estimators for cure probability and latency as functions of a covariate, plus bootstrap bandwidth selectors and a nonparametric covariate-significance test (López-Cheda et al., 2024).

These developments suggest that contemporary cure modeling is increasingly concerned with identifiability under weak observability, robustness to censoring assumptions, and separation of incidence from latency.

3. CURE in multimodal and medical AI

In medical AI, “CURE” is also used as an acronym for several technically unrelated frameworks. “Counterfactual Understanding via Retrieval-aware Multimodal Modeling for Time-to-Event Survival Prediction” defines CURE as a two-phase framework for time-to-event counterfactual survival prediction under right-censoring. It fuses clinical, paraclinical, demographic, and multi-omics information through bottlenecks, mixture-of-experts modules, and cross-attention, then predicts treatment-specific survival by mixing over latent treatment-response and baseline-survival subgroups: $\Delta=\mathbf{1}\{T\le C\}$2 On METABRIC, it reports $\Delta=\mathbf{1}\{T\le C\}$3 and IBS $\Delta=\mathbf{1}\{T\le C\}$4; on TCGA-LUAD, $\Delta=\mathbf{1}\{T\le C\}$5 and IBS $\Delta=\mathbf{1}\{T\le C\}$6, outperforming the listed baselines (Nguyen et al., 23 Feb 2026).

“CURE: Confidence-driven Unified Reasoning Ensemble Framework for Medical Question Answering” addresses medical multiple-choice QA without fine-tuning. Its primary model, Qwen3-30B-A3B-Instruct, first answers a confidence prompt with “Sure” or “Not Sure.” Low-confidence questions are routed to Phi-4 14B and Gemma 2 12B, and Qwen3 then performs collaborative chain-of-thought reasoning over their answers. The reported accuracies are $\Delta=\mathbf{1}\{T\le C\}$7 on MedQA, $\Delta=\mathbf{1}\{T\le C\}$8 on MedMCQA, and $\Delta=\mathbf{1}\{T\le C\}$9 on PubMedQA, with an average of $\Pr(T=\infty)=p>0$0 (Elshaer et al., 16 Oct 2025).

“CURE: Curriculum-guided Multi-task Training for Reliable Anatomy Grounded Report Generation” fine-tunes MedGemma-4B-IT on phrase grounding, grounded report generation, and anatomy-grounded report generation, using an error-aware curriculum that reweights datasets and categories according to IoU and CXRFEScore. The method improves grounding accuracy by $\Pr(T=\infty)=p>0$1 IoU, boosts report quality by $\Pr(T=\infty)=p>0$2 CXRFEScore, and reduces hallucinations by $\Pr(T=\infty)=p>0$3 (Messina et al., 21 Jan 2026).

A plausible implication is that, in the medical-AI papers using the CURE acronym, reliability is pursued through explicit structure: latent subgroup retrieval in survival prediction, confidence-aware routing in QA, and anatomy-level grounding in report generation.

4. CURE in vision, unlearning, and generative modeling

Outside medicine, the acronym names several methods in representation editing, restoration, and graph-manifold regularization. “CURE: Centroid-guided Unsupervised Representation Erasure for Facial Recognition Systems” is an unsupervised machine-unlearning framework for face embeddings. It clusters teacher embeddings with K-means, assigns forget samples to farthest centroids, and optimizes a student model with pseudo-label, cosine, contrastive, feature-matching, and distribution losses. The paper also introduces the Unlearning Efficiency Score,

$\Pr(T=\infty)=p>0$4

with $\Pr(T=\infty)=p>0$5 in experiments. On CASIA-WebFace, CURE reports forget accuracy $\Pr(T=\infty)=p>0$6, retain accuracy $\Pr(T=\infty)=p>0$7, and the highest UES, $\Pr(T=\infty)=p>0$8, among the compared methods (Shivam et al., 23 Sep 2025).

“CURE: Concept Unlearning via Orthogonal Representation Editing in Diffusion Models” is a training-free concept-unlearning method for text-to-image diffusion models. Its Spectral Eraser computes SVDs of forget and retain token-embedding matrices, constructs spectrally expanded projectors, and edits cross-attention key and value weights by

$\Pr(T=\infty)=p>0$9

All operations are closed form, and the paper reports concept erasure in about $S(t\mid x)=1-p(x)+p(x)S_0(t\mid x),$0 seconds while modifying $S(t\mid x)=1-p(x)+p(x)S_0(t\mid x),$1 of parameters (Biswas et al., 19 May 2025).

“CURE: Controllable Unified Image Restoration for Complex Degradations” adds an identity embedding, ratio control, intermediate loss, and permutation-invariant loss to existing text-guided restoration backbones. Restoration intensity is controlled by

$S(t\mid x)=1-p(x)+p(x)S_0(t\mid x),$2

with $S(t\mid x)=1-p(x)+p(x)S_0(t\mid x),$3. On CCDD-11, OneRestore (Text) improves from $S(t\mid x)=1-p(x)+p(x)S_0(t\mid x),$4 to $S(t\mid x)=1-p(x)+p(x)S_0(t\mid x),$5 under CURE training, and identity-prompt fidelity rises from roughly $S(t\mid x)=1-p(x)+p(x)S_0(t\mid x),$6–$S(t\mid x)=1-p(x)+p(x)S_0(t\mid x),$7 dB to about $S(t\mid x)=1-p(x)+p(x)S_0(t\mid x),$8 dB (Kim et al., 3 Jul 2026).

An earlier usage, “CURE: Curvature Regularization For Missing Data Recovery,” augments the low-dimensional manifold model with a biharmonic curvature term,

$S(t\mid x)=1-p(x)+p(x)S_0(t\mid x),$9

and introduces WeCURE as a weighted variant. The reported numerical experiments show that CURE and WeCURE significantly outperform LDMM and WNLL, respectively, in image inpainting and semi-supervised learning (Dong et al., 2019).

5. CURE in education and composite processing

In undergraduate STEM education, CURE denotes a Course-Based Undergraduate Research Experience. The cited physics-education study distinguishes five elements commonly associated with CUREs: use of scientific practices, discovery, broad relevance, collaboration, and iteration. It then compares two experimentation-based introductory physics labs, one “CURE-like” (PEPPER, using cosmic-ray muon detectors and explicit particle-physics framing) and one standard experimentation-based lab (SALT). Using hierarchical linear modeling, the paper finds no statistically significant difference in post-survey scores between PEPPER and SALT on experimental critical thinking skills, self-efficacy, belonging, perceived agency, or recognition under multiple imputation. The authors conclude that increased levels of broad relevance may not inherently improve gains in student learning or attitudes (Markert et al., 30 May 2026).

In composite manufacturing, “cure” refers to the thermally driven crosslinking of thermoset prepregs. “Numerical Simulation Informed Rapid Cure Process Optimization of Composite Structures using Constrained Bayesian Optimization” models degree of cure $1-p(x)$0, exothermic heat generation, cure shrinkage, and process-induced deformation in flat and L-shaped laminates. It formulates cure-cycle optimization as minimizing deformation subject to full-cure constraints and solves the problem with constrained Bayesian optimization using Gaussian-process surrogates. Relative to a genetic algorithm, the reported deformation and final degree-of-cure errors are below $1-p(x)$1, while computational efficiency exceeds $1-p(x)$2 because cBO converges in fewer than $1-p(x)$3 simulations versus more than $1-p(x)$4 for GA (Limaye et al., 30 May 2025).

These two uses are unrelated methodologically. One concerns curricular design in laboratory education; the other concerns thermo-chemo-mechanical optimization in manufacturing.

6. Cross-disciplinary interpretation

The cited literature does not support treating CURE as a single framework. In survival analysis, it is a substantive concept centered on long-term survivors and defective event-time distributions; in engineering, “cure” denotes resin crosslinking; in education, CURE denotes a course design format; and in machine learning, the acronym labels several unrelated optimization or representation-learning methods (Delhelle et al., 2024, Limaye et al., 30 May 2025, Markert et al., 30 May 2026, Shivam et al., 23 Sep 2025, Elshaer et al., 16 Oct 2025, Messina et al., 21 Jan 2026, Kim et al., 3 Jul 2026, Biswas et al., 19 May 2025, Dong et al., 2019).

Several limitations and controversies recur across these domains. In cure modeling, assuming independent censoring when censoring is dependent can seriously bias survival and cure estimates, and fully parametric approaches remain sensitive to misspecification (Delhelle et al., 2024). In physics education, the cited results challenge the claim that broad relevance by itself yields superior outcomes over other well-designed experimentation-based labs (Markert et al., 30 May 2026). In medical QA, the binary “Sure”/“Not Sure” gate is explicitly described as simplistic (Elshaer et al., 16 Oct 2025). In diffusion-model concept unlearning, the forgetting-strength parameter must balance suppression against collateral damage (Biswas et al., 19 May 2025). In anatomy-grounded report generation, improved entailment is accompanied by the practical difficulty of evaluating hallucination with auxiliary models and heterogeneous annotations (Messina et al., 21 Jan 2026).

This suggests that interpreting “CURE” requires immediate attention to disciplinary context, observables, and target variable. The same label may refer to cure fraction $1-p(x)$5 in survival data, degree of cure $1-p(x)$6 in composites, a curricular format in laboratory pedagogy, or an acronymic method in machine learning.