Hierarchical Soft Mixture-of-Experts (HoME)

• HoME is a hierarchical model family that recursively combines expert outputs using soft probabilistic splits for continuous, end-to-end trainable predictions.
• It employs probabilistic gating and Bayesian estimation techniques to manage expert specialization, reduce overfitting, and prevent expert collapse.
• Applications of HoME span adversarial generation, multi-task recommendation, and 3D segmentation, showcasing its versatility in both academic and industrial settings.

Searching arXiv for recent and foundational papers on Hierarchical Soft Mixture-of-Experts. Hierarchical Soft Mixture-of-Experts (HoME) denotes a family of models in which expert outputs are combined through multi-level, differentiable routing rather than a single flat gate. In the classical hierarchical mixture of experts formulation, internal nodes implement soft probabilistic splits and leaves implement local experts, so the overall prediction is a path-weighted mixture over leaves. In more recent literature, the same organizing principle appears in tree-structured conditional density models, hierarchical routing models for regression, adversarial generators, industrial multi-task recommenders, graph models for high-level synthesis, and token-routing modules for long-context 3D segmentation. Across these variants, the defining property is not a single canonical topology but a recursive or multi-level decomposition in which routing remains soft and end-to-end trainable (Bishop et al., 2012, Wang et al., 2024, Płotka et al., 8 Jul 2025).

1. Core architectural idea

The foundational HoME/HME architecture is a binary tree with two node types: gating nodes at internal positions and expert nodes at the leaves. For regression, a classical expert is a Gaussian conditional model,

$$p(t \mid x, W_j, T_j) = \mathcal{N}(t \mid W_j x,\, T_j^{-1} I),$$

while each gating node uses a binary latent switch variable with logistic probability

$$p(z_i \mid x, v_i) = \sigma(v_i^T x)^{z_i}\,[1-\sigma(v_i^T x)]^{1-z_i}, \qquad \sigma(a)=\frac{1}{1+\exp(-a)}.$$

After marginalizing the gating variables, the conditional distribution becomes a mixture over experts whose coefficient for expert $j$ is the product of gate probabilities along the path to that expert (Bishop et al., 2012).

An equivalent recursive view, common in soft decision tree treatments, writes the subtree output at node $m$ as

$$y_m(x) = \begin{cases} \alpha_m(x) y_{ml}(x) + (1-\alpha_m(x)) y_{mr}(x), & \text{if } m \text{ is internal}\ f_m(x), & \text{if } m \text{ is leaf} \end{cases}$$

with

$\alpha_m(x) = \text{sigmoid}(w_m^T x).$

Under this recursion, every path contributes proportionally to its routing probability; the model therefore defines a soft hierarchical partitioning of the input space rather than a hard partition (İrsoy et al., 2018).

The adjective “soft” is central. HoME does not usually commit to one branch deterministically. Instead, neighboring regions overlap, a single input can assign nonzero mass to multiple experts, and the resulting predictor is continuous. This differentiable routing mechanism is what makes the architecture compatible with gradient-based optimization and distinguishes it from CART-style trees or hard expert selection (Bishop et al., 2012).

2. Probabilistic formulations and estimation

In its original formulation, HME was trained by maximum likelihood, typically using EM. That training regime is described as prone to severe overfitting, susceptible to likelihood singularities when components collapse onto points, and lacking a principled mechanism for choosing tree size or topology. The Bayesian treatment in “Bayesian Hierarchical Mixtures of Experts” replaces point estimation with priors over expert weights, gating weights, and precisions, and optimizes a rigorous variational lower bound on the marginal probability of the data. The resulting lower bound is then used for model order selection (Bishop et al., 2012).

The Bayesian model places Gaussian priors on gating and expert weights and Gamma priors on hyperparameters and expert precisions, with broad hyperparameters $a=10^{-2}$ and $b=10^{-4}$. Because the sigmoid gates break conjugacy, the variational treatment introduces the Jaakkola–Jordan bound

$$\sigma(x) \ge F(x,\xi) = \sigma(\xi)\exp\left\{\frac{x-\xi}{2} - \lambda(\xi)(x^2-\xi^2)\right\}, \qquad \lambda(\xi)=\frac{\tanh(\xi/2)}{4\xi},$$

which restores an exponential-quadratic form and yields tractable mean-field updates. This turns evidence maximization into a practical training criterion and provides an explicit Bayesian complexity control mechanism (Bishop et al., 2012).

A distinct but closely related line is “Hierarchical Routing Mixture of Experts” (HRME), which preserves the binary-tree, soft-routing structure but changes what internal nodes do. In HRME, non-leaf nodes are classifier experts, leaf nodes are regression experts, and routing is learned from the joint structure of input and output. Internal classifiers are trained using thresholded target labels, the threshold is learned by grid search, and leaf responsibilities are products of routing probabilities along root-to-leaf paths. Training proceeds through a recursive EM procedure that grows the tree top-down and keeps a split only when the objective improves (Zhao et al., 2019).

These formulations share a common statistical premise: multimodal conditional structure is easier to model by routing examples into simpler local regions than by fitting a single global regressor. What differs is whether the hierarchy is interpreted primarily as a conditional density model, a soft decision tree, or a routing mechanism over jointly structured $(x,y)$ space.

3. Routing behavior, regularization, and specialization

Because HoME models become exponentially expressive as depth increases, regularization is not incidental. “Dropout Regularization in Hierarchical Mixture of Experts” argues that standard unitwise dropout is structurally mismatched to a hierarchy, since parent gates determine how descendant subtrees contribute. The proposed alternative samples a Bernoulli variable at each internal node and, when the node is dropped, routes entirely through the right subtree:

$$p(z_i \mid x, v_i) = \sigma(v_i^T x)^{z_i}\,[1-\sigma(v_i^T x)]^{1-z_i}, \qquad \sigma(a)=\frac{1}{1+\exp(-a)}.$$0

The method intentionally always drops the left subtree, so the model can be interpreted as an ensemble over different tree complexities. On synthetic regression, MNIST, and CIFAR-10, this tree-faithful dropout is reported to reduce overfitting, improve generalization, and produce smoother fits, especially for deep trees (İrsoy et al., 2018).

A second regularization theme is explicit encouragement of expert diversity. In “Adversarial Mixture Of Experts with Category Hierarchy Soft Constraint,” the inference gate uses only sub-category IDs from the query side, while a constraint gate uses the corresponding top-category ID. A hierarchical soft constraint penalizes differences between the inference and constraint gate distributions on the top-$$p(z_i \mid x, v_i) = \sigma(v_i^T x)^{z_i}\,[1-\sigma(v_i^T x)]^{1-z_i}, \qquad \sigma(a)=\frac{1}{1+\exp(-a)}.$$1 active experts, so sibling sub-categories are encouraged to choose similar expert mixtures without being forced to be identical. An adversarial regularizer then samples inactive experts and pushes them to disagree with active experts, increasing expert diversity and reducing collapse (Xiao et al., 2020).

The reported empirical pattern is that hierarchical soft sharing is especially helpful when some categories are small or sparse. On the in-house JD search dataset, the paper reports AUC values of $$p(z_i \mid x, v_i) = \sigma(v_i^T x)^{z_i}\,[1-\sigma(v_i^T x)]^{1-z_i}, \qquad \sigma(a)=\frac{1}{1+\exp(-a)}.$$2 for DNN, $$p(z_i \mid x, v_i) = \sigma(v_i^T x)^{z_i}\,[1-\sigma(v_i^T x)]^{1-z_i}, \qquad \sigma(a)=\frac{1}{1+\exp(-a)}.$$3 for MoE, $$p(z_i \mid x, v_i) = \sigma(v_i^T x)^{z_i}\,[1-\sigma(v_i^T x)]^{1-z_i}, \qquad \sigma(a)=\frac{1}{1+\exp(-a)}.$$4 for Adv-MoE, $$p(z_i \mid x, v_i) = \sigma(v_i^T x)^{z_i}\,[1-\sigma(v_i^T x)]^{1-z_i}, \qquad \sigma(a)=\frac{1}{1+\exp(-a)}.$$5 for HSC-MoE, and $$p(z_i \mid x, v_i) = \sigma(v_i^T x)^{z_i}\,[1-\sigma(v_i^T x)]^{1-z_i}, \qquad \sigma(a)=\frac{1}{1+\exp(-a)}.$$6 for Adv-HSC-MoE; on the Amazon review dataset, the corresponding values are $$p(z_i \mid x, v_i) = \sigma(v_i^T x)^{z_i}\,[1-\sigma(v_i^T x)]^{1-z_i}, \qquad \sigma(a)=\frac{1}{1+\exp(-a)}.$$7, $$p(z_i \mid x, v_i) = \sigma(v_i^T x)^{z_i}\,[1-\sigma(v_i^T x)]^{1-z_i}, \qquad \sigma(a)=\frac{1}{1+\exp(-a)}.$$8, $$p(z_i \mid x, v_i) = \sigma(v_i^T x)^{z_i}\,[1-\sigma(v_i^T x)]^{1-z_i}, \qquad \sigma(a)=\frac{1}{1+\exp(-a)}.$$9, $j$0, and $j$1. The paper also reports cleaner category-wise separation in t-SNE projections of gate outputs under the hierarchical and adversarial variants (Xiao et al., 2020).

More broadly, these results establish a recurrent HoME principle: hierarchical routing is often coupled with explicit mechanisms to prevent expert coalescence, expert monopoly, or poorly conditioned sharing.

4. Application-specific realizations

HoME has been instantiated in multiple settings that preserve soft hierarchical routing while redefining what counts as an “expert.”

Setting	Hierarchical structure	Reported outcome
Adversarial generation	Tree of gating nodes and local generators	High quality and diversity on MNIST, FashionMNIST, UTZap50K, Oxford Flowers, and CelebA
E-commerce ranking	Sub-category inference gate plus top-category constraint gate	Adv-HSC-MoE reports AUC $j$2 on JD and $j$3 on Amazon
High-level synthesis	Low-level MoE at node, block, and graph granularities plus high-level MoE aggregation	Offline MSE $j$4; online geometric mean speedup $j$5
3D medical segmentation	Grouped local SMoE followed by global SMoE refinement inside a Mamba backbone	PANORAMA mDSC $j$6; pretrained mDSC $j$7

In “Hierarchical Mixtures of Generators for Adversarial Learning,” the proposed Hierarchical Mixture of Generators (HMoG) is explicitly presented as a GAN analogue of Hierarchical Soft Mixture-of-Experts. Internal nodes compute soft routing probabilities over the latent code $j$8, leaves are local generators, and the recursive output at an internal node is a weighted sum of its children. Because the gates are sigmoidal or softmax-based, the whole generator is a continuous convex combination of leaf outputs, and the resulting tree can be trained jointly by backpropagation under a Wasserstein GAN objective. The paper reports experiments on five image data sets—MNIST, FashionMNIST, UTZap50K, Oxford Flowers, and CelebA—and states that the learned hierarchical structure also leads to knowledge extraction (Ahmetoğlu et al., 2019).

In “Hierarchical Mixture of Experts: Generalizable Learning for High-Level Synthesis,” the hierarchy is not a classical binary tree but a two-level MoE over representation granularities. The low-level MoEs operate at node, basic-block, and graph levels, and a high-level MoE learns to aggregate those three views for final prediction. The main reported result is an offline MSE of $j$9, compared with HARP’s $m$0, and an online geometric mean speedup of $m$1, compared with HARP’s $m$2. The paper also states that the average speedup improvement is about $m$3 higher than the baseline (Li et al., 2024).

In “Mamba Goes HoME,” HoME becomes a two-level token-routing layer for long-context 3D medical image segmentation. A first-stage local SMoE partitions the token sequence into local groups and routes grouped representations to local experts; a second-stage global SMoE then refines and fuses information across groups. Integrated into a Mamba-based U-shaped network, this design is reported to achieve strong results across CT, MRI, and ultrasound datasets, including PANORAMA, AMOS, FeTA 2022, and MVSeg 2023 (Płotka et al., 8 Jul 2025).

5. Contemporary architectural variants and gating functions

Recent HoME work often departs from the strict “binary tree of local regressors” pattern while keeping a hierarchical, soft routing logic. “HoME: Hierarchy of Multi-Gate Experts for Multi-Task Learning at Kuaishou” is a large-scale industrial multi-task recommender in which the hierarchy is organized by semantic task groups rather than by recursive spatial partition. The paper identifies three anomalies in standard MoE systems—expert collapse, expert degradation, and expert underfitting—and addresses them through expert normalization plus Swish, a hierarchy mask mechanism, and feature-gate plus self-gate mechanisms. The resulting structure has at least two levels of expert abstraction: a meta level with shared, interaction-category, and watch-time-category experts, and a task level with task-specific experts. The paper reports an average offline improvement over MMoE of $m$4 GAUC, together with online gains such as $m$5 Average Play-time on Kuaishou Single Page and $m$6 Play-time on Kuaishou Double Page (Wang et al., 2024).

A different modernization concerns the gating function itself. “On Expert Estimation in Hierarchical Mixture of Experts: Beyond Softmax Gating Functions” keeps the classic two-level HMoE structure but compares three gate combinations: Softmax–Softmax (SS), Softmax–Laplace (SL), and Laplace–Laplace (LL). The paper argues theoretically that Softmax gating induces undesirable parameter interactions, while Laplace gating removes part or all of those interactions depending on the level at which it is used. The density estimation rate remains near-parametric, $m$7, but the expert-parameter convergence story changes: for over-specified experts, SS and SL depend on higher-order algebraic multiplicities, whereas LL yields the faster rate $m$8 for the expert parameter $m$9. Empirically, the paper reports that Laplace–Laplace is generally best on multimodal clinical prediction and image classification, while Softmax–Laplace performs best on some latent-domain discovery tasks (Nguyen et al., 2024).

These contemporary variants extend HoME in two directions. First, hierarchy is increasingly used to encode semantic, representational, or token-level structure rather than only tree depth. Second, gate design has become an object of theoretical analysis in its own right, rather than a fixed softmax default.

6. Terminology, scope, and recurrent misconceptions

The literature uses several nearby names—HME, HMoE, HoME, HRME, HMoG, and Soft-MoE—and they are not interchangeable. Classical HME and many HMoE formulations refer to a tree of probabilistic gates and local experts. HMoG is explicitly a GAN analogue in which leaves are local generators rather than regressors or classifiers. HRME is a hierarchical routing model for regression in which internal nodes are classifier experts trained from thresholded targets. Kuaishou’s HoME expands the acronym to “Hierarchy of Multi-Gate Experts” and uses semantic task-group hierarchy rather than a binary decision tree (Ahmetoğlu et al., 2019, Zhao et al., 2019, Wang et al., 2024).

A common misconception is that any soft mixture of experts is automatically hierarchical. “Robust Exploration in Directed Controller Synthesis via Reinforcement Learning with Soft Mixture-of-Experts” is useful precisely because it clarifies the distinction. Its proposed method is a flat Soft-MoE with a prior-confidence gating mechanism: every expert is scored once, mixture weights are computed once at the initial state $$y_m(x) = \begin{cases} \alpha_m(x) y_{ml}(x) + (1-\alpha_m(x)) y_{mr}(x), & \text{if } m \text{ is internal}\ f_m(x), & \text{if } m \text{ is leaf} \end{cases}$$0, and the final policy is a single weighted sum of all experts. The paper states explicitly that there is no multi-level routing tree, no recursive expert selection, and no nested gating architecture. It is therefore better interpreted as a flat Soft-MoE baseline or building block, not a true hierarchical MoE (Ubukata et al., 22 Feb 2026).

Another misconception is that hierarchy always implies hard routing or discrete specialization. Across the works summarized here, the defining mechanism is usually the opposite: sigmoid, softmax, or related continuous gates keep routing differentiable, preserve nonzero contribution from multiple experts, and allow gradient-based joint training. This is true in classical Bayesian HME, in tree-faithful dropout for HMoE, in HMoG for adversarial generation, in industrial multi-task HoME, and in Mamba-HoME for 3D segmentation (Bishop et al., 2012, İrsoy et al., 2018, Płotka et al., 8 Jul 2025).

Taken together, these strands suggest that HoME is best viewed not as a single frozen architecture but as a modeling principle: organize experts across levels, keep routing soft, and use the hierarchy to encode the structure—statistical, semantic, or spatial—that a flat mixture would obscure.