NoAH: Diverse Models, Methods & Applications

• NoAH is a versatile term with distinct meanings in hydrology, computer vision, anomalous diffusion, multimodal evaluation, and optimization.
• Key applications include enhancing land-surface model calibration, parameter-efficient vision tuning, serverless scheduling, and benchmark evaluation.
• Researchers must rely on citation context to disambiguate NoAH’s usage, ensuring technical precision across diverse domains.

In the cited literature, “NoAH”, “NOAH”, and “Noah” do not denote a single object. The term appears as the Noah land surface model and its Noah‑MP extension in Earth system modeling, as several distinct machine-learning methods and benchmarks, as the Noah effect in anomalous diffusion theory, as a negative class label in multimodal affect recognition, and as the name of algorithms in serverless computing, underwater swarm optimization, and attributed hypergraph generation (Pelissier et al., 2020, Zhang et al., 2022, Lee et al., 9 Nov 2025, Aghion et al., 2020).

1. Terminological scope

The diversity of usages is explicit across recent arXiv literature.

Usage of “NoAH/NOAH/Noah”	Domain	Representative paper
Noah / Noah‑MP land surface model	Hydrology, land–atmosphere modeling	(Pelissier et al., 2020)
Neural prOmpt seArcH (NOAH)	Vision transfer learning, NAS	(Zhang et al., 2022)
Non-glObal Attentive Head (NOAH)	Image classification	(Li et al., 2024)
NOAH benchmark	Video LLM evaluation	(Lee et al., 9 Nov 2025)
NoAH negative class	Ambivalence/hesitancy recognition	(Bekhouche et al., 16 Mar 2026)
Noah effect / Noah exponent $L$	Anomalous diffusion	(Bera et al., 18 Dec 2025)
Nauplius Optimisation for Autonomous Hydrodynamics (NOAH)	Underwater swarm optimization	(Ramesh et al., 17 Oct 2025)
NOAH serverless heuristic	Cloud scheduling	(Stein, 2018)
NoAH attributed hypergraph model	Hypergraph generation	(Chun et al., 26 Sep 2025)

This multiplicity is not merely orthographic. In some papers, “Noah” denotes a long-established physical model; in others, it is an acronym; in others, it denotes a statistical effect or a class label. A practical implication is that citation context is indispensable for disambiguation.

2. Noah in land-surface and hydrologic modeling

In Earth system modeling, Noah denotes the Noah land surface model (Noah LSM), and Noah‑MP denotes its multiparameterization extension. In the hybrid soil-moisture study, Noah is treated as a parametric land surface model (PBM) whose state vector includes soil moisture in multiple layers, soil temperature, snowpack, and related land-surface states; its soil moisture dynamics are governed by infiltration, runoff, percolation or drainage, and evapotranspiration. The paper casts the calibrated physical model as $x_{t+1}=f(x_t)$, and then augments it with a Gaussian-process correction $y(v_t)=f(v_t)+\mathrm{GP}(v_t)$ for top-layer soil moisture. The GP is trained on one-step-ahead Noah residuals, while unobserved states are still propagated by Noah alone. A central design choice is the zero-mean GP prior, which makes the hybrid revert to pure Noah when inputs lie far from the training climatology. In one-year leave-one-out cross-validation over nine AmeriFlux sites, the method reports upwards of a 3-fold reduction in the RMSE, with the largest gain illustrated at Blodgett Forest, where the GP removes a Noah spring snowmelt oscillatory artifact (Pelissier et al., 2020).

A related hydrologic use appears in GRACE data fusion over India, where NOAH refers to GLDAS NOAH v2.1. In that setting, NOAH-simulated terrestrial water storage is the sum of soil moisture storage in four layers from 0–200 cm depth, snow water storage, and canopy water storage; the model does not simulate surface water storage, and it does not account for deep groundwater storage and human intervention. A CNN is trained to learn the mismatch between GRACE total water storage anomalies and NOAH output. The corrected fields achieve a country-average correlation coefficient of 0.94 and Nash-Sutcliff efficient of 0.87, corresponding to 14\% and 52\% improvement respectively over the original NOAH TWSA; the learned mismatch also correlates with in situ groundwater storage anomalies over much of India (Sun et al., 2019).

The same physical backbone is extended again in NoahMP‑AI, which uses the full Noah‑MP v5.0 model as a physics-based feature generator and feeds surface soil moisture, latent heat flux, and sensible heat flux into a 3D U‑Net to predict SMAP soil moisture during Texas extremes. In the reported drought case, Noah‑MP alone gives negative $R^2$, whereas the hybrid model increases $R^2$ from -0.7 to 0.5 during drought conditions, while preserving physical consistency and spatial coherence (Mbarak et al., 15 Jun 2025).

A fourth line of work concerns parameter selection within Noah‑MP rather than hybridization. Over the Iberian Peninsula, 70 one-year simulations using 35 Noah‑MP combinations for 2005 and 2010 identify 10 groups (A to J) and show that the options with greatest impact are canopy stomatal resistance (CRS), surface exchange coefficient for heat (SFC), soil-moisture factor controlling stomatal resistance (BTR), runoff and groundwater (RUN), surface resistance to evaporation/sublimation (RSF), and also dynamic vegetation (DVEG). Among several reasonable configurations, experiment s27I—with Jarvis CRS, Chen97 SFC, CLM-type BTR, BATS RUN, and adjusted Sellers RSF for wet soils—is reported to provide a particularly good characterization of heat fluxes over the Iberian Peninsula (Donaire-Montaño et al., 30 Oct 2025).

3. NOAH in vision architectures and parameter-efficient adaptation

In computer vision, NOAH can denote Neural prOmpt seArcH, a NAS framework for parameter-efficient tuning of large vision models. The method treats Adapter, LoRA, and VPT as a unified search space of prompt modules on top of a frozen ViT‑B/16 backbone, then uses AutoFormer with weight sharing and evolutionary search to discover a dataset-specific configuration under a parameter budget. On VTAB‑1k, the searched subnet reaches 75.5 average accuracy with about 0.43M trainable parameters, exceeding LoRA (74.5), Adapter (73.9), and VPT (72.0); it also reports large gains on domain-shifted targets such as ImageNet‑V2, ImageNet‑Sketch, ImageNet‑A, and ImageNet‑R (Zhang et al., 2022).

A distinct usage is NOAH as Non-glObal Attentive Head, a classification head that replaces the standard global average pooling plus linear classifier in CNNs, ViTs, and MLP-like vision models. Its core operation is Pairwise Object Category Attention (POCA). After channel splitting, each block computes

$$\mathbf{A}_n=\mathrm{softmax}(\mathbf{W}_{kn}*\mathbf{F}_{kn}),\quad \mathbf{V}_n=\mathbf{W}_{vn}*\mathbf{F}_{vn},\quad \mathbf{Z}_n=\mathbf{A}_n\odot\mathbf{V}_n,$$

then sums the local category-specific maps over space and groups. The head is designed to preserve spatially dense, category-specific evidence rather than collapse it immediately into a global vector. On ImageNet, the reported gains are especially large for lightweight models, including 3.14\% top-1 for MobileNetV2 (0.5x), 5.3\% for DeiT‑Tiny (0.5x), and 1.9\% for gMLP‑Tiny (0.5x); on MS‑COCO, it reports 3.6\% mAP improvement for ResNet101 and 1.1\% for ViT‑Large (Li et al., 2024).

The term also appears institutionally in Huawei Noah’s Ark Lab. In a quantization study on DINO‑WM, the paper reports that W8A8 remains robust, group-wise weight quantization can stabilize low-bit rollouts, activation quantization granularity yields inconsistent benefits, and quantization sensitivity is highly asymmetric between encoder and predictor modules. A notable systems-level finding is that aggressive low-bit quantization can degrade the alignment between the planning objective and task success in world-model-based planning (Fu et al., 2 Feb 2026).

4. NOAH and NoAH in multimodal reasoning and affect recognition

In Video LLM evaluation, NOAH denotes “Benchmarking Narrative Prior driven Hallucination and Omission in Video LLMs.” The benchmark constructs 9,000 composite videos from 1,000 target videos by inserting semantically controlled clips at start, middle, or end positions, yielding over 60K evaluation samples across one captioning task and three QA tasks: Existence, Temporal, and Narrative. The study introduces caption-level and event-level hallucination and omission metrics—such as CHR, COR, EHR, EOR, and IEOR—and reports three main findings: most Video LLMs exhibit hallucinations and omissions driven by narrative priors, the error patterns vary with architecture and with insertion similarity and position, and reliance on narrative priors intensifies when fewer frames are sampled. The benchmark is presented as the first standardized evaluation of narrative prior-induced hallucination and omission in Video LLMs (Lee et al., 9 Nov 2025).

In a different multimodal context, NoAH denotes the negative class in ambivalence/hesitancy recognition: “No A/H present.” In the BAH dataset of the ABAW10 Ambivalence/Hesitancy Challenge, labels are video-level and binary, with $y=1$ for A/H present and $y=0$ for NoAH. The proposed ConflictAwareAH framework uses pre-trained video, audio, and text encoders and constructs pairwise conflict features

$$\mathbf{c}_{va}=|\mathbf{v}-\mathbf{a}|,\quad \mathbf{c}_{vt}=|\mathbf{v}-\mathbf{t}|,\quad \mathbf{c}_{at}=|\mathbf{a}-\mathbf{t}|,$$

treating large cross-modal differences as evidence for A/H and small differences as evidence for NoAH. Relative to text alone, the full multimodal model improves F1‑NoAH by +4.6 points and halves the class-performance gap; the reported performance reaches 0.694 Macro F1 on the labeled test split and 0.715 on the private leaderboard (Bekhouche et al., 16 Mar 2026).

These two usages are conceptually unrelated but structurally parallel: in both cases, the term is tied to failure of naive single-channel inference. In the Video LLM benchmark, the failure mode is narrative coherence overriding visual grounding; in ambivalence recognition, the failure mode is text-dominant inference overcalling A/H when cross-modal consistency would support NoAH.

5. NoAH as the Noah effect in anomalous diffusion

In stochastic-process theory, NoAH refers to the Noah effect in Mandelbrot’s sense: anomalous transport driven by fat‑tailed increment statistics. The decomposition used in recent work separates anomalous diffusion into Joseph, Noah, and Moses effects. In continuous time, the exponents are defined through the scaling of increment velocities: $$\langle |v(t)| \rangle \sim t^{\,M-\tfrac12}, \qquad \langle v^2(t) \rangle \sim t^{\,2L+2M-2},$$ while the Hurst exponent satisfies

$x_{t+1}=f(x_t)$0

Here $x_{t+1}=f(x_t)$1 quantifies long-range correlations, $x_{t+1}=f(x_t)$2 the Noah effect, and $x_{t+1}=f(x_t)$3 nonstationarity or aging (Aghion et al., 2020, Bera et al., 18 Dec 2025).

Empirically, the decomposition has been applied to systems spanning 12 orders of magnitude in length and 8 orders of magnitude in time. In that analysis, the Noah effect is identified with fat‑tailed probability density of increments. Many systems display $x_{t+1}=f(x_t)$4, indicating no substantial Noah effect, whereas clearer Noah signatures were reported for the searching black-winged kite and for some stork movement regimes, where rare long displacements dominate variance more strongly (Vilk et al., 2021).

The most recent theoretical extensions substantially enlarge the role of the Noah effect. In Variable Speed Generalized Lévy Walks (VGLWs), anomalous diffusion is decomposed into the same three effects, but the study finds that the Noah exponent $x_{t+1}=f(x_t)$5 is unbounded from above, unlike in previously studied Lévy-walk-type models. This indicates that the strength of the Noah effect in VGLWs can exceed the range usually associated with classical Lévy walks (Bera et al., 18 Dec 2025). In two-state random walks that alternate between a CTRW rest state and a standard Lévy walk motion state, the reported conclusion is equally striking: classical Lévy walks alone possess only the Joseph effect, but both Noah and Moses effects emerge from stochastic switching with the CTRW phase (Bera et al., 29 May 2026).

6. NOAH in optimization, serverless systems, and hypergraph generation

In underwater robotics, NOAH stands for Nauplius Optimisation for Autonomous Hydrodynamics. The algorithm is inspired by barnacle nauplii and integrates three mechanisms: current-aware drift, irreversible settlement, and colony-based communication. Agents evolve under a flow field $x_{t+1}=f(x_t)$6, a colony potential field $x_{t+1}=f(x_t)$7, and a settlement probability that depends on relative fitness, distance to colonies, flow shear, crowding, and energy. A central design feature is that once the settlement flag becomes $x_{t+1}=f(x_t)$8, the agent is permanently stationary. The validation study reports an 86% success rate for permanent anchoring scenarios (Ramesh et al., 17 Oct 2025).

In cloud systems, NOAH denotes a noncooperative online allocation heuristic for the serverless scheduling problem. The paper frames serverless as a joint allocation and per-event scheduling problem rather than ordinary VM allocation or request load balancing. For each function type, NOAH uses an M/M/c allocation model and the Erlang C waiting-time formula to choose the smallest number of logical instances $x_{t+1}=f(x_t)$9 satisfying

$y(v_t)=f(v_t)+\mathrm{GP}(v_t)$0

where $y(v_t)=f(v_t)+\mathrm{GP}(v_t)$1 is a function-specific waiting-time threshold. It then colocates allocations to reduce setup overhead and lets each host compare estimated queueing delay against setup time before spawning a new container. Simulation results show that, under high system utilization, NOAH contains instance churn more effectively than the default Apache OpenWhisk heuristic and the reviewed game-theoretic load balancer (Stein, 2018).

In hypergraph generation, NoAH is a stochastic model for attributed hypergraphs, paired with NoAHFit, a fitting procedure for a single observed hypergraph. The generator uses a core–fringe node hierarchy and forms each hyperedge as a seed-centered core group followed by fringe attachments, with probabilities given by multiplicative attribute-affinity matrices. Across nine datasets across four different domains, the paper reports that NoAH with NoAHFit reproduces structure–attribute interplay more accurately than eight baseline hypergraph generative models, obtaining the best average rank overall (1.5) across six metrics (Chun et al., 26 Sep 2025).

7. Disambiguation and recurring patterns

The cited literature supports four broad disambiguation rules. In Earth system science, “Noah” almost always denotes the Noah LSM / Noah‑MP family (Pelissier et al., 2020). In vision, “NOAH” denotes either Neural prOmpt seArcH or Non-glObal Attentive Head (Zhang et al., 2022, Li et al., 2024). In multimodal evaluation and affective computing, the term can denote either a benchmark name or the absence class “No A/H present” (Lee et al., 9 Nov 2025, Bekhouche et al., 16 Mar 2026). In stochastic-process theory, “Noah” denotes the Noah effect and its exponent $y(v_t)=f(v_t)+\mathrm{GP}(v_t)$2 (Bera et al., 18 Dec 2025).

A further institutional usage appears in Huawei Noah’s Ark Lab, where “Noah” is not an acronym for a model but part of a research-lab name. In the cited study, that institutional label is associated with an empirical map of post-training quantization behavior in world models, including the observation that encoder and predictor quantization sensitivities are highly asymmetric and that aggressive low-bit quantization can break the alignment between planning objective and task success (Fu et al., 2 Feb 2026).

Taken together, these usages show that NoAH/NOAH/Noah functions as a highly overloaded research term. Its meaning is therefore inseparable from domain context: land-surface physics, parameter-efficient tuning, classification heads, multimodal evaluation, anomalous-diffusion theory, serverless scheduling, underwater robotics, and attributed hypergraph generation each assign the name a different technical identity.