Naomi: A Cross-Domain Overview

• Naomi is a cross-domain term referring to systems in VLTI adaptive optics, non-autoregressive sequence imputation, edge AI/ML workflows, spatial HIV synthesis, and scholarly analysis of manufactured doubt.
• Its applications include enhancing interferometric image quality with NAOMI in astronomy, reducing prediction errors in time series with a divide-and-conquer imputation technique, and streamlining modular AI/ML pipelines for edge deployments.
• Additionally, Naomi underpins high-dimensional Bayesian inference in epidemiological models and inspires formal models of propaganda that quantify how selective evidence distorts policy-maker beliefs.

In contemporary research literature, Naomi or NAOMI denotes several unrelated entities that share a label but belong to distinct technical domains. The term is used for an adaptive optics upgrade in Very Large Telescope Interferometer instrumentation, a non-autoregressive sequence-imputation model, a modular AI/ML workflow system for the network edge, and a spatial evidence synthesis model for district-level HIV estimation in sub-Saharan Africa; it also refers to Naomi Oreskes, whose historical analysis of manufactured doubt motivates formal models of propaganda in epistemic networks (Matter et al., 2016, Liu et al., 2019, Čop et al., 2024, Howes et al., 21 Aug 2025, Weatherall et al., 2018).

1. Principal referents of the name

The shared label covers both acronymic systems and a personal name. In the astronomical and engineering literature, NAOMI is an instrument or platform name. In machine learning and epidemiology, it is a model name. In history and philosophy of science, Naomi refers to Naomi Oreskes, whose work is cited as a substantive intellectual source rather than as an acronymic system.

Domain	Referent	Source
VLTI instrumentation	Adaptive Optics system for the ATs / adaptive optics module for interferometry	(Matter et al., 2016, Woillez et al., 2014)
Sequence modeling	Non-AutOregressive Multiresolution Imputation	(Liu et al., 2019)
Edge AI/ML systems	Democratized AI/ML workflow system for the network edge	(Čop et al., 2024)
HIV epidemiology	Spatial evidence synthesis model for district-level HIV indicators	(Howes et al., 21 Aug 2025)
History and philosophy of science	Naomi Oreskes as the inspiration for a formal model of the “tobacco strategy”	(Weatherall et al., 2018)

This multiplicity matters because the same token appears in citation graphs, software repositories, and instrumentation roadmaps with entirely different semantics. The term therefore functions as a cross-domain homonym rather than a unified research program.

2. NAOMI in VLTI interferometry and adaptive optics

Within the VLTI, NAOMI is the adaptive optics system designed for the Auxiliary Telescopes (ATs). In the MATISSE context, it is introduced as one of two planned VLTI upgrades—“NAOMI, the Adaptive Optics system for the ATs, and GRA4MAT”—required to “fully exploit” MATISSE, especially for protoplanetary disks around low-mass young stars (Matter et al., 2016). Its functional role is to provide real-time wavefront correction for the 1.8 m ATs, improving the image quality and stability of the beams supplied to mid-infrared interferometry and fringe tracking. The paper ties this to coupling efficiency, point-spread-function stability, reduced residual tip-tilt and higher-order aberrations, and ultimately to sensitivity in a thermal-background-limited regime.

In that architecture, NAOMI is upstream of both the GRAVITY fringe tracker and MATISSE. GRAVITY fringe tracking operates in K band, while MATISSE observes in L & M bands (2.8–5.0 µm) and N band (~8–13 µm). MATISSE reaches angular resolutions of about 3 mas in L band and about 10 mas in N band on the longest baselines, governed by the usual relation

$\theta \approx \frac{\lambda}{B}.$

NAOMI does not alter $B$ or $\lambda$, so it does not change the raw interferometric resolution; rather, it affects how effectively that resolution is exploited by stabilizing the wavefront and increasing the near-infrared Strehl ratio. The same paper associates external fringe tracking with a gain of about 2 magnitudes at MATISSE wavelengths relative to operation without a fringe tracker, and describes NAOMI as part of the chain that makes such performance scientifically usable on AT-fed observations (Matter et al., 2016).

The PRIMA astrometry paper frames NAOMI more explicitly as an answer to unstable injection into the fringe sensor units. At that stage, only tip/tilt was corrected, “providing a performance of 50 mas rms on bright targets,” while injection fluctuations were “dominated by the uncorrected higher order modes of the atmospheric turbulence,” with 48% Strehl at 2.2 µm for 0.8 arcsec seeing. The paper states that “to reduce the injection fluctuation an adaptive optic system must be installed on the ATs: this is the NAOMI project,” and links that requirement to the aim of reaching the limiting magnitude required by the PRIMA exoplanet program, namely $K \sim 7$ for the primary star (Woillez et al., 2014).

In both instrument programs, NAOMI is therefore an infrastructural rather than stand-alone scientific system. Its significance lies in enabling stable beam delivery, efficient fringe tracking, and scientifically useful interferometric observables—absolute and differential visibilities, differential phases, and closure phases—on targets that would otherwise be limited by seeing-driven PSF instability.

3. NAOMI as Non-Autoregressive Multiresolution Imputation

In machine learning, NAOMI stands for Non-AutOregressive Multiresolution Imputation, a deep generative model for imputing missing values in long spatiotemporal sequences (Liu et al., 2019). The model is defined for a sequence

$$X = (x_1, x_2, \dots, x_T), \qquad x_t \in \mathbb{R}^d,$$

with a binary mask

$$M = (m_1, m_2, \dots, m_T), \qquad m_t \in \{0,1\},$$

and aims to learn a conditional generative model

$p_\theta(X \mid X \odot M, M).$

Its stated motivation is that deep autoregressive models suffer from error propagation and short-range bias, which become particularly severe for long gaps and arbitrary missingness patterns.

The core design choice is non-autoregressive, multiresolution decoding. Instead of imputing one time step at a time in temporal order, NAOMI fills missing values by a divide-and-conquer, coarse-to-fine schedule. It identifies gaps bounded by known points, predicts midpoint values, inserts them as newly known points, updates contextual encodings, and recursively refines the sequence at finer temporal scales. This makes the dependency structure interval-based rather than strictly left-to-right, and allows each imputation step to use both past and future context.

The implementation described in the paper uses bidirectional RNNs as encoders and a decoder network operating on selected hidden states. A reconstruction objective is combined with adversarial training: $$\min_\theta \max_\phi \; \mathcal{L}_{\text{rec}}(\theta) + \lambda \, \mathcal{L}_{\text{adv}}(\theta,\phi).$$ The paper emphasizes that the model does not rely on teacher forcing in the standard autoregressive sense; instead, the recursive imputation schedule used during training matches the test-time schedule more closely.

Empirically, the reported result is a 60% reduction in average prediction error relative to autoregressive counterparts, together with improved generalization on long-range sequences. Evaluations are described on benchmark datasets from deterministic and stochastic dynamics, including physical systems, multi-agent trajectories, and UCI time-series. The broader methodological significance is that NAOMI places sequence imputation at the intersection of non-autoregressive modeling, hierarchical temporal structure, and adversarially trained generative models, while remaining explicitly targeted at arbitrary missing patterns rather than only contiguous gaps.

4. NAOMI as a democratized AI/ML workflow system for the network edge

A different use of the name appears in networking and systems research, where NAOMI is a solution for democratizing AI/ML workflows at the network edge (Čop et al., 2024). The motivating context is the increasing availability of SDR hardware, open 5G stacks, and edge/cloud resources to actors beyond traditional mobile network operators and hyperscale cloud providers. The system is designed to implement the full AI/ML lifecycle near the network—data preparation, model training, model management, deployment and inference, and continuous operation—while remaining modular, scalable, distributed, and hardware-architecture-independent.

The architecture is built on mainstream open-source components deployed atop Kubernetes. The paper specifies MinIO for storage, Ray for training and data processing, MLflow for model management, Ray Serve for deployment and inference, Flyte for workflow orchestration, and Prometheus + Grafana for monitoring. It is intended to run on distributed clusters of heterogeneous devices, including mixed ARM and x86 nodes, and to support “self-evolving” AI services via automated retraining and redeployment triggered by monitored metrics.

The design is explicitly aligned with the O-RAN AI/ML workflow architecture, whose logical components are data preparation, model training, model management, model inference, and continuous operation. NAOMI is presented as an implementation of this workflow using mainstream MLOps tools rather than bespoke framework-specific subsystems. The paper’s requirements analysis emphasizes openness, ease of use, modularity, self-evolving operation, heterogeneity and virtualization, and distributed services and scalability.

The reported performance comparison uses the O-RAN Software Community AI/ML Framework as a representative baseline. The paper states that NAOMI performs up to 40% better in deployment time and up to 73% faster in AI/ML workflow execution for larger datasets, while performing inference and utilizing resources on par with its counterpart. In the reported setup, the distributed cluster includes a virtual machine and Raspberry Pi 5 nodes, which serves as an existence proof for heterogeneous edge deployment rather than a purely cloud-centric configuration.

In this sense, NAOMI is best understood as an edge-native MLOps stack shaped by 6G and O-RAN assumptions: AI models consume RAN telemetry and KPI data, training and retraining occur continuously, and deployed models must be manageable on self-hosted, heterogeneous infrastructure.

5. Naomi as a spatial evidence synthesis model for HIV indicators

In HIV epidemiology, Naomi is a spatial evidence synthesis model used to produce district-level HIV epidemic indicators in sub-Saharan Africa (Howes et al., 21 Aug 2025). It is described as part of a UNAIDS-supported annual process and is used operationally by more than 35 countries via web-based software. Its purpose is to generate small-area, age–sex specific estimates for policy and programmatic use when direct measurements are sparse or biased.

The model jointly represents several linked outcomes for each district–sex–age stratum $i \equiv (x,s,a)$: $$p_i = p_{x,s,a} \in [0,1], \qquad a_i = a_{x,s,a} \in [0,1], \qquad \lambda_i = \lambda_{x,s,a} > 0, \qquad K_i = K_{x,s,a} \in [0,1].$$ These are, respectively, HIV prevalence, ART coverage, HIV incidence rate, and proportion recently infected. The model also links to ANC prevalence and ANC ART coverage for pregnant women and to cross-district ART attendance probabilities $B$0. Its data inputs include nationally representative household survey data, routine programme data, ANC testing data, and external offsets from Spectrum/EPP describing age patterns, fertility ratios, and incidence/prevalence ratios.

Methodologically, Naomi is formulated as an extended latent Gaussian model (ELGM) with high-dimensional latent field $B$1 and hyperparameters $B$2. In the Malawi case study, the latent field has dimension $B$3 and the hyperparameter vector has $B$4 components. The joint posterior is written as

$B$5

where $B$6 is the precision matrix of the latent Gaussian field. Age effects are modeled with AR(1) structure, spatial effects with BYM2, ICAR, and IID components, and household-survey aggregates are linked through generalized binomial working likelihoods based on design weights and Kish effective sample sizes.

The paper on fast approximate Bayesian inference focuses on improving inference for Naomi. Previous operational inference used empirical Bayes with a Laplace approximation in TMB. The new method extends adaptive Gauss-Hermite quadrature to high-dimensional hyperparameter spaces by applying PCA to the inverse curvature at the Laplace mode and constructing quadrature nodes along the dominant principal directions. In Malawi, 87% explained variance is achieved with $B$7 principal components, reducing a notional $B$8 grid to $B$9 nodes. Reported runtimes are 0.015 hours for empirical Bayes Laplace, 1.4 hours for GPCA-AGHQ on a laptop, and 71 hours for NUTS HMC on a cluster.

Against NUTS as a gold standard, GPCA-AGHQ improves latent-field accuracy relative to empirical Bayes. The reported RMSE for posterior means decreases from 0.066 to 0.049, and the RMSE for posterior standard deviations decreases from 0.16 to 0.042. Distributional agreement also improves, with average KS decreasing from 0.081 to 0.072 and third-order MMD decreasing from 0.0036 to 0.0024. The practical interpretation is that Naomi is both a substantive epidemiological model and a vehicle for deterministic approximate Bayesian inference in high-dimensional latent Gaussian systems.

6. Naomi Oreskes and the epistemic-network model of manufactured doubt

In the philosophy-of-science literature, Naomi refers to Naomi Oreskes, identified as the central historical and conceptual inspiration for a formal model of propaganda in epistemic networks (Weatherall et al., 2018). The paper “How to Beat Science and Influence People: Policy Makers and Propaganda in Epistemic Networks” translates phenomena documented by Oreskes and Erik Conway in Merchants of Doubt into an agent-based model. The focus is the “tobacco strategy”, which the paper formalizes as a two-pronged mechanism for distorting policy-maker belief.

The two components are selective sharing and biased production. In selective sharing, independently produced studies that happen to support the preferred position are aggressively disseminated, while contrary studies are de-emphasized. In biased production, an interested actor funds or runs studies and selectively publishes only those with favorable outcomes. The model embeds these tactics in an epistemic network derived from Bala–Goyal and Zollman-style frameworks, with scientists, policy makers, a propagandist, and, in one variant, a journalist.

Scientists conduct experiments in a two-armed bandit setting, policy makers update on the evidence they receive, and all updating is Bayesian: $\lambda$0 A semi-analytic description of policy-maker belief is given by

$\lambda$1

where $\lambda$2 captures the average truth-directed force of scientific evidence and $\lambda$3 captures the average misleading force of the propagandist’s evidence. This produces a “tug-of-war” dynamic in which policy makers can become maximally wrong even when the science is unbiased and all non-industry agents are rational Bayesians.

The paper identifies conditions that make the strategy especially effective: small effect sizes, low per-study sample size, many scientists with limited direct access by policy makers, and communication structures that amplify skewed subsets of evidence. It also models journalistic “fair” or “balanced” reporting. Under the paper’s results, sharing all evidence produces the most accurate policy-maker beliefs; random subsampling slows convergence; and “fair” reporting, by forcing one study from each side, never improves epistemic outcomes relative to the other regimes and often makes them substantially worse.

Here, “Naomi” is not an acronymic system but the name of the historian whose empirical case studies provide the mechanism later abstracted into a formal model. The significance of this use lies in the passage from archival historical narrative to mathematical representation of how genuine but selectively curated evidence can sever public and policy belief from scientific consensus.