LEAF: Multifaceted Research in AI & Plant Science

• LEAF is a term used to denote a family of distinct research objects across AI benchmarks and plant science, emphasizing modular design and domain realism.
• It spans disciplines including federated learning, fog computing, audio processing, and medical imaging, often with rigorous performance metrics and energy optimization.
• In plant science, leaf models analyze geometric morphogenesis, hydraulics, and sensing, achieving high accuracies and validating theories through detailed empirical studies.

In contemporary technical literature, LEAF denotes a family of unrelated but highly specific research objects rather than a single concept. On arXiv, the name is used for a federated-learning benchmark, a fog-computing simulator, a learnable audio frontend, a medical image segmentation framework, a convex-optimization method, a chart question-answering architecture, a Text-to-SQL system, and several domain-specific systems in agriculture and plant science (Caldas et al., 2018, Wiesner et al., 2021, Schlüter et al., 2022, Huang et al., 24 Jul 2025, Nguyen et al., 8 Jun 2026, Chaudhry et al., 2019, Tan et al., 10 May 2026). In parallel, the ordinary noun leaf remains central in plant morphogenesis, hydraulics, sensing, and phenotyping, where it names the biological object rather than an acronym (Young, 2010, Luo et al., 2021).

1. Disambiguation across research domains

Name	Domain	Representative role
LEAF	Federated learning	Modular benchmarking framework (Caldas et al., 2018)
LEAF	Fog computing	Large Energy-Aware Fog Computing simulator (Wiesner et al., 2021)
LEAF	Audio classification	Learnable audio frontend (Schlüter et al., 2022)
LEAF	Medical imaging	Latent diffusion segmentation framework (Huang et al., 24 Jul 2025)
LEAF-SQL	Text-to-SQL	Level-wise skeleton search (Tan et al., 10 May 2026)
LEAF-QA / LEAF-Net	Figure QA	Dataset and “Locate, Encode & Attend” architecture (Chaudhry et al., 2019)

Several later systems extend the same naming pattern with domain-specific compounds, including LeafInst for forestry leaf instance segmentation, ReLeaf for cross-domain leaf segmentation benchmarking, NeuraLeaf for neural parametric 3D leaf modeling, and Leafeon for mmWave leaf-water-content sensing (Luo et al., 4 Mar 2026, Martinko et al., 5 May 2026, Yang et al., 17 Jul 2025, Cardamis et al., 2024). This distribution suggests that LEAF functions in the literature both as an acronymic brand for algorithmic frameworks and as a biologically literal label in plant-focused work.

2. Leaf form, growth, and hydraulics in plant science

A foundational use of leaf concerns geometric morphogenesis. The growth-algorithm model of leaf shape represents a leaf as a two-dimensional expanding sheet, discretized into rows indexed by $j=1,\dots,N$, with axial growth along the midrib and position-dependent lateral elongation. In that model, axial growth follows $y(t+1)=(1+g_y)y(t)$, row length evolves as $L_j(t+1)=(1+g_L(j))L_j(t)$, and vein-centered attenuation is introduced through $f_j=\exp(-a|d_j|^n)$, so that lobes and serrations emerge from spatial modulation around veins (Young, 2010). The same framework states that varying $R=g_y/g_{L0}$, the location of maximum lateral growth $j_a$, the exponent $k$, the slope field $s(j)$, and the attenuation parameters $a,n$ is sufficient to traverse broad-to-narrow, entire-to-lobed, and rounded-to-pointed leaf series (Young, 2010).

A distinct mathematical account argues that leaf growth is conformal. There, the contour at one time is mapped to a later contour by a conformal transformation, and the induced interior displacement field is compared against particle-image-velocimetry measurements. The predicted displacement field agrees with the measured field to 92%, reaches 97% in the best specimen, and yields pixelwise correlations above 85% across the leaf surface (Alim et al., 2016). The same study reports that area-growth correlations are lower in raw form, at 40–60%, but rise to 60–80% after spatial averaging, and interprets the results as evidence for locally isotropic growth at the supracellular scale (Alim et al., 2016).

Leaf hydraulics introduces a third, mechanistic usage. A spatially explicit capacitive model treats the hydraulic network as a distributed RC circuit with xylem resistances, stomatal sinks, and storage capacitances, and derives a continuum equation for the monocot one-dimensional case (Luo et al., 2021). In the excised-leaf setting, both the average xylem potential and transpiration decay exponentially with time constant

$\tau = C(R_c + R_a),$

showing that storage capacitance and the resistances linking storage and atmosphere govern short-term robustness to intermittent drought (Luo et al., 2021). The same model predicts a monotonic base-to-tip decline in water potential, identifies distal regions as disproportionately disadvantaged under stress, and explicitly notes that lumped models reproduce average dynamics but miss distal vulnerability and spatial heterogeneity (Luo et al., 2021).

3. Sensing, manipulation, and reconstruction of biological leaves

Leaf-focused engineering systems increasingly target direct measurement, manipulation, and geometric reconstruction. In robotics, an integrated actuation–perception framework for leaf retrieval combines an Intel RealSense D435i, Open3D point-cloud processing, a Kinova Gen-2 arm, and a custom cutting end-effector. Offline detection on avocado point clouds reached 80.0% average indoors and 79.8% average outdoors, mean localization errors were 8.28 mm, 14.38 mm, and 15.54 mm along the three reported axes, and integrated indoor retrieval over 46 trials produced 21 successful cuts, of which 4 were described as clean cuts suitable for stem water potential measurement (Campbell et al., 2022). The paper also reports perception latency between 0.5 s and 11.0 s per cloud and overall retrieval times from 6.1 s to 62.5 s (Campbell et al., 2022).

For non-contact physiological sensing, Leafeon uses a COTS TI AWR1843Boost mmWave FMCW radar, electronic beam steering from −10° to +10° in 2° steps, and a learned fusion network, LM-Net, to estimate modified relative water content $y(t+1)=(1+g_y)y(t)$0 (Cardamis et al., 2024). The reported in-lab mean absolute error is 3.17% for Avocado, 3.41% for Rubra, and 5.87% for Bull bay, and the abstract states an MAE reduction of up to 55.7% relative to state-of-the-art approaches (Cardamis et al., 2024). The same work attributes the gain to multi-perspective sensing of angle-dependent surface and volumetric scattering rather than single-angle RSS alone (Cardamis et al., 2024).

NeuraLeaf addresses 3D geometry rather than physiology. It disentangles a leaf into a planar 2D base shape represented by a neural signed distance function and a 3D deformation generated by a skeleton-free blend-skinning model with up to 1000 control points (Yang et al., 17 Jul 2025). The accompanying DeformLeaf dataset contains approximately 300 pairs of flattened 2D base shapes and 3D deformed scans, and on single-leaf reconstruction the model reports 2.1 mm Chamfer error and 0.973 normal consistency, outperforming PCA, B-spline, and a human-style neural parametric baseline (Yang et al., 17 Jul 2025). This suggests that the biological near-planarity of flattened leaves can be exploited as a strong inductive bias for learned 3D reconstruction.

4. Leaf recognition, segmentation, and phenotyping in computer vision

Controlled-environment leaf classification has been a persistent benchmark problem. One multimodal pipeline combines refined color images, vein images, xy-projection histograms, handcrafted shape and texture features, Fourier descriptors, neural encoders, and an RBF-kernel SVM, reaching 99.58% ± 0.31% test accuracy on Flavia under random stratified 10-fold cross-validation (Quach et al., 2021). A different nine-layer CNN with transfer learning and aggressive augmentation reports 99.81 ± 0.26% on Flavia under the “10 × All” protocol with fixed-rotation test-time augmentation and 99.40 ± 0.09% on Foliage (Wick et al., 2017). Both works are explicit about the controlled-background regime and the risk of overfitting on small leaf datasets (Quach et al., 2021, Wick et al., 2017).

Recent work shifts from classification to leaf-level instance segmentation under field conditions. LeafInst introduces the Poplar-leaf benchmark with 1,202 branch-scale RGB images and 19,876 pixel-level annotated leaf instances, and pairs it with a unified anchor-free segmentation model built from AFPN, DASP, DARH, and TCFU (Luo et al., 4 Mar 2026). On Poplar-leaf validation it reports 68.4 seg/mAP, and on the test split 70.0 seg/mAP, while zero-shot transfer to PhenoBench yields 52.7 box mAP and 50.2 seg/mAP (Luo et al., 4 Mar 2026). The paper explicitly positions the benchmark around scale variation, illumination changes, and irregular morphology in open-field forestry scenes (Luo et al., 4 Mar 2026).

ReLeaf systematizes cross-domain evaluation. It aggregates 17,946 single-plant patches from LSC, Komatsuna, GrowliFlower, and PhenoBench, introduces the CropAndWeedAndLeaf benchmark with 345 patches across 23 species, and identifies a YOLO26 Medium at 768² px configuration as the best trade-off for real-world precision-agriculture tasks (Martinko et al., 5 May 2026). A model trained on all four existing datasets reaches a mean mAP50–95 of 83.9% across their test sets and 40.2% on the new benchmark, while average cross-domain transfer from laboratory to real imagery is only 17.7%, versus 29.8% for real-to-lab transfer (Martinko et al., 5 May 2026). A common misconception in this area is that high in-domain laboratory accuracy transfers cleanly to field deployment; the benchmark results argue the opposite (Martinko et al., 5 May 2026).

5. LEAF as benchmark, simulator, frontend, and energy optimizer

In machine learning infrastructure, LEAF is best known as a federated benchmark. The benchmark defines a modular framework comprising open-source federated datasets, a rigorous evaluation framework, and reference implementations, with datasets such as FEMNIST, Shakespeare, CelebA, Sentiment140, Reddit, and Synthetic organized by natural client partitions rather than IID reshuffling (Caldas et al., 2018). Its evaluation emphasizes not only micro- and macro-averaged performance but also client-level heterogeneity and fairness measures such as the 10th-percentile accuracy, alongside systems measures such as rounds, bytes, and participation fraction (Caldas et al., 2018). This framing helped establish client-partition realism as a benchmark requirement rather than an implementation detail.

LEAF also names a simulator for fog and edge systems. Large Energy-Aware Fog Computing models infrastructure as a graph $y(t+1)=(1+g_y)y(t)$1 and applications as a DAG $y(t+1)=(1+g_y)y(t)$2, combines analytical and discrete-event modeling, and uses component-wise power models such as

$y(t+1)=(1+g_y)y(t)$3

for compute nodes and $y(t+1)=(1+g_y)y(t)$4 for links (Wiesner et al., 2021). In the smart-city evaluation, a 24-hour scenario with approximately 46,500 taxis executing approximately 330,000 tasks finished in <50 seconds on a single 1.4 GHz Intel Core i5 core, and a Fog 4 deployment saved approximately 2300 Wh compared with Fog 6 (Wiesner et al., 2021). The paper presents LEAF not merely as a simulator of compute nodes, but as a holistic energy model spanning compute, network, and applications (Wiesner et al., 2021).

In signal processing, LEAF is a learnable audio frontend based on Gabor filters, Gaussian pooling, and PCEN, whereas EfficientLEAF replaces PCEN and uses inhomogeneous kernel sizes and strides for efficiency (Schlüter et al., 2022). The central empirical claim is negative as well as positive: EfficientLEAF matches LEAF at 3% of the cost, but both fail to consistently outperform a fixed mel filterbank across six audio classification tasks (Schlüter et al., 2022). In mobile augmented reality, LEAF names an optimization algorithm that jointly adapts CPU frequency, model size, and radio allocation to minimize per-frame energy under latency and accuracy constraints, while the companion AIO mechanism regulates image offloading frequency (Wang et al., 2022). The analytical model reports validation mean absolute percentage errors of 6.1% ± 3.4%, 7.6% ± 4.9%, 6.9% ± 3.9%, and 3.7% ± 2.6% across major configuration axes, and LEAF reduces per-frame energy by up to 40% and latency by 35% relative to the cited baseline FACT at $y(t+1)=(1+g_y)y(t)$5 Mbps (Wang et al., 2022).

6. LEAF in contemporary AI and mathematical theory

Several recent AI systems adopt LEAF for task-specific structured reasoning. LEAF-QA contributes 250,000 densely annotated chart images and approximately 2 million question-answer pairs, while LEAF-Net implements “Locate, Encode & Attend” through chart element localization, question/answer encoding by chart-localized text, and stacked attention (Chaudhry et al., 2019). The oracle version reports 67.42% overall on LEAF-QA test, 72.72% on DVQA Test-Familiar, and 81.15% validation accuracy on FigureQA (Chaudhry et al., 2019). LEAF-SQL treats SQL skeleton prediction as a coarse-to-fine tree search over Base, Expanded, and Detailed skeletons, using a Skeleton Formulation Agent and a Skeleton Evaluation Agent, and achieves 71.6% execution accuracy on the hidden BIRD test set (Tan et al., 10 May 2026).

In medical imaging, LEAF stands for Latent Diffusion with Efficient Encoder Distillation for Aligned Features. The framework replaces $y(t+1)=(1+g_y)y(t)$6-prediction with direct latent segmentation-map prediction and adds cosine-based feature alignment to a pretrained transformer encoder, while keeping inference-time architecture and parameter count unchanged (Huang et al., 24 Jul 2025). Reported results include 89.5 / 81.5 Dice/IoU on REFUGE2, 95.2 / 90.9 on CVC-ClinicDB, 80.2 / 71.0 on QaTa-Covid19, and 90.5 / 84.1 on ISIC 2018, with lower variance under $y(t+1)=(1+g_y)y(t)$7-prediction than under $y(t+1)=(1+g_y)y(t)$8-prediction (Huang et al., 24 Jul 2025). In continual information extraction, LEAF denotes a mixture-of-experts architecture with LoRA experts, semantic-aware routing, label-description contrastive learning, and two-level distillation for few-shot continual event detection; it reaches 54.9% F1 on MAVEN task 5 in the 4-way 10-shot setting and 61.6% on ACE-2005 task 5 in the 2-way 10-shot setting (Dao et al., 29 Sep 2025).

Optimization provides another meaning. LEAF, as a learning-enabled ADMM framework, learns a scalar Moreau envelope with an input convex neural network and uses its gradient in MEL-ADMM and sMEL-ADMM updates (Nguyen et al., 8 Jun 2026). The framework preserves convexity and smoothness, provides convergence guarantees for the surrogate problem, and reports microgrid runtimes of 0.5 ± 0.1 ms for sMEL-ADMM versus 11.5 ± 0.4 ms for IPOPT at $y(t+1)=(1+g_y)y(t)$9 and 1% optimality gap, as well as 1.3 ± 0.2 ms versus 26.1 ± 0.2 ms at $L_j(t+1)=(1+g_L(j))L_j(t)$0 and 0.01% gap (Nguyen et al., 8 Jun 2026). A final, non-acronymic use appears in graph theory, where the “leaf function” $L_j(t+1)=(1+g_L(j))L_j(t)$1 of Penrose P2 graphs denotes the maximum number of leaves in an induced subtree of order $L_j(t+1)=(1+g_L(j))L_j(t)$2, and the cited work derives exact and recursive formulae while constructing infinite families of fully leafed caterpillar subtrees (Porrier et al., 2023).

LEAF therefore functions less as a unitary concept than as a recurrent naming convention for technically ambitious frameworks whose commonality is usually structural rather than disciplinary. In some cases, the name signals learnability or efficiency; in others, it marks domain realism, structured search, or direct engagement with biological leaves. The term’s encyclopedic significance lies precisely in this polysemy: it is a compact label under which arXiv research has assembled benchmark design, geometric morphogenesis, hydraulic transport, leaf phenotyping, signal processing, multimodal reasoning, and optimization theory (Caldas et al., 2018, Luo et al., 2021, Schlüter et al., 2022, Nguyen et al., 8 Jun 2026).