ENCORE: A Multidisciplinary Research Framework

• ENCORE is a diverse research paradigm encompassing strong gravitational lens cosmography, semi-supervised segmentation, numerical reasoning, and other domains with significant empirical achievements.
• It employs advanced methodologies such as adaptive thresholding, time-delay lens modeling, isotropic basis expansions, and ensemble learning to optimize performance and reduce systematics.
• The framework enables actionable insights across disciplines, from precision cosmological measurements and program repair to robust web censorship detection and enhanced NLP reasoning.

ENCORE

ENCORE refers to a diverse set of research frameworks, methodologies, and systems across computer science, cosmology, natural language processing, and semi-supervised learning. The following article surveys the principal instantiations bearing the ENCORE designation, focusing on their theoretical foundations, algorithmic constructions, domain-specific advances, and empirical achievements, with citations to primary literature.

1. ENCORE in Strong Gravitational Lens Cosmography: Supernova Encore

The term ENCORE designates a strongly lensed Type Ia supernova (SN) at $z=1.949$ discovered behind the galaxy cluster MACS J0138.0–2155 ($z_\ell=0.336$), notable as the second multiply-imaged SN in the same host, MRG–M0138, after SN Requiem (Pierel et al., 2024). Both events offer rare opportunities for time-delay cosmography: leveraging measured time delays between multiple SN images to infer the Hubble constant ($H_0$) independent of local distance ladders or CMB analyses.

Discovery and Identification

ENCORE was detected in JWST/NIRCam F150W imaging on 2023 Nov 17; subsequent NIRSpec IFU confirmed its normal, fast-declining SN Ia nature (rest-frame phase ≈30 d past maximum) (Pierel et al., 2024, Dhawan et al., 2024). The lensing cluster yields three visible SN images; modeling predicts additional delayed images in future years, providing extended baselines for cosmographic inference (Bazzanini et al., 23 Jun 2026).

Lens Modeling and Time-Delay Measurement

Multiple independent strong-lensing mass models (parametric and free-form) have been constructed, incorporating multiband imaging (HST, JWST), deep spectroscopy (VLT/MUSE), and stellar kinematics of cluster members (Acebron et al., 12 Mar 2025, Ertl et al., 12 Mar 2025). The standard formalism utilizes the lens equation,

$$\boldsymbol\beta = \boldsymbol\theta - \nabla \psi(\boldsymbol\theta),$$

and Fermat potential,

$$\phi(\boldsymbol\theta) = \frac{1}{2}|\boldsymbol\theta-\boldsymbol\beta|^2 - \psi(\boldsymbol\theta),$$

to relate the observed time-delays,

$$\Delta t_{ij} = \frac{D_{\Delta t}}{c}\,\Delta\phi_{ij},\quad D_{\Delta t}=(1+z_\ell)\frac{D_\ell D_s}{D_{\ell s}},$$

directly to $H_0$ through $D_{\Delta t}\propto H_0^{-1}$ (Suyu et al., 15 Sep 2025, Pierel et al., 15 Sep 2025, Bazzanini et al., 23 Jun 2026).

The measured delay between images 1b and 1a, $\Delta t_{1b,1a}=-39.8_{-3.3}^{+3.9}$ days, combined with ensemble mass models, yields

$$H_0 = 66.9_{-8.1}^{+11.2}\;\mathrm{km\,s^{-1}\,Mpc^{-1}}$$

(Suyu et al., 15 Sep 2025, Pierel et al., 15 Sep 2025, Bazzanini et al., 23 Jun 2026).

Systematics Control and Kinematic Anchoring

Redshift catalogs (107 objects, 50 cluster members, 13 lensed background images) derived from VLT/MUSE enable detailed Faber–Jackson calibrations,

$z_\ell=0.336$0

for member galaxies, breaking degeneracies in mass modeling (Granata et al., 2024). This kinematic anchoring is essential for suppressing the mass-sheet and galaxy truncation degeneracies, bolstering the credibility of derived $z_\ell=0.336$1 posteriors.

Future Prospects

Forthcoming delayed images of SN Encore and sibling SN Requiem, with predicted time delays $z_\ell=0.336$28–10 years, are expected to constrain $z_\ell=0.336$3 to uncertainties $z_\ell=0.336$4 provided systematics remain subdominant (Pierel et al., 2024, Suyu et al., 15 Sep 2025). The presence of two lensed SNe in a single host is unprecedented and enables cross-calibration and systematics checks unique among time-delay cosmography systems.

2. ENCORE in Semi-Supervised Semantic Segmentation: Ensemble-of-Confidence Reinforcement

ENCORE has also been introduced as "Ensemble-of-Confidence Reinforcement," a dynamic thresholding framework for pseudo-label selection in semi-supervised semantic segmentation (Ghamsarian et al., 12 May 2025). Unlike conventional methods that apply static, class-agnostic confidence thresholds, ENCORE adaptively estimates and updates class-wise thresholds in response to model feedback, optimizing the reliability of pseudo-labels drawn from an unlabeled dataset.

Core Algorithmic Mechanism

Let $z_\ell=0.336$5 be the number of semantic classes. At iteration $z_\ell=0.336$6, for each class $z_\ell=0.336$7, ENCORE computes a set of accepted pixels $z_\ell=0.336$8 as those for which the teacher’s pseudo-label equals $z_\ell=0.336$9 and the maximum softmax confidence exceeds the class-specific threshold $H_0$0. The class-wise true-positive rate $H_0$1 is estimated as the proportion where the student (on a strongly augmented view) agrees with the teacher. The threshold is updated via: $H_0$2 where $H_0$3 is a learning rate (Ghamsarian et al., 12 May 2025). This adaptive mechanism improves segmentation performance, especially for underrepresented or difficult classes, and diminishes reliance on hand-tuned hyperparameters.

3. ENCORE in Numerical Reasoning for Natural Language Processing

Encore (Enhancing NumeriCal reasOning with Reliable procEsses) is an approach to numerical reasoning in NLP tasks, which generates verifiable, structured reasoning traces directly supported by the input evidence (Wang et al., 2024).

Formula Decomposition and Structured Annotation

Given a gold answer formula $H_0$4, Encore decomposes it into operands $H_0$5, operators $H_0$6, and a located formula $H_0$7 (operand–evidence mappings), forming a tuple $H_0$8. Training objectives jointly optimize answer prediction and reasoning sequence generation: $H_0$9 This approach outperforms LLM-generated rationales by eliminating spurious content and delivers improved execution accuracy and exact match performance on hybrid QA datasets. Ablation confirms operands and located formulas are most critical (Wang et al., 2024).

4. ENCORE for Efficient Cosmological $$\boldsymbol\beta = \boldsymbol\theta - \nabla \psi(\boldsymbol\theta),$$0-Point Correlation Functions

ENCORE is the name of an algorithm and corresponding code for efficiently estimating galaxy $$\boldsymbol\beta = \boldsymbol\theta - \nabla \psi(\boldsymbol\theta),$$1-point correlation functions (NPCFs) up to $$\boldsymbol\beta = \boldsymbol\theta - \nabla \psi(\boldsymbol\theta),$$2 (Philcox et al., 2021).

Isotropic Basis Expansion and Algorithmic Scaling

NPCFs

$$\boldsymbol\beta = \boldsymbol\theta - \nabla \psi(\boldsymbol\theta),$$3

are expanded in separable isotropic bases constructed from spherical harmonics, $$\boldsymbol\beta = \boldsymbol\theta - \nabla \psi(\boldsymbol\theta),$$4: $$\boldsymbol\beta = \boldsymbol\theta - \nabla \psi(\boldsymbol\theta),$$5 Computation reduces to $$\boldsymbol\beta = \boldsymbol\theta - \nabla \psi(\boldsymbol\theta),$$6 one-point convolutions, yielding $$\boldsymbol\beta = \boldsymbol\theta - \nabla \psi(\boldsymbol\theta),$$7 runtime for $$\boldsymbol\beta = \boldsymbol\theta - \nabla \psi(\boldsymbol\theta),$$8 galaxies, a substantial reduction from the brute-force $$\boldsymbol\beta = \boldsymbol\theta - \nabla \psi(\boldsymbol\theta),$$9. FFT-based implementations achieve $$\phi(\boldsymbol\theta) = \frac{1}{2}|\boldsymbol\theta-\boldsymbol\beta|^2 - \psi(\boldsymbol\theta),$$0 scaling (Philcox et al., 2021). GPU acceleration is provided for high-order terms.

Edge correction, random catalogs, and orthonormality of the basis are natively handled. The approach enables routine estimation of the 4-, 5-, and 6-point functions from survey-scale datasets.

5. ENCORE as Lightweight Measurement of Web Censorship

ENCORE is also the designation of a web-scale censorship measurement platform leveraging cross-origin requests from unmodified browsers to infer filtering at scale, without the need for client-side installations (Burnett et al., 2014).

Measurement Primitives and Architecture

ENCORE injects innocuous cross-origin HTML elements (e.g., <img>, <link>, <iframe>, <script>) into pages and uses side-channel signals (onload/onerror events, timing, DOM mutations) to detect resource accessibility. The system architecture consists of:

• Origin web servers embedding ENCORE snippets.
• A centralized coordination server distributing measurement tasks.
• A collection server aggregating client-reported results.

Filtering is statistically detected using binomial hypothesis tests on success rates per region/resource, robust to network loss.

Deployment and Case Studies

Pilot and global deployments yielded 141,626 measurements from 170 countries, detecting well-known censorship patterns (e.g., blocking of youtube.com, twitter.com) and offering near-real-time, fine-grained inference of domain-level filtering. Ethical concerns are addressed via restricted targeting and institutional oversight (Burnett et al., 2014).

6. ENCORE in Automatic Program Repair

ENCORE represents an ensemble-learning approach based on convolutional neural machine translation (NMT) models for generate-and-validate (G&V) program repair (Lutellier et al., 2019).

Model Construction and Ensemble Strategy

ENCORE formulates patch generation as a token-level translation task from buggy to fixed statements using stackable 1D convolutional NMT models. Diverse models are obtained through hyper-parameter randomization; inference ensembles are constructed by selecting candidates with minimal negative log-likelihood across models rather than aggregating by majority vote.

Empirical Results

ENCORE demonstrates state-of-the-art performance over benchmarks such as Defects4J and QuixBugs, fixing 42 Java bugs (16 previously unsolved) and extending to Python, C++, and JavaScript without architecture modification—solving 67 bugs in total (Lutellier et al., 2019). Ablation reveals the superiority of convolutional architectures and empirical gains from ensemble diversity.

7. ENCORE for Fine-Grained Entity Typing via Coreference Contrastive Pre-Training

EnCore pre-trains entity encoders for fine-grained entity typing by optimizing mention representations to cluster coreferring mentions, using the intersection of two coreference systems for high-precision chains and a contrastive InfoNCE loss (Mtumbuka et al., 2023).

Model Formulation

Coreferring mention pairs are contrastively maximized, negatives drawn from other documents, and head-token masking enhances context reliance. Combined pre-training with MLM yields substantial gains in micro/macro F1 across all evaluated FET datasets. Empirical analysis shows greatest gains for deeper label granularity, with EnCore outperforming MLM-only and prior FET systems.

8. ENCORE in Early Universe Cosmology: The Tachyonic Encore Mechanism

"Tachyonic ENCORE" refers to a dynamical mechanism during (p)reheating where a spectator axion field, initially frozen near its hilltop, induces late-time turns and tachyonic phases in the multi-field inflationary trajectory, yielding a scale-invariant enhancement of the curvature perturbation, suppression of tensor-to-scalar ratio, and $$\phi(\boldsymbol\theta) = \frac{1}{2}|\boldsymbol\theta-\boldsymbol\beta|^2 - \psi(\boldsymbol\theta),$$1 non-Gaussianity (Gorgulho et al., 10 Jun 2026).

Theoretical Analysis

The curvature power spectrum after the ENCORE phase is

$$\phi(\boldsymbol\theta) = \frac{1}{2}|\boldsymbol\theta-\boldsymbol\beta|^2 - \psi(\boldsymbol\theta),$$2

resulting in

$$\phi(\boldsymbol\theta) = \frac{1}{2}|\boldsymbol\theta-\boldsymbol\beta|^2 - \psi(\boldsymbol\theta),$$3

where $$\phi(\boldsymbol\theta) = \frac{1}{2}|\boldsymbol\theta-\boldsymbol\beta|^2 - \psi(\boldsymbol\theta),$$4 is the entropic tilt. This dynamics accommodates otherwise disfavored inflationary potentials within CMB bounds (Gorgulho et al., 10 Jun 2026).

---

The ENCORE paradigm thus constitutes a multi-domain motif denoting empirically validated systems, algorithms, and mechanisms at the intersection of astrophysics, learning theory, software engineering, NLP, and systems measurement. Each instantiation advances its field by integrating domain-specific innovation (e.g., lensing cosmography, pseudo-label selection, isotropic basis expansions) with contemporary methodological rigor.