Neo: Diverse Scientific and Technical Uses

• Neo is a polysemous term encompassing diverse methods and models, from scene outpainting in computer vision to near-Earth object analysis in astronomy.
• Each usage of 'Neo' signifies a refined framework, such as improved test-time adaptation in machine learning and advanced query optimization in databases.
• Neo serves as a concise marker for innovative constructs across disciplines, enabling practical advancements in fields like genetics, cosmology, and arithmetic theory.

“Neo” is a polysemous technical label that appears in contemporary research as an acronym, a retained laboratory shorthand, and a productive prefix. In the literature represented here, it denotes methods for faithful field-of-view extrapolation, target-driven visual navigation, learned query optimization, native vision-language modeling, test-time adaptation, and non-equilibrium sampling; it also denotes near-Earth objects in astronomy, the Nuclear–Electronic Orbital framework in chemical dynamics, the neomycin resistance cassette in mouse genetics, Novel Epitope Optimization in cancer-vaccine modeling, neo-Newtonian cosmology, and neo balcobalancing numbers in arithmetic research (Yu et al., 2023, Wu et al., 2019, Marcus et al., 2019, Diao et al., 16 Oct 2025, Murphy et al., 7 Oct 2025, Thin et al., 2021, Grav et al., 2023, Li et al., 2023, Dudas et al., 6 May 2025, Basava, 2024, Fabris et al., 2012, Tekcan, 14 Apr 2025).

1. Terminological scope

The term spans several naming regimes. In some cases it is an explicit acronym: NEO as NeRF-Enhanced Outpainting, Next Expected Observations, Novel Epitope Optimization, or No-Optimization Test-Time Adaptation through Latent Re-Centering. In other cases it is a conventional domain label, most prominently near-Earth object in astronomy. Elsewhere it is a prefix marking a modified variant of an older framework, as in neo-Newtonian cosmology and neo balcobalancing numbers. In mouse genetics, Neo denotes the neomycin resistance cassette retained from gene-targeting constructs (Yu et al., 2023, Wu et al., 2019, Basava, 2024, Murphy et al., 7 Oct 2025, Grav et al., 2023, Fabris et al., 2012, Tekcan, 14 Apr 2025, Dudas et al., 6 May 2025).

Form	Expansion or meaning	Domain
NEO	NeRF-Enhanced Outpainting	Computer vision (Yu et al., 2023)
NEO	Next Expected Observations	Visual navigation (Wu et al., 2019)
Neo	Neural Optimizer	Database systems (Marcus et al., 2019)
NEO	native VLM family	Vision-language modeling (Diao et al., 16 Oct 2025)
NEO	No-Optimization Test-Time Adaptation through Latent Re-Centering	Machine learning (Murphy et al., 7 Oct 2025)
NEO	Non Equilibrium Sampling on the Orbit of a Deterministic Transform	Monte Carlo methods (Thin et al., 2021)
NEO	near-Earth object	Astronomy and planetary defense (Grav et al., 2023)
NEO	Nuclear–Electronic Orbital	Chemical dynamics (Li et al., 2023)
NEO	Novel Epitope Optimization	Cancer-vaccine modeling (Basava, 2024)
Neo	neomycin resistance cassette	Mouse genetics (Dudas et al., 6 May 2025)
neo-Newtonian	pressure-gravitating Newtonian cosmology	Cosmology (Fabris et al., 2012)
neo balcobalancing	modified balancing-number family	Number theory (Tekcan, 14 Apr 2025)

This distribution suggests that “Neo” functions less as a single concept than as a recurring marker of methodological redesign. In acronymic uses, it usually names a full pipeline or framework. In prefixal uses, it denotes a modified descendant of an established classical construction.

2. Vision, navigation, and multimodal modeling

In computer vision, NeRF-Enhanced Outpainting (NEO) defines a scene-specific solution to faithful field-of-view extrapolation. Its central distinction from conventional outpainting is that the added image region is required to match the actual scene geometry and appearance rather than merely appear plausible. The pipeline trains a NeRF on pre-captured views, renders paired narrow- and extended-FOV images across dense pose samples, and then trains a deterministic 2D outpainting model for pose-free inference. On Replica, Gibson, HM3D, and ScanNet, the reported NEO scores are 25.94 / 0.868 / 0.217, 23.53 / 0.822 / 0.263, 21.54 / 0.731 / 0.338, and 22.40 / 0.793 / 0.168 in PSNR / SSIM / LPIPS, respectively, outperforming practical baselines while remaining below the Oracle NeRF upper bound (Yu et al., 2023).

In embodied navigation, NeoNav uses Next Expected Observations (NEO) as a one-step imagination mechanism for target-driven action selection. The model is a conditional variational Bayesian system whose posterior is conditioned on the current multi-view observation and the target view, while the latent prior is action-conditioned through a mixture-of-posteriors construction. The paper reports that on AVD cross-scene generalization with stop action, NeoNav achieves 19.9 / 5.9 average success / SPL, compared with 14.2 / 1.8 for TD-A3C and 13.3 / 1.9 for I2A; without stop action on AVD, NeoNav reaches 47.7 / 25.5, and on AI2-THOR with stop it reports 16.7 / 7.9 (Wu et al., 2019). The method positions “Neo” not as a geometric prior, as in outpainting, but as an explicit predictive latent variable for forward dynamics.

A third usage appears in large multimodal models. The NEO family of native VLMs is built around Native-RoPE, Multi-Head Native Attention, and a Pre-Buffer and Post-LLM training partition that later dissolves into a single monolithic backbone. The reported training footprint is 390M image-text examples, and the evaluated models include NEO-2.2B and NEO-9B, built on Qwen3-1.7B and Qwen3-8B. On general multimodal benchmarks, the 2.2B model reports MMMU 48.6, MMBench-EN 76.0, and MMVet 49.6, while the 9B model reports MMMU 54.6, MMBench-EN 82.1, and MMVet 53.6 (Diao et al., 16 Oct 2025). Here the term marks a native, early-fusion alternative to modular vision-language architectures.

3. Learned optimization, inference, and adaptation

In database systems, Neo denotes a learned query optimizer that replaces classical hand-crafted cost-model components with a deep value model and a learned search procedure. It is bootstrapped from an expert optimizer such as PostgreSQL, then updated from observed execution latencies. The plan representation combines a query-level encoding with a tree-structured plan-level encoding, and search is executed as DNN-guided best-first enumeration over join order, physical operators, and access paths. On the JOB benchmark executed on PostgreSQL, Neo’s plans take 60% of the time of PostgreSQL’s own plans; on MS SQL Server/JOB and Corp they are reported as 10% faster on average than the commercial optimizer’s plans; and the system required less than two hours to outperform PostgreSQL’s plans and half a day to match or exceed commercial optimizers across engines (Marcus et al., 2019).

In test-time adaptation, NEO stands for No-Optimization Test-Time Adaptation through Latent Re-Centering. The method assumes that the dominant target-domain shift in latent space can be approximated by a translation and simply subtracts the target embedding mean, $z' = \phi(x) - \mu_T$, before the linear classifier head. Because this is equivalent to a bias correction in the classifier, the method is hyperparameter-free in its core form and adds almost no compute beyond inference. The paper reports that on ViT-Base and ImageNet-C, accuracy increases from 55.6% to 59.2% after adapting on one batch of 64 samples; with 512 samples, NEO beats all 7 comparison methods on ImageNet-C, ImageNet-R and ImageNet-S and 6/7 on CIFAR-10-C. On Raspberry Pi and Jetson Orin Nano, it reports 63% lower inference time and 9% lower memory usage relative to the baselines (Murphy et al., 7 Oct 2025).

A more formal methodological use appears in NEO: Non Equilibrium Sampling on the Orbit of a Deterministic Transform. There, the target distribution is written as $\pi(x)\propto \rho(x)L(x)$, and an invertible map $T$ is used to generate forward and backward orbit points that are combined through orbit weights. NEO-IS yields an unbiased estimator of the normalizing constant $Z$, while NEO-MCMC uses iterated sampling-importance-resampling on those orbits. The paper states that NEO-MCMC is uniformly geometrically ergodic and gives explicit mixing-time estimates under mild conditions; with conformal Hamiltonian integrators, NEO-IS achieves state-of-the-art performance on difficult benchmarks and NEO-MCMC can explore highly multimodal targets (Thin et al., 2021). In this usage, “Neo” signals deliberate use of dynamics that do not preserve the target distribution and are corrected by importance weighting.

4. Near-Earth objects and planetary defense

In astronomy, NEO means near-Earth object, defined as an asteroid or comet whose orbit brings it within 1.3 au of the Sun. A central subset is the near-Earth asteroid population, divided into Apollos, Amors, Atens, and Atiras according to semimajor axis and perihelion or aphelion constraints. As of June 2023, the MPC catalog includes >32,100 NEAs; about one quarter are Potentially Hazardous Asteroids under the paper’s MOID-based definition, and current completeness for D ≥ 140 m objects is estimated at ~38% at the end of 2022. The Known Object Model developed for NEO Surveyor projects that, after the nominal five-year mission, overall completeness should reach ~76% for D ≥ 140 m and ~82% for PHAs (Grav et al., 2023).

Operationally, NEO work is structured by triage, survey modeling, and follow-up. The digest2 code classifies short-arc tracklets and returns a D2 pseudo-probability for the NEO class; the operative threshold for NEOCP posting is D2 = 65. The paper reports that 94% of simulated NEOs reached D2 = 100 at least once and 99.6% achieved D2 ≥ 65 at least once during a simulated 10-year interval (Keys et al., 2019). In Rubin simulations, LSST is projected to submit ~129 new candidates per night in its first year under current NEOCP criteria, with ~8.3% purity; a self-recovery predictor reduces the external follow-up list to ~64 candidates per night with ~8.4% purity (Wagg et al., 2024). The NEOExchange portal coordinates follow-up on the Las Cumbres Observatory network and reported 39,300 measurements to the MPC from July 2014–July 2018, with nearly 9,000 measurements contributing to confirmation of ~1,300 NEOs (Lister et al., 2021).

Orbit prediction and dynamical risk assessment form another layer of NEO usage. NEOPROP is a two-module SSA-NEO tool with an Analytical Module for preliminary orbit determination and analytical MOID computation, and a Numerical Module for refined propagation, close approaches, and virtual-asteroid analysis. The paper reports agreement with reference systems generally within about 1% for key quantities and highlights the importance of adaptive step control during close encounters (Zuccarelli et al., 2016). At the population-dynamics level, simulations of tidal disruption at Earth find that catastrophic s-class events produce fragment families with cumulative magnitude distribution $N(H)\propto 10^{(0.55\pm0.04)H}$, a mean coherence time of $\bar{\tau}_c=(14.7\pm0.6)\times 10^3$ years, detectability lifetimes of roughly 2,000–12,000 years, and an ≈5× higher Earth collision probability than the background NEO population (Schunová et al., 2014).

Long-term phase-space structure is similarly encoded through secular dynamics. A semi-analytical map of the NEO region finds that the apsidal secular resonances $\nu_2$, $\nu_3$, $\nu_4$, $\nu_5$, and $\pi(x)\propto \rho(x)L(x)$0 occur well inside the NEO region, while nodal resonances such as $\pi(x)\propto \rho(x)L(x)$1, $\pi(x)\propto \rho(x)L(x)$2, and $\pi(x)\propto \rho(x)L(x)$3 are mainly present at smaller semimajor axis (Fenucci et al., 2023). Survey architecture interacts strongly with these orbital properties. A comparative study of LSST, Sentinel, NEOCam, and a CubeSat-5 constellation finds that, for 140–1000 m general NEOs, the instantaneous integrated fraction in the search volume is roughly ~16%, ~28%, and ~41% for LSST under low-, medium-, and high-brightness assumptions, whereas NEOCam and Sentinel are both near ~7.5% in the same simplified framework (Myhrvold, 2015). Across these studies, NEO is not a single instrument or algorithm but the astronomical object class around which survey completeness, classification, follow-up, and dynamical theory are organized.

5. Physical theory and multicomponent dynamics

In chemical dynamics, NEO abbreviates the Nuclear–Electronic Orbital framework, in which electrons and selected light nuclei, typically protons, are treated quantum mechanically on the same footing, while heavier nuclei remain classical. The paper on the electronic Born–Oppenheimer approximation within RT-NEO-TDDFT introduces a scheme in which the electronic density is quenched to the instantaneous ground state at each time step, while the quantum nuclear density is propagated in real time. This removes the fast electronic time scale from explicit propagation and allows an order-of-magnitude larger time step. For HCN, the BO-RT-NEO approach remains accurate up to 0.387 fs, whereas full RT-NEO required Δtq ≤ 0.024 fs for accurate spectra; in semiclassical cavity simulations, BO-RT-NEO is stable at 0.194 fs versus 0.005 fs for the non-BO case, and the overall speedup for spectra is reported as ≥8× (Li et al., 2023).

The same work also resolves a qualitative artifact in vibrational polariton simulations. Previous semiclassical RT-NEO-TDDFT calculations could show an unphysical asymmetric vibrational Rabi splitting, even for small splittings. Under the BO approximation, the electrons remain in the ground state, the spurious asymmetry disappears, and the lower and upper polariton peaks become symmetric about the cavity frequency (Li et al., 2023). In malonaldehyde, both RT-NEO-Ehrenfest and BO-RT-NEO-Ehrenfest describe proton delocalization during intramolecular proton transfer, whereas a classical proton treatment does not.

A structurally different but equally important physical use is neo-Newtonian cosmology. This framework modifies Newtonian cosmology so that pressure acts as an active source of gravity, reproducing the Friedmann and Raychaudhuri equations of general relativity for a homogeneous and isotropic universe. Its defining replacements include the continuity law $\pi(x)\propto \rho(x)L(x)$4, the Euler equation with inertial term $\pi(x)\propto \rho(x)L(x)$5, and a Poisson equation of the form $\pi(x)\propto \rho(x)L(x)$6 (Fabris et al., 2012). The framework retains Euclidean space, absolute time, and a scalar gravitational potential, yet captures the gravitational role of pressure and the correct sub-horizon linear perturbation dynamics for cosmological fluids. Here “neo” signifies a minimal extension of Newtonian theory rather than a newly coined acronym.

6. Biomedicine, genetics, and arithmetic sequences

In computational immunology, NEO stands for Novel Epitope Optimization, an ensemble deep-learning system for prioritizing patient-specific neoepitopes in personalized cancer vaccines. The reported training corpus includes an NCI cohort of 70 individuals and 2,433,281 unique epitopes. The architecture combines an FFNN branch over non-sequential features with an RNN with LSTM layers over peptide sequences and aggregates their outputs by fixed-weight averaging. Reported performance includes AUC = 0.9166, recall = 91.67 percent as a headline result, and inference on 40,000 neoepitopes in 69.89 milliseconds (Basava, 2024). In this setting, “NEO” is a systems-level optimizer over epitope binding and immunogenicity signals.

In mouse genetics, Neo refers to the neomycin resistance cassette inserted during ES-cell targeting. A study of Arc/Arg3.1 knockout mice compares Neo+ and Neo− lines and concludes that severe maternal and social deficits are specific to the retained cassette rather than to Arc deletion alone. In the reported maternal assay, 93% of Arc/Arg3.1−/− Neo+ dams failed to retrieve all pups within 30 min, whereas the corresponding Neo− lines did not show the same severe phenotype. The same work reports altered ERK signaling, dysregulated gene expression, and reduced oxytocin fiber density in PFC and NAC, while noting that the Neo cassette product itself did not produce cytotoxicity (Dudas et al., 6 May 2025). Here “Neo” is neither acronym nor prefix but the standard shorthand for a selectable marker with major phenotypic consequences.

A final mathematical use appears in neo balcobalancing numbers. The paper defines neo balcobalancing numbers, neo Lucas-balcobalancing numbers, neo balcobalancers, and neo Lucas-balcobalancers, derives their general terms in terms of classical balancing numbers, and proves that the defining square condition reduces to the Pell-type equation $\pi(x)\propto \rho(x)L(x)$7. The resulting sequences satisfy a common third-order recurrence,

$\pi(x)\propto \rho(x)L(x)$8

and the paper develops Binet formulas, relations with Pell and Pell–Lucas numbers, connections with triangular and square triangular numbers, Pythagorean triples, Cassini identities, and formulas for sums of first $\pi(x)\propto \rho(x)L(x)$9 terms (Tekcan, 14 Apr 2025). In this arithmetic setting, “neo” denotes a modified descendant of classical balancing-number theory.

Across these literatures, “Neo” functions as a compact indicator of reformulation. Sometimes it names a concrete pipeline or model family; sometimes it marks an object class central to planetary defense; sometimes it denotes a retained genetic cassette or a mathematically altered sequence class. The common thread is not semantic uniformity but technical repurposing: each usage identifies a new construction built against, within, or beyond an established framework.