CLoE: Diverse Cross-Domain Research

• CLoE is a polysemous term representing distinct research frameworks in lunar planetary science, precision cosmology, medical imaging, and clinical NLP.
• In lunar studies, the Center for Lunar Origins and Evolution uses crater chronology and impact models to constrain early Solar System dynamics and planetary evolution.
• In cosmology and medical imaging, modular code ecosystems and novel training strategies enhance systematic error control and robustness in data analysis.

CLoE is a polysemous research label rather than a single established concept. In the literature represented here, the term and its capitalization variants denote several unrelated entities: the Center for Lunar Origins and Evolution in lunar planetary science, the Cosmology Likelihood for Observables in Euclid software ecosystem in precision cosmology, Curriculum Learning on Endoscopy for Mayo Endoscopic Subscore classification, and Consistency Learning of Experts for missing-modality segmentation; one clinical NLP benchmark is also described with the variant spelling CLoE/CLUE for Clinical Language Understanding Evaluation (Burns et al., 2012, Collaboration et al., 23 Mar 2026, Ozdemir et al., 18 Aug 2025, Tong et al., 10 Mar 2026, Goodwin et al., 2022).

1. Nomenclature and scope

In the cited literature, capitalization is not uniform. CLOE is used in all-caps for the Euclid cosmology ecosystem and for the Center for Lunar Origins and Evolution, whereas CLoE appears in title case for medical-imaging frameworks. A related but distinct spelling, CLUE, is used in clinical NLP (Goodwin et al., 2022).

Usage	Expansion	Domain
CLOE	Center for Lunar Origins and Evolution	Lunar planetary science
CLOE	Cosmology Likelihood for Observables in Euclid	Cosmological inference software
CLoE	Curriculum Learning on Endoscopy	Endoscopic image analysis
CLoE	Consistency Learning of Experts	Multimodal medical segmentation
CLUE/CLoE	Clinical Language Understanding Evaluation	Clinical NLP benchmarking

This multiplicity of meanings has practical consequences for citation and retrieval. In astrophysics, “CLOE” identifies an NLSI lunar-science consortium; in Euclid papers it refers to a theory-and-likelihood stack; in medical imaging it denotes task-specific training frameworks rather than institutions or software ecosystems (Burns et al., 2012, Collaboration et al., 10 Oct 2025, Ozdemir et al., 18 Aug 2025, Tong et al., 10 Mar 2026).

2. Center for Lunar Origins and Evolution

In lunar and planetary science, CLOE is the Center for Lunar Origins and Evolution, one of two astrophysics-focused teams within the NASA Lunar Science Institute, alongside LUNAR (Burns et al., 2012). Its work is identified with the “Understanding New Worlds” theme of New Worlds, New Horizons in Astronomy & Astrophysics, and its stated scientific role is to use the Moon’s geologic record as a probe of planetary-system formation and evolution (Burns et al., 2012).

CLOE treats the Moon primarily as a geologic and dynamical archive rather than as an observing platform. The key empirical resource is the Moon’s preserved cratering chronology and large basin record, especially the imprint of the Late Heavy Bombardment, described as occurring about $0.7\ \mathrm{Gyr}$ after Solar System formation (Burns et al., 2012). Within this framework, lunar basin ages and impact-flux models are used to constrain models of giant-planet migration, dynamical instability, resonance chains, and planetesimal scattering (Burns et al., 2012).

A central result associated with this program is the argument that successful reconstructions of the early Solar System often require an initial system of five giant planets, including an additional ice giant that is later ejected (Burns et al., 2012). The cited resonance-chain example is $(3:2, 3:2, 4:3, 5:4)$, and the lunar bombardment record is used as the timing constraint on the instability that produced the present architecture (Burns et al., 2012). The same work explicitly links these solar-system reconstructions to exoplanet observations such as high eccentricities, resonant systems, and free-floating planets, thereby placing lunar chronology in a broader comparative-planetology context (Burns et al., 2012).

Institutionally, this version of CLOE is not a software package or algorithm. It is a consortium, directed by William Bottke at Southwest Research Institute, whose significance lies in tying lunar geology, dynamical simulations, and exoplanetary context into a single research program (Burns et al., 2012).

3. Cosmology Likelihood for Observables in Euclid

In contemporary cosmology, CLOE denotes Cosmology Likelihood for Observables in Euclid, the Euclid Consortium’s software ecosystem for turning theory and survey specifications into likelihoods and cosmological constraints (Bonici et al., 22 May 2026). The 2025–2026 Euclid preparation papers describe it as the mission’s unified framework for weak lensing, photometric galaxy clustering, galaxy–galaxy lensing, and spectroscopic galaxy clustering, with support for both harmonic-space and configuration-space summary statistics (Collaboration et al., 10 Oct 2025, Collaboration et al., 23 Mar 2026).

The ecosystem is structured around separable layers. In the 2026 architecture, cloelib is the theory-computation layer, cloelike is the likelihood layer, and euclidlib handles official Euclid data products (Bonici et al., 22 May 2026, Bonici et al., 22 May 2026). The code-implementation paper describes CLOE as a modular Python code that computes theoretical predictions and evaluates them against survey data in a unified likelihood, while the likelihood papers describe it as the default analysis code for Euclid cosmology (Collaboration et al., 23 Mar 2026, Bonici et al., 22 May 2026).

At the level of statistical structure, the likelihood is Gaussian, with the standard form

$$-2 \ln \mathcal{L}(\mathbf{d}\,|\,\boldsymbol{\theta}) = (\mathbf{d}-\mathbf{t}(\boldsymbol{\theta}))^\mathrm{T} \mathbf{C}^{-1} (\mathbf{d}-\mathbf{t}(\boldsymbol{\theta})) + b,$$

where $\mathbf{d}$ is the data vector, $\mathbf{t}(\boldsymbol{\theta})$ the theory prediction, and $\mathbf{C}$ the covariance (Bonici et al., 22 May 2026). The theory side implements standard FLRW background quantities, matter power spectra, tracer kernels, Limber-projected $C_\ell$, and spectroscopic multipoles $P_\ell(k)$, while extended modelling includes intrinsic alignments, magnification bias, modified gravity, massive neutrinos, baryonic feedback, and spectroscopic purity (Collaboration et al., 10 Oct 2025, Collaboration et al., 10 Oct 2025, Collaboration et al., 11 Oct 2025).

Several Euclid preparation papers characterize the maturity of this usage of CLOE. The forecasts paper reports a figure of merit for the dark energy parameters $w_0$ and $w_a$ exceeding 400 when all primary probes are combined (Collaboration et al., 10 Oct 2025). The validation paper states that all the summary statistics of interest always differ less than $(3:2, 3:2, 4:3, 5:4)$0 from the chosen benchmarks and that CLOE predictions are statistically compatible with simulated benchmark data (Collaboration et al., 10 Oct 2025). A separate systematics paper finds that intrinsic-alignment mis-modelling can bias $(3:2, 3:2, 4:3, 5:4)$1 by up to $(3:2, 3:2, 4:3, 5:4)$2 and that spectroscopic-sample purity is a major driver of full-shape constraints, illustrating that the Euclid meaning of CLOE is not just a codebase but a vehicle for systematic-error control (Collaboration et al., 11 Oct 2025).

This usage is therefore both infrastructural and methodological. It is a mission-scale cosmological inference stack, organized as a reusable Python ecosystem and explicitly designed to support standard and beyond-$(3:2, 3:2, 4:3, 5:4)$3CDM analyses on Euclid data (Collaboration et al., 23 Mar 2026, Collaboration et al., 10 Oct 2025).

4. Medical-imaging uses of CLoE

In medical imaging, CLoE appears in at least two unrelated framework names. The first is “Curriculum Learning on Endoscopic Images for Robust MES Classification”, where CLoE is a training framework for Mayo Endoscopic Subscore classification in ulcerative colitis (Ozdemir et al., 18 Aug 2025). The second is “CLoE: Expert Consistency Learning for Missing Modality Segmentation”, where CLoE denotes a consistency-driven approach to multimodal medical-image segmentation with arbitrary missing modalities at inference (Tong et al., 10 Mar 2026).

In the endoscopic setting, CLoE addresses two specific difficulties: label noise from inter-observer variability and the ordinal nature of MES labels (Ozdemir et al., 18 Aug 2025). The method uses image quality, estimated by a MobileNetV2 trained on Boston Bowel Preparation Scale labels, as a proxy for annotation confidence; images are partitioned into clean and noisy subsets using a threshold $(3:2, 3:2, 4:3, 5:4)$4, and training proceeds in a three-stage curriculum: clean-only, mixed clean-plus-noisy, and noisy-focused fine-tuning (Ozdemir et al., 18 Aug 2025). The framework is combined with ResizeMix, and on LIMUC the reported best result is ConvNeXt-Tiny with 82.51% top-1 accuracy and QWK 0.8935 (Ozdemir et al., 18 Aug 2025). This usage of CLoE is therefore an optimization schedule and robustness strategy for ordinal image classification, not a benchmark or general software ecosystem (Ozdemir et al., 18 Aug 2025).

In missing-modality segmentation, CLoE expands to Consistency Learning of Experts and is built around modality-specific experts whose predictions are regularized by a dual consistency objective (Tong et al., 10 Mar 2026). The two main components are Modality Expert Consistency (MEC), which enforces global agreement among expert predictions, and Region Expert Consistency (REC), which emphasizes agreement in foreground-critical regions (Tong et al., 10 Mar 2026). These consistency scores are mapped to reliability weights by a lightweight gating network, and the resulting weights recalibrate features before fusion (Tong et al., 10 Mar 2026). On BraTS 2020, the reported average Dice scores are 88.09% for whole tumor, 80.23% for tumor core, and 65.06% for enhancing tumor; on MSD Prostate, CLoE reaches 80.12% average Dice on the peripheral zone (Tong et al., 10 Mar 2026).

The two medical-imaging meanings share a concern with robustness under imperfect supervision or incomplete inputs, but they operate at different levels. One is a curriculum strategy for ordinal endoscopy classification, the other a consistency-and-gating architecture for segmentation (Ozdemir et al., 18 Aug 2025, Tong et al., 10 Mar 2026).

5. Clinical-language variant and adjacent acronym usage

A closely related but distinct usage appears in clinical NLP. The paper “Clinical Language Understanding Evaluation” presents CLUE, a benchmark built from MIMIC-III for disease staging, computational phenotyping, mortality prediction, and remaining length-of-stay prediction (Goodwin et al., 2022). In the supplied description, this benchmark is explicitly referred to as “CLoE (or CLUE: Clinical Language Understanding Evaluation)”, indicating a variant spelling in circulation (Goodwin et al., 2022).

The benchmark defines standardized 8:1:1:1 train, validation, calibration, and test splits at the patient level, stratified on confounders such as age, sex, race, ICU, admission source, admission type, severity, and insurance (Goodwin et al., 2022). Its task structure includes four broad clinical language understanding tasks instantiated as six specific prediction problems, and it provides a software toolkit with HuggingFace integration (Goodwin et al., 2022). Task metrics are heterogeneous by design: disease staging and remaining length-of-stay use MAE, phenotyping uses micro-averaged F1, and mortality uses macro-averaged F1 (Goodwin et al., 2022).

This benchmark is not normally titled “CLoE” in the cited paper; the title and canonical acronym are CLUE (Goodwin et al., 2022). Its relevance here is terminological: it shows that nearby spellings can refer to yet another, unrelated object—a standardized evaluation suite in clinical NLP rather than a lunar center, cosmology framework, or medical-imaging method (Goodwin et al., 2022).

6. Comparative significance and disambiguation in practice

Across these usages, CLoE spans four distinct ontological types. It can denote an institutional research center in lunar science, a mission-scale cosmological inference ecosystem, a task-specific training framework for endoscopy, or a consistency-driven segmentation architecture for multimodal imaging (Burns et al., 2012, Bonici et al., 22 May 2026, Ozdemir et al., 18 Aug 2025, Tong et al., 10 Mar 2026). A plausible implication is that the acronym should never be interpreted without field context.

The Euclid and lunar-science meanings are the most infrastructural. The former is a software and likelihood stack supporting photometric and spectroscopic cosmological inference, with publicly described theory, implementation, validation, forecasting, and systematics papers (Collaboration et al., 10 Oct 2025, Collaboration et al., 23 Mar 2026, Collaboration et al., 10 Oct 2025). The latter is a NASA-supported consortium that uses the Moon’s cratering record to constrain early Solar-System dynamics and to compare Solar-System history with exoplanetary architectures (Burns et al., 2012).

The medical-imaging meanings are narrower and algorithmic. Both are formulated around robustness, but one handles ordinal label uncertainty and the other missing modalities and expert disagreement (Ozdemir et al., 18 Aug 2025, Tong et al., 10 Mar 2026). The clinical-language variant adds a benchmark-centered meaning whose principal contribution is standardization rather than model architecture (Goodwin et al., 2022).

For scholarly usage, the decisive identifier is therefore not the acronym alone but the paired expansion and citation. “CLOE” in a Euclid context typically refers to Cosmology Likelihood for Observables in Euclid; in lunar science it refers to the Center for Lunar Origins and Evolution; in medical-imaging papers, title case CLoE usually names a method rather than an institution (Burns et al., 2012, Bonici et al., 22 May 2026, Ozdemir et al., 18 Aug 2025, Tong et al., 10 Mar 2026).