Mixture-of-Experts Models

CNN- and Transformer-based architectures have achieved strong performance in medical image segmentation, but CNNs are limited in modeling long-range dependencies, while Transformers often suffer from quadratic computational and memory complexity. State space models, especially Mamba-based networks, offer an efficient alternative with linear sequence complexity. However, existing Mamba segmentation models still face two limitations: pixel-wise directional scanning can disrupt local 2D spatial structure, and simple summation-based fusion of scan directions cannot adapt well to diverse object sizes, shapes, and boundaries. To address these issues, we propose \textit{Patch-MoE Mamba}, a patch-ordered mixture-of-experts state space architecture for medical image segmentation. It introduces a hierarchical patch-ordered scanning mechanism that preserves local spatial neighborhoods while capturing multi-scale context, and an MoE-based directional fusion module that adaptively combines multiple Mamba scanner outputs using four directional experts, a learnable concatenation expert, and residual directional aggregation. Experiments on five public polyp segmentation benchmarks and the ISIC 2017/2018 skin lesion segmentation datasets demonstrate the effectiveness and generality of Patch-MoE Mamba.

Mixture-of-Experts (MoE) architectures enable efficient model scaling, yet expert routing behavior across underrepresented languages remains poorly understood. We analyze routing dynamics in two architecturally distinct MoE models -- a pure Transformer (Qwen3-30B-A3B) and a hybrid Mamba-Transformer (Nemotron-3-Nano-30B-A3B) -- using Hebrew as a morphologically rich, low-resource testbed. Both pre-trained models exhibit \emph{deep-layer routing collapse}: usage entropy drops sharply in final layers and tokens concentrate on a narrow expert subset, a pattern largely absent for English. Continual pre-training (CPT) on balanced bilingual data substantially corrects this imbalance, increasing entropy and shifting routing toward shared, language-agnostic experts; supervised fine-tuning (SFT) alone achieves less complete correction. Extending the analysis to Japanese reveals quantitatively consistent collapse signatures, providing cross-linguistic evidence that the phenomenon is a systematic consequence of pre-training underrepresentation rather than any language-intrinsic property. Routing improvements correlate with consistent downstream benchmark gains, positioning routing entropy and expert specialization as principled diagnostics for multilingual capacity in MoE systems.

Large-scale pre-trained Vision-Language models (VLMs), such as CLIP, exhibit strong zero-shot generalization, yet remain highly vulnerable to imperceptible adversarial perturbations, raising serious safety concerns for open-world deployment. To enhance robustness without requiring downstream task-specific retraining, we propose TAME, a novel test-time defense. Building upon our prior Test-Time Adversarial Prompt Tuning (TAPT), TAME introduces an architectural reformulation by replacing TAPT's single adaptive prompt with an input-conditioned Mixture-of-Experts (MoE) framework, enabling more expressive and adaptive defense. Specifically, TAME maintains a bank of learnable expert prompts and employs an input-dependent routing mechanism to aggregate a customized prompt mixture for each unlabeled test sample at inference time. This test-time defense mechanism is driven by three unsupervised objectives: (1) multi-view prediction entropy minimization, (2) layer-wise alignment of visual token statistics to precomputed clean and adversarial reference distributions, and (3) MoE regularization for balanced expert utilization and prompt diversity. We evaluated TAME on 11 benchmark datasets, including ImageNet and 10 additional zero-shot datasets. The results show that TAME improves the zero-shot adversarial robustness of the original CLIP by at least 49.1% under AutoAttack while largely preserving generalization on clean samples. TAME also consistently outperforms existing adversarial prompt tuning methods across multiple prompt designs, yielding an average robustness gain of at least 30.2%.

Class-incremental learning (CIL) requires models to learn new classes sequentially while preserving prior knowledge. Recently, approaches that combine pre-trained models with mixture-of-experts (MoE) have received increasing attention in CIL: they typically expand experts during learning and employ a router to assign weights across experts. However, existing MoE methods often overlook routing drift induced by expert expansion. Once new experts are introduced, the router may reassign samples from earlier classes to newly added experts, thereby perturbing previously established expert compositions and causing interference even when old experts remain frozen. We argue that expandable MoE in CIL requires two complementary properties: stable old-class routing for knowledge preservation and sufficient capacity utilization for new-class adaptation. To this end, we propose Stable Routing for MoE (StaR-MoE), a routing-level framework for expandable MoE in CIL. By incorporating sensitivity-aware routing alignment, StaR-MoE aligns current old-class routing behavior with historical routing distributions through sensitivity-guided constraints. Complementarily, StaR-MoE introduces asymmetric capacity regularization to encourage effective utilization of the expanded expert pool without compromising class-specific routing specialization. Extensive experiments across four standard CIL benchmarks demonstrate that StaR-MoE consistently improves both average and last accuracy over state-of-the-art methods, highlighting the importance of stable routing.

Model merging enables the reuse of fine-tuned models without joint training or access to original data. Dynamic merging further improves flexibility by selectively activating task-relevant parameters and efficiently composing experts across multiple tasks. However, existing dynamic methods either maintain a full shared model with tiny experts or allocate excessive capacity to experts, leading to suboptimal accuracy--efficiency trade-offs. To address this, we propose DiDi-Merging, a slim dynamic merging framework that leverages differentiable rank allocation to balance shared and expert parameters. By formulating parameter budgeting as differentiable rank optimization in low-rank modules and introducing a data-free refinement step to recover task fidelity, DiDi-Merging matches prior dynamic baselines at only 1.24x the parameters of a single fine-tuned model and surpasses them at 1.4x, substantially more compact than methods requiring > 2x storage. DiDi-Merging applies across vision, language, and multimodal tasks.

Heterogeneous LoRA-rank methods address system heterogeneity in federated fine-tuning of foundation models by assigning client-specific ranks based on computational capabilities. However, these methods achieve only marginal computational savings, as dense feed-forward computations dominate. Sparse Mixture-of-Experts (SMoE) provides a promising alternative through conditional computation, yet we identify that its naive application to heterogeneous federated settings introduces two critical discordances: (i) expert utilization imbalance and (ii) non-differentiability of Top-K routing. Our convergence analysis demonstrates that these discordances lead to degraded convergence, particularly for resource-constrained clients. To address these challenges, we propose Universally Balanced Sparse Mixture-of-Experts (UB-SMoE), which introduces Dynamic Modulated Routing (DMR) to rebalance expert utilization, and Universal Pseudo-Gradient (PG) to reconstruct learning signals for non-activated experts. These mechanisms form a self-reinforcing cycle that maintains expert viability across heterogeneous clients. Experiments on benchmarks show that UB-SMoE achieves up to $45.0\%$ computational reduction on low-resource clients while improving their performance by $8.7 \times$ compared to existing heterogeneous LoRA-rank methods.

Knowledge editing (KE) provides a lightweight alternative to repeated fine-tuning of LLMs. However, most existing KE methods target dense feed-forward layers, while modern LLMs increasingly adopt Mixture-of-Experts (MoE) architectures for their superior memory footprint and inference efficiency. This mismatch leaves a growing class of production models without principled editing tools. We propose a MEMIT-like framework for knowledge editing in MoE-based LLMs. Our method exploits the tensor structure of MoE layers to formulate the editing objective faithfully at the per expert level, and applies the Woodbury matrix identity to avoid materializing or inverting the full stacked matrix of expert weights. The resulting update reduces to inversions of fixed low-rank matrices and requires no additional backward passes. Empirically, our approach matches the editing quality of strong baselines on the main KE metrics while accelerating the editing procedure by up to 6x, owing to the batched MEMIT-style formulation and the low-dimensional inversions enabled by the Woodbury identity. These results show that closed-form, parameter-modifying KE can be extended efficiently beyond dense layers, opening a path toward scalable knowledge editing in modern sparse LLM architectures.

Multimodal clinical prediction faces three challenges: multiple foundation models (FMs) with complementary strengths per modality, pervasive missing modalities at training and test time, and sample-specific variation in modality contributions. We introduce MedMIX, a multimodal framework that combines intra-modality expert fusion, learned inter-modality fusion, and training-only large--small model collaboration for robust medical prediction under incomplete modalities. Within each modality, MedMIX aggregates complementary embeddings from multiple small expert models; across modalities, it performs learned fusion over available modalities; and during training, it leverages large teacher models to improve deployed representations without additional inference cost. Across three heterogeneous benchmarks (OpenI, MIMIC-IV-MM, and MMIST-ccRCC), MedMIX achieves consistently strong performance while remaining robust under controlled missing-modality perturbations, and further demonstrates sustained robustness under cross-cohort shift on MIMIC-III. These results highlight MedMIX as a practical framework that unifies within-modality expert collaboration, sample-specific cross-modality fusion, and efficient large--small model collaboration while remaining robust to incomplete modalities.

Heterogeneous Graph Prompt Learning (HGPL)has emerged as a promising paradigm for bridging the gap between the objectives of pre-training foundation models and their downstream applications in heterogeneous graph settings. However, existing HGPL methods are primarily designed for in-domain scenarios, whereas real-world deployments often span multiple domains, and the data used for pre-training and downstream tasks may originate from different distributions. Consequently, the applicability of current HGPL approaches is limited to in-domain settings, and their performance typically degrades when application domains shift. To address this serious limitation, we develop CHoE, a cross-domain HGPL method built upon an expert network. During pre-training, we introduce and train structure-conditioned experts, and during prompt tuning, we adopt a structure-aware expert routing and load balancing mechanism to select structurally compatible experts for each meta-path view. In addition, we design a prompt-based semantic fusion module to integrate representations across multiple views for downstream prediction. Extensive experiments show that CHoE consistently improves performance in few-shot cross-domain applications, outperforming all baseline approaches.

Mixture-of-Experts (MoE) networks promise favorable accuracy-compute trade-offs, yet practical vision deployments are hindered by expert collapse and limited end-to-end efficiency gains. We study when sparse top-$k$ routing with hard capacity constraints helps in vision classification, evaluated under multi-seed protocols on four benchmarks (CIFAR-10/100, Tiny-ImageNet, ImageNet-1K). We observe a \emph{compute-leverage pattern}: positive accuracy gaps require a substantial fraction $ρ$ of total FLOPs to be routed; at ImageNet scale this is necessary but not sufficient, as multi-expert routing ($k \geq 2$) is additionally required. Two controlled experiments isolate these factors. A hidden-size sweep on CIFAR-10 yields both predicted sign reversals across standard and depthwise backbones, ruling out backbone family as the active variable. An ImageNet-1K ablation that varies only top-$k$ -- holding architecture, initialization, and $ρ$ fixed -- reverses the gap from positive to negative across all five seeds. A per-sample variant of Soft MoE that softmaxes over experts rather than the batch rescues CIFAR-100 above the dense baseline, identifying batch-axis dispatch as the dominant failure mode in per-sample CNN settings. Code and aggregate results: https://github.com/libophd/sparse-moe-vision-rho.

Mixture-of-Experts (MoE) models rely on balanced expert utilization to fully realize their scalability. However, existing load-balancing methods are largely heuristic and operate on noisy mini-batch assignment statistics, introducing bias relative to population-level objectives. We propose $φ$-balancing, a principled framework that directly targets population-level expert balance by minimizing a strictly convex, symmetric, and differentiable potential of the expected routing distribution. Using convex duality, we derive an equivalent min-max formulation and obtain a simple online algorithm via mirror descent, yielding an efficient EMA-based routing adjustment with negligible overhead. Across large-scale pretraining and downstream fine-tuning, $φ$-balancing consistently outperforms prior Switch-style and loss-free baselines, demonstrating more stable and effective expert utilization.

Scaling Scientific Machine Learning (SciML) toward universal foundation models is bottlenecked by negative transfer: the simultaneous co-training of disparate partial differential equation (PDE) regimes can induce gradient conflict, unstable optimization, and plasticity loss in dense neural operators. In particular, broadband open-channel fluid dynamics and boundary-dominated porous media flows impose incompatible spectral and geometric demands on a single dense parameter path. We introduce Shodh-MoE, a sparse-activated latent transformer architecture for multi-physics transport. Shodh-MoE operates on compressed 16^3 physical latents produced by a physics-informed autoencoder with an intra-tokenizer Helmholtz-style velocity parameterization, restricting decoded states to divergence-free velocity manifolds. The model guarantees exact mass conservation, achieving a physically verifiable velocity divergence of ~2.8 x 10^-10 (evaluated post-hoc in FP64) on 128^3 grids. A Top-1 soft-semantic router dynamically assigns localized latent patches to expert subnetworks, enabling specialized parameter paths for distinct physical mechanisms while preserving shared experts for universal symmetries. In a 20,000-step distributed pretraining run over mixed three-dimensional physical tensors, routing telemetry shows autonomous domain bifurcation: held-out validation tokens from the open-channel domain route exclusively to Expert 0, while porous-media tokens route exclusively to Expert 1. The model converges simultaneously across both regimes, achieving latent validation MSEs of 2.46 x 10^-5 and 9.76 x 10^-6, and decoded physical MSEs of 2.48 x 10^-6 and 1.76 x 10^-6. These results support sparse expert routing as a practical architectural mechanism for mitigating multi-physics interference in universal neural operators.

Generative Recommendation (GenRec) models reformulate recommendation as a sequence generation task, representing items as discrete Semantic IDs used symmetrically as both inputs and prediction targets. We identify a critical dual-stage information bottleneck in this design: (1) the Input Bottleneck, where lossy quantization degrades fine-grained semantics, while popularity bias skews the learned representations toward frequent items, and (2) the Output Bottleneck, where imprecise discrete targets limit supervision quality. To address these issues, we propose AsymRec, an asymmetric continuous-discrete framework that decouples input and output representations. Specifically, Multi-expert Semantic Projection (MSP) maps continuous embeddings into the Transformer's hidden space via expert-specialized projections, preserving semantic richness and improving generalization to infrequent items. Multi-faceted Hierarchical Quantization (MHQ) constructs high-capacity, structured discrete targets through multi-view and multi-level quantization with semantic regularization, preventing dimensional collapse while retaining fine-grained distinctions. Extensive experiments demonstrate that AsymRec consistently outperforms state-of-the-art generative recommenders by an average of 15.8 %. The code will be released.

As the misuse of AI-generated images grows, generalizable image detection techniques are urgently needed. Recent state-of-the-art (SOTA) methods adopt aligned training datasets to reduce content, size, and format biases, empowering models to capture robust forgery cues. A common strategy is to employ reconstruction techniques, e.g., VAE and DDIM, which show remarkable results in diffusion-based methods. However, such reconstruction-based approaches typically introduce limited and homogeneous artifacts, which cannot fully capture diverse generative patterns, such as GAN-based methods. To complement reconstruction-based fake images with aligned yet diverse artifact patterns, we propose a GAN-based upsampling approach that mimics GAN-generated fake patterns while preserving content, size, and format alignment. This naturally results in two aligned but distinct types of fake images. However, due to the domain shift between reconstruction-based and upsampling-based fake images, direct mixed training causes suboptimal results, where one domain disrupts feature learning of the other. Accordingly, we propose a Separate Expert Fusion (SEF) framework to extract complementary artifact information and reduce inter-domain interference. We first train domain-specific experts via LoRA adaptation on a frozen foundational model, then conduct decoupled fusion with a gating network to adaptively combine expert features while retaining their specialized knowledge. Rather than merely benefiting GAN-generated image detection, this design introduces diverse and complementary artifact patterns that enable SEF to learn a more robust decision boundary and improve generalization across broader generative methods. Extensive experiments demonstrate that our method yields strong results across 13 diverse benchmarks. Codes are released at: https://github.com/liyih/SEF_AIGC_detection.

Mixture-of-Experts (MoE) architectures enhance the efficiency of large language models by activating only a subset of experts per token. However, standard MoE employs a fixed Top-K routing strategy, leading to redundant computation and suboptimal inference latency. Existing acceleration methods either require costly retraining with architectural changes or suffer from severe performance drop at high sparsity due to train-inference mismatch. To address these limitations, we propose BEAM (Binary Expert Activation Masking), a novel method that learns token-adaptive expert selection via trainable binary masks. With a straight-through estimator and an auxiliary regularization loss, BEAM induces dynamic expert sparsity through end-to-end training while maintaining model capability. We further implement an efficient custom CUDA kernel for BEAM, ensuring seamless integration with the vLLM inference framework. Experiments show that BEAM retains over 98% of the original model's performance while reducing MoE layer FLOPs by up to 85%, achieving up to 2.5$\times$ faster decoding and 1.4$\times$ higher throughput, demonstrating its effectiveness as a practical, plug-and-play solution for efficient MoE inference.

Mixture-of-Experts (MoE) models scale capacity by combining specialized experts, but most existing approaches assume centralized access to training data. In practice, data are distributed across clients and cannot be shared due to privacy constraints, making unified MoE training challenging. We propose MetaMoE, a privacy-preserving framework that unifies independently trained, domain-specialized experts into a single MoE using public proxy data as surrogates for inaccessible private data. Central to MetaMoE is diversity-aware proxy selection, which selects client-domain-relevant and diverse samples from public data to effectively approximate private data distributions and supervise router learning. These proxies are further used to align expert training, improving expert coordination at unification time, while a context-aware router enhances expert selection across heterogeneous inputs. Experiments on computer vision and natural language processing benchmarks demonstrate that MetaMoE consistently outperforms recent privacy-preserving MoE unification methods. Code is available at https://github.com/ws-jiang/MetaMoE.

We present EnergyLens, an end-to-end framework for energy-aware large language model (LLM) inference optimization. As LLMs scale, predicting and reducing their energy footprint has become critical for sustainability and datacenter operations, yet existing approaches either require production-level code and expensive profiling or fail to accurately capture multi-GPU energy behavior. As a result, practitioners lack tools for deciding which optimizations to prioritize and for selecting among existing deployment configurations when exhaustive profiling is impractical. EnergyLens addresses this gap with an intuitive einsum-based interface that captures LLM specifications including fusion, parallelism, and compute-communication overlap, combined with load-imbalance-aware MoE modeling and an empirically driven communication energy model for multi-GPU settings. We validate EnergyLens on Llama3 and Qwen3-MoE across tensor-parallel and expert-parallel configurations, achieving mean absolute percentage errors (MAPEs) between 9.25% and 13.19% for multi-GPU prefill and decode energy, and 12.97% across SM allocations for Megatron-style overlap. Our energy-driven exploration reveals up to 1.47x and 52.9x energy variation across configurations in prefill and decode efficiency and motivates distributed serving. We further show that compute-communication overlap is difficult to optimize with intuition alone, but EnergyLens correctly identifies Pareto-optimal overlap configurations.

Recent frontier large language models predominantly rely on Mixture-of-Experts (MoE) architectures. Despite empirical progress, there is still no principled understanding of how hyperparameters should scale with network width $N$, expert width $N_e$, number of experts $M$, sparsity $K$, and depth $L$ to ensure both stability and optimal performance at scale. We take a principled step toward resolving this gap by analyzing three different scaling regimes: (I) co-scaling $N\asymp N_e$, (II) co-scaling $N\asymp M\asymp K$, and (III) full proportional scaling of $N, N_e, M$, and $K$. For each regime, we develop a novel Dynamical Mean Field Theory (DMFT) description of the limiting training dynamics of MoEs that provides a formal foundation for our analysis. Within this framework, we derive the unique parameterization for SGD and Adam satisfying all maximal-update ($μ$) desiderata. We then show that the resulting $μ$P prescription does not reliably induce monotonic improvement with scale or robust learning-rate transfer. We trace these pathologies to scale-dependent observables in the aggregation dynamics, which motivates a refined set of desiderata that we term maximal scale stability. Guided by this principle, we derive a Maximally Scale-Stable Parameterization (MSSP) for both SGD and Adam in all three scaling regimes, and characterize the corresponding limiting dynamics - qualitatively distinct from the $μ$P limit - through a separate DMFT analysis. Experiments verify that MSSP robustly recovers learning rate transfer and monotonic improvement with scale across regimes. Combined with existing depth-scaling theory, these results provide a complete scaling prescription for MoE architectures as a function of width, depth, expert width, and number of experts.

Sparse Mixture-of-Experts (MoE) layers route tokens through a handful of experts, and learning-free compression of these layers reduces inference cost without retraining. A subtle obstruction blocks every existing compressor in this family: three experts can each be pairwise compatible yet form an irreducible cycle when merged together, so any score that ranks experts on pairwise signals is structurally blind to which triples are jointly mergeable. We show the obstruction is a precise mathematical object, the harmonic kernel of the simplicial Laplacian on a 2-complex whose vertices are experts, whose edges carry KL merge barriers, and whose faces carry triplet barriers; Hodge-decomposing the edge-barrier signal isolates the kernel exactly. We turn the diagnostic into a selection objective: HodgeCover greedily covers the harmonic-critical edges and triplet-critical triangles, and a hybrid variant of HodgeCover pairs it with off-the-shelf weight pruning on survivors. On three open-weight Sparse MoE backbones under aggressive expert reduction, HodgeCover matches state-of-the-art learning-free baselines on the expert-reduction axis, leads on the aggressive-compression frontier of the hybrid axis, and uniquely balances retained mass across all four Hodge components. These results show that exposing the harmonic kernel of a learned MoE structure changes which compressor wins at the regime that matters most.

We study dense and mixture-of-experts (MoE) transformers in a tiny-scale pretraining regime under a shared LLaMA-style decoder training recipe. The sparse model replaces dense feed-forward blocks with Mixtral-style routed experts. Dense baselines are modestly width-resized to tightly match either active or total parameter budgets, while tokenizer, data, optimizer, schedule, depth, context length, normalization style, and evaluation protocol are held fixed. Our best sparse recipe uses four experts, top-2 routing, Switch-style load balancing, and router z-loss. In a three-seed full-data comparison, the dense active-match model reaches 1.6545 +/- 0.0012 best validation loss, the MoE reaches 1.5788 +/- 0.0020, and the dense total-match model reaches 1.5608 +/- 0.0025. This yields a matched-active gap of 0.0758 +/- 0.0021 in the MoE's favor and a matched-total gap of 0.0180 +/- 0.0020 in the dense model's favor. Across training, the matched-active advantage grows while the matched-total dense advantage narrows sharply. In this sub-25M-parameter regime, MoE therefore improves validation loss under active-parameter matching but does not surpass dense training at equal total stored capacity.

Sparse Mixture-of-Experts (MoE) models offer a powerful way to scale model size without increasing compute, as per-token FLOPs depend only on k active experts rather than the total pool of E experts. Yet, this asymmetry creates an MoE efficiency paradox in practice: adding more experts balloons memory and communication costs, making actual training inefficient. We argue that this bottleneck arises in part because current MoE training allocates too many experts from the beginning, even though early-stage data may not fully utilize such capacity. Motivated by this, we propose EMO, a simple progressive training framework that treats MoE capacity as expandable memory and grows the expert pool over the course of training. EMO explicitly models sparsity in scaling law to derive stage-wise compute-optimal token budgets for progressive expansion. Empirical results show that EMO matches the performance of a fixed-expert setup in large-scale experiments while improving wall-clock efficiency. It offers a surprisingly simple yet effective path to scalable MoE training, preserving the benefits of large expert pools while reducing both training time and GPU cost.

Model merging aims to combine multiple task-specific expert models into a single model without joint retraining, offering a practical alternative to multi-task learning when data access or computational budget is limited. Existing methods, however, face two key limitations: (1) they overlook the valuable inductive bias of strong anchor models and estimate the merged weights from scratch, and (2) they rely on a shared hyperparameter setting across different modules of the network, lacking a global optimization strategy. This paper introduces Bayesian Model Merging (BMM), a plug-and-play bi-level optimization framework, where the inner level formulates the model merging as an activation-based Bayesian regression under a strong prior induced by an anchor model, yielding an efficient closed-form solution; and the outer level leverages a Bayesian optimization procedure to search module-specific hyperparameters globally based on a small validation set. Furthermore, we reveal a key alignment between activation statistics and task vectors, enabling us to derive a data-free variant of BMM that estimates the Gram matrix for regression without any auxiliary data. Across extensive benchmarks, including up to 20-task merging in vision and 5-task merging in language, BMM consistently outperforms all plug-and-play anchor baselines (e.g., TA, WUDI-Merging, and TSV). In particular, on the ViT-L/14 benchmark for 8-task merging, a single merged model reaches 95.1, closely matching the average performance of eight task-specific experts (95.8).

This paper systematically diagnoses the training failure modes of Token-Choice sparse Mixture-of-Experts (MoE) on video Diffusion Transformers. Starting from a pretrained dense model of about 5 billion parameters, we convert it into an MoE architecture following three laws: routed experts exactly clone the original FFN weights, shared experts are initialized to zero for verification and then to extremely small non-zero noise for actual training, while only the gating networks start from random initialization. Experiments reveal a hierarchy of five failure modes: (1) linear routers suffer global soft saturation with complete expert homogenization; (2) MLP routers introduce selective deadlock, where roughly one-third of layers degenerate into a single-expert mode that cannot be prevented by increasing the auxiliary loss; (3) cross-attention routers exhibit preliminary self-recovery, yet about nine layers remain stubbornly deadlocked; (4) deadlocked layers display a U-shaped distribution, concentrated in shallow visual processing layers and deep semantic integration layers; (5) bfloat16 mixed precision causes tiny weight updates to be truncated to zero by hardware. Based on routing decision time series over 65 million tokens across 5,000 training steps, we propose the Functional Redundancy Hypothesis: deadlock is a rational waiting strategy before the shared expert matures within the gate-shared expert-routed expert triadic system. This hypothesis is supported by the theory of functional redundancy in systems biology. On the engineering side, we summarize the Three Laws of dense-to-MoE conversion and provide a complete solution for the bfloat16 precision trap. We calibrate the current capability boundary of the Token-Choice paradigm and outline a three-step evolutionary roadmap from visual unification to a world model.

Sparse Mixture-of-Experts (SMoE) models enable scaling language models efficiently, but training them remains challenging, as routing can collapse onto few experts and auxiliary load-balancing losses can reduce specialization. Motivated by these hurdles, we study how routing decisions in SMoEs are formed mechanistically. First, we reveal a geometric coupling between routers and their corresponding experts. For a given token, the router weights for the selected expert and the expert weights processing it receive gradients along the same input direction, differing only in scalar coefficients. Thus, matched router--expert directions accumulate the same routed token history. This theoretical coupling also appears empirically in routing dynamics. In a $1$B SMoE trained from scratch, higher router scores predict stronger expert neuron activations, showing that routing decisions are mirrored inside the selected expert. Next, we analyze the effects of auxiliary load balancing on the router--expert geometric coupling, showing that such losses break this structure by spreading input-directed gradients across router weights, making distinct router directions nearly three times more similar to each other. Last, we demonstrate the centrality of geometric coupling for effective routing with a parameter-free online K-Means router, in which each expert maintains a running average of the hidden states routed to it and tokens are assigned based on cosine similarity. Compared with auxiliary-loss and loss-free balancing, this router achieves the lowest load imbalance with only a modest perplexity increase, indicating that geometric coupling captures a substantial part of what the router learns. Overall, our results explain how routers form assignment geometry that supports an effective division of labor.

The Segment Anything Model 2 (SAM2) has emerged as a foundation model for universal segmentation. Owing to its generalizable visual representations, SAM2 has been successfully applied to various downstream tasks. However, extending SAM2 to the RGB-D video salient object detection (RGB-D VSOD) task encounters three challenges including limited spatial modeling of linear LoRA, insufficient employment of SAM's multi-scale features, and dependence of initialization on explicit prompts. To address the issues, we present Multi-Modal Mixture-of-Experts with Memory-Augmented SAM (M$^4$-SAM), which equips SAM2 with modality-related PEFT, hierarchical feature fusion, and prompt-free memory initialization. Firstly, we inject Modality-Aware MoE-LORA, which employs convolutional experts to encode local spatial priors and introduces a modality dispatcher for efficient multi-modal fine-tuning, into SAM2's encoder. Secondly, we deploy Gated Multi-Level Feature Fusion, which hierarchically aggregates multi-scale encoder features with an adaptive gating mechanism, to balance spatial details and semantic context. Finally, to conduct zero-shot VSOD without manual prompts, we utilize a Pseudo-Guided Initialization, where a coarse mask is regarded as a pseudo prior and used to bootstrap the memory bank. Extensive experiments demonstrate that M$^4$-SAM achieves the state-of-the-art performance across all evaluation metrics on three public RGB-D VSOD datasets.

Mixture-of-Experts (MoE) architectures have become standard in large language models, yet many of their core design choices - expert count, granularity, shared experts, load balancing, token dropping - have only been studied one or two at a time over narrow configuration ranges. It remains an open question whether these choices can be optimized independently, without considering interactions. We present the first systematic study of over 2,000 pretraining runs spanning models up to 6.6B total parameters, in which we exhaustively vary total experts, expert dimension, heterogeneous expert sizing within a single layer, shared expert size and load-balancing mechanisms. We find that at every active-parameter scale that we study, performance consistently improves with total MoE parameters even at extreme active expert parameter ratios like 128.Further, the optimal expert size is nearly invariant to total parameter count and depends only on active parameter count. Third, we see that other choices like shared experts, heterogeneous experts and load-balancing settings have small effects relative to expert count and granularity, although dropless routing yields a consistent gain. Overall, our results suggest a simpler recipe: focus on expert count and granularity, other choices have minimal effect on final quality.

All-in-one image restoration seeks to recover clean images from inputs affected by diverse and unknown degradations using a unified framework. Recent methods have shown strong performance by identifying degradation characteristics to guide the restoration process. However, many of them treat degradations as discrete categories, which limits their ability to model the continuous relational structure that arises in composite degradations. To address this issue, we propose a multimodal large language model (MLLM)-guided image restoration framework that exploits multimodal embeddings as guidance for low-level restoration. Specifically, MLLM-derived features are injected into an encoder-decoder architecture through an MLLM-guided fusion block (MGFB) to enhance degradation-aware representations. In addition, we incorporate a mixture-of-frequency-experts (MoFE) module that adaptively combines frequency experts using MLLM-guided contextual cues. To further improve expert routing, we design an MLLM-guided router with a relational alignment loss that encourages routing patterns consistent with the embedding-space relationships of degraded inputs. Extensive experiments on multiple benchmarks show that the proposed method achieves strong performance across diverse restoration settings and establishes a new state of the art on the challenging CDD11 dataset, outperforming previous methods by up to 1.35 dB.

While Mixture-of-Experts (MoE) scales model capacity without proportionally increasing computation, its massive total parameter footprint creates significant storage and memory-access bottlenecks, which hinder efficient end-side deployment that simultaneously requires high performance, low computational cost, and small storage overhead. To achieve these properties, we present DECO, a sparse MoE architecture designed to match the performance of dense Transformers under identical total parameter budgets and training tokens. DECO utilizes the differentiable and flexible ReLU-based routing enhanced by learnable expert-wise scaling, which adaptively balances the contributions of routed and shared experts. Furthermore, we introduce NormSiLU, an activation function that normalizes inputs prior to SiLU operators, producing a more stable trend of routed-expert activation ratio and a higher intrinsic sparsity level. We also identify an empirical advantage in using non-gated MLP experts with ReLU-based routing, indicating the possibility of MoE architecture simplification. Experiments demonstrate that DECO, activating only 20% of routed experts, matches dense performance and outperforms established MoE baselines. Our specialized acceleration kernel delivers a 2.93$\times$ speedup on Jetson AGX Orin compared with dense inference. Code and checkpoints are available at https://github.com/thunlp/DECO.

We propose a novel adaptive Mixture-of-Experts (MoE) framework for time series forecasting that addresses the optimization problem arising from small gating weights by incorporating expert-specific losses, which provide each expert with a direct learning signal independent of the gate-assigned weight. Specifically, the overall objective comprises the base forecasting loss and expert-specific losses, allowing individual expert prediction errors to directly influence parameter updates alongside the aggregate forecasting error. The framework also encourages different experts to learn from different temporal segments of the data. The proposed framework is further combined with a partial online learning strategy that enables efficient incremental updates of model parameters. By integrating expert-level loss information with partial online optimization, the proposed method improves forecasting performance while retaining computational efficiency. Empirical results across economic, tourism, and energy datasets with different sampling frequencies show that the proposed approach generally outperforms state-of-the-art supervised neural forecasting models, including Transformer-based architectures such as PatchTST, as well as zero-shot time-series foundation models such as TimeMoE. Furthermore, ablation studies confirm the effectiveness of the expert-specific loss integration strategy, highlighting its contribution to enhancing predictive performance.

Low-Rank Adaptation (LoRA) dominates parameter-efficient fine-tuning of large language models, yet most variants target dense architectures. Mixture-of-Experts (MoE) models scale parameters at near-constant per-token compute, and their sparse activation patterns create untapped opportunities for more efficient adaptation. We propose Hot-Experts Layer-level Low-Rank Adaptation (HELLoRA), which attaches LoRA modules only to the most frequently activated experts at each layer. This simple mechanism reduces trainable parameters and adapter-induced FLOPs while improving downstream performance, an effect we attribute to a form of structured regularization that preserves pretrained expert specialization. To stress-test HELLoRA under extreme parameter budgets, we further compose it with LoRI to form HELLoRI, which freezes the up-projection and sparsifies the down-projection. Across three MoE backbones, namely OlMoE-1B-7B, Mixtral-8x7B, and DeepSeekMoE, and three task families covering mathematical reasoning, code generation, and safety alignment, HELLoRA consistently outperforms strong PEFT baselines. Relative to vanilla LoRA on OlMoE, HELLoRA uses 15.7% of the trainable parameters, reduces adapter FLOPs by 38.7%, achieves 1.9x the training throughput, and improves accuracy by 9.2%. On DeepSeekMoE, HELLoRA outperforms LoRA while using only 23.2% of its trainable parameters. These results demonstrate that activation-aware adapter placement is an effective and practical route to scaling PEFT for MoE language models.

Standard Mixture-of-Experts (MoE) transformers route tokens to expert subnetworks within each layer, but the layer structure itself remains monolithic. We introduce Mixture of Layers (MoL), which replaces full-width transformer blocks (d_model) with K parallel thin blocks at reduced dimensionality (d_thin << d_model), connected via learned down/up projections and composed via top-k block routing. Scaling sparse block routing to many blocks creates an attention coverage problem, as each block sees fewer tokens. We address this by introducing hybrid attention, which pairs one shared softmax block for global context with Gated DeltaNet linear attention in routed blocks.

Real-world model deployment across multiple domains requires multimodal models to operate under two complementary regimes: (1) multi-task pretraining, tasks are co-available at design time where related tasks could borrow representational strength from one another, (2) continual adaptation, in which new tasks emerge after deployment with previously unseen modality combinations. However, neither regime alone suffices: the pretraining task set is never exhaustive, while bypassing joint training forfeits the transfer gains and efficiency among co-trainable tasks. Sparse Mixture-of-Experts (MoE) is a natural fit for this dual requirement: sparse activation enables modular capacity expansion as new tasks arrive, while routing decouples modality-level computation from task-level composition. In this work, we propose a scalable MoE framework for multitask pretraining and continual learning across flexible modality combinations. The framework is designed to support training on multimodal tasks with diverse modality configurations by leveraging modality-specific routers that process tokens from each modality across tasks. Furthermore, it enables continual learning over sequential multimodal tasks within a fixed-capacity MoE by compressing accumulated expert knowledge into low-rank memory subspaces, while expanding only the lightweight routers. We validate the effectiveness of our method on multiple healthcare multimodal benchmarks. It demonstrates competitive multitask pretraining performance while alleviating catastrophic forgetting and improving parameter efficiency.

Mixture-of-experts (MoE) architectures enable trillion-parameter LLMs with sparsely activated experts. Expert parallelism (EP) is a widely adopted MoE training strategy, but it suffers from severe all-to-all communication bottlenecks, which is exaggerated by the limited inter-node network bandwidth as the growing model size requires distributing experts across GPU nodes. Prior work focused on overlapping these all-to-all communications with feed-forward network (FFN) and self-attention computations, which often leaves residual network-bound stalls due to inherent imbalance in attention and FFN layers' computation-communication ratios. We present DisagMoE, a disaggregated MoE training system that jointly optimizes model placement and scheduling for maximal efficiency. DisagMoE separates attention and FFN layers into disjoint GPU groups, introduces a multi-stage pipeline with uni-directional, many-to-many communications, and employs a computation-communication roofline model to balance GPU and network bandwidth allocation among the attention and FFN groups. DisagMoE is implemented on Megatron-LM, and evaluation shows that DisagMoE improves training efficiency across multiple MoE models with up to 1.8x speedup on 16-node 8xH800 clusters.

Mixture-of-Experts (MoE) models achieve remarkable performance by sparsely activating specialized experts, yet their massive parameters in experts pose significant challenges for deployment. While low-rank quantization offers a promising route to compress MoE models, existing methods still incur nonnegligible memory overhead and inference latency. To address these limitations, we propose \textsc{TileQ}, a fine-tuning-free post-training quantization (PTQ) method that employs 2D-tiling structured low-rank quantization to share low-rank factors across both input and output dimensions of MoE experts. Furthermore, we introduce an efficient inference technique for \textsc{TileQ} that fuses multiple low-rank expert computations into a single-pass operation, significantly improving hardware utilization. Experiments show that \textsc{TileQ} cuts down additional memory usage up to 10$\times$ and reduces inference latency to $\sim$5% while preserving state-of-the-art accuracy.

Looped language models repeat a set of transformer layers through depth, reducing memory costs and providing natural early-exit points at loop boundaries. However, looped models do not scale as favorably as standard transformers with unique layers. We compare standard and Mixture-of-Experts (MoE) transformers, with and without looping, and find two main results. First, we find Looped-MoE models scale better than the standard baseline while dense looped models do not. We trace this to routing divergence between loops: in Looped-MoE models, different experts are activated on each pass through the same shared layers, recovering expressivity without additional parameters. Our second finding is that looped models have better compute-quality trade-offs with early exits than standard models. Because each loop ends with the same layers that produce the final output, loop boundaries are superior exit points, as confirmed by earlier output convergence at these points. In sum, we provide a clear direction for scaling looped models: a Looped-MoE model with early exits can not only beat standard transformers at scale, but also enable significant memory and inference savings with minimal degradation in quality.

Mixture-of-Experts (MoE) language models organize knowledge into explicitly routed expert modules, making expert-level representations traceable and analyzable. By analyzing expert activation patterns in MoE large language models (LLMs), we find that a subset of experts is consistently activated across diverse knowledge domains. These common experts encode cross-domain, generalizable knowledge that is closely related to model generalization, naturally raising the question of how such identifiable expert knowledge can be practically reused. Motivated by this observation, we propose XPERT, a framework that extracts, consolidates, and reuses expert knowledge from pre-trained MoE LLMs to support more effective training of language models across different model scales. XPERT identifies cross-domain experts via inference-only analysis, refines their representations through tensor decomposition, and adapts the extracted knowledge to reuse in downstream models. Experiments on language understanding and dialogue generation benchmarks show that models benefiting from reused expert knowledge achieve consistently stronger performance and faster convergence compared to strong baselines. These results highlight MoE LLMs as structured and reusable knowledge sources, and demonstrate the value of expert-level knowledge reuse for improving model training.

Structured pruning and knowledge distillation (KD) are typical techniques for compressing large language models, but it remains unclear how they should be applied at pretraining scale, especially to recent mixture-of-experts (MoE) models. In this work, we systematically study MoE compression in large-scale pretraining, focusing on three key questions: whether pruning provides a better initialization than training from scratch, how expert compression choices affect the final model after continued training, and which training strategy is most effective. We have the following findings: First, across depth, width, and expert compression, pruning a pretrained MoE consistently outperforms training the target architecture from scratch under the same training budget. Second, different one-shot expert compression methods converge to similar final performance after large-scale continual pretraining. Motivated by this, we introduce a simple partial-preservation expert merging strategy that improves downstream performance across most benchmarks. Third, combining KD with the language modeling loss outperforms KD alone, particularly on knowledge-intensive tasks. We further propose multi-token prediction (MTP) distillation, which yields consistent gains. Finally, given the same training tokens, progressive pruning schedules outperform one-shot compression, suggesting that gradual architecture transitions lead to better optimization trajectories. Putting it all together, we compress Qwen3-Next-80A3B to a 23A2B model that retains competitive performance. These results offer practical guidance for efficient MoE compression at scale.

Many biological systems evolve through continuous local dynamics while switching between latent regimes defined by learning, stimulus context, internal state, or developmental stage. These processes are often observed only as unpaired longitudinal snapshots: the same cells, neurons, or animals are not tracked as matched trajectories, even though population states are sampled across successive stages. This creates two coupled challenges. First, trajectories must respect curved low-dimensional manifolds embedded in high-dimensional biological measurements. Second, the model must identify when the transport mechanism itself changes. We introduce FLUX (FLow matching for Unpaired longitudinal data with miXture-of-experts), a geometry-aware longitudinal flow-matching framework for joint transport modeling and unsupervised regime discovery. FLUX learns a data-dependent metric from pooled labeled and unlabeled observations, uses that metric to construct geometry-aware conditional paths between adjacent marginals, and decomposes the resulting velocity field into sparse expert vector fields selected by a Straight-Through Gumbel-Softmax router. Across manifold controls, a regime-switching Lorenz system, widefield cortical calcium imaging during associative learning, and embryoid body single-cell differentiation, FLUX reconstructs longitudinal transport while recovering interpretable regime structure. Ablations show that mixture-of-experts routing alone is insufficient: FLUX without geometric learning can fit local transport but fails or weakens regime discovery when regimes are encoded in local dynamics. These results suggest that geometry-aware velocity decomposition provides a general strategy for discovering latent biological state transitions from unpaired longitudinal snapshots.

Load imbalance is a long-standing challenge in Mixture-of-Experts (MoE) training and is exacerbated in reinforcement learning (RL) for LLMs, where hot experts can shift frequently across micro-batches. Existing MoE training systems rely on historical loads to predict future expert demand, making them less effective under sharp fluctuations. We propose ReLibra, an MoE RL training system that exploits a unique opportunity in RL's rollout-training workflow, routing replay, to enable fine-grained load balancing at micro-batch granularity. Because rollout and training process the same tokens with the same MoE parameters, the token-to-expert routing decisions are known before training starts. Leveraging this information, ReLibra places two MoE load-balancing mechanisms at inter- and intra-batch timescales, matching their communication patterns to hierarchical network bandwidths. At the inter-batch timescale, ReLibra performs expert reordering to redistribute experts for batch-level cross-node balancing; at the intra-batch timescale, it dynamically performs expert replication within a node to absorb micro-batch-level load fluctuations. Experiments on diverse MoE LLMs and RL workloads show that ReLibra improves training throughput by up to 1.6$\times$ over Megatron-LM and by up to 1.2$\times$ over EPLB, even when EPLB is given oracle loads. Moreover, ReLibra remains within 6%-10% of the throughput of an idealized balanced baseline.

Mixture of Experts (MoE) architecture has become the standard for state-of-the-art large language models, owing to its computational efficiency through sparse expert activation. However, sparsity through finer expert granularity is becoming increasingly difficult to achieve due to fundamental training challenges such as expert collapse and load imbalance. In this work, we explore and leverage intra-expert activation sparsity as a complementary and underexplored dimension of sparsity in MoE models. Surprisingly, substantial intra-expert sparsity is readily available in existing pre-trained MoE models, without any modification to the activation function or model parameters, providing up to 90% sparsity within each expert without significant accuracy loss. We explore intra-expert activation sparsity across eight off-the-shelf MoE models ranging from 1B to 400B parameters, and extend the MoE execution pipeline of vLLM to leverage intra-expert activation sparsity by skipping the computations of inactive neurons, on top of its existing optimizations, achieving up to 2.5 times speedup in MoE layer execution and 1.2 times end-to-end speedup compared to the original dense vLLM baseline.

Sparse MoE models achieve a good balance between capacity and compute by routing each token to a small subset of experts. However, in most MoE architectures, once a token is routed, the selected experts process it independently and their outputs are combined via a weighted sum. This leaves open whether enabling communication among them could improve performance. While prior work has raised this question, direct interaction among the active routed experts remains underexplored. In this paper, we propose SDG-MoE (Signed Debate Graph Mixture-of-Experts), a novel architecture that adds a lightweight, iterative deliberation step before final aggregation. SDG-MoE introduces three components: (i) two learned interaction matrices over the active experts, a support graph $A^+$ and a critique graph $A^-$, capturing reinforcing and corrective influences; (ii) a signed message-passing step that updates expert representations before aggregation; and (iii) a disagreement-gated Friedkin-Johnsen-style anchoring that controls deliberation strength while preventing expert drift. Together, these enable a structured deliberation process where interaction strength scales with disagreement and specialization is preserved. We also provide a theoretical analysis establishing stability conditions on expert states and showing that deliberation adds only low-order overhead over the active set. In controlled three-seed pretraining experiments, SDG-MoE improves validation perplexity over both an unsigned graph communication baseline and vanilla MoE, outperforming the strongest baseline by 19.8%, and gives the best external perplexity on WikiText-103, C4, and Paloma among the compared systems.

Continual learning enables vision-language models to accumulate knowledge and adapt to evolving tasks without retraining from scratch. However, in multi-domain task-incremental learning, large domain shifts intensify the stability-plasticity dilemma. Most existing methods rely on fixed architectures with statically allocated parameters, which limits adaptation to new domains and aggravates catastrophic forgetting. To address these challenges, we propose DIMoE-Adapters, a Dynamic Incremental Mixture-of-Experts Adapters framework that introduces a dynamic expert evolution paradigm to balance stability and plasticity. This paradigm is implemented through two collaborative components: Self-Calibrated Expert Evolution (SCEE) and Prototype-Guided Expert Selection (PGES). SCEE constructs and evolves a sparse expert pool through expert optimization dynamics, improving plasticity while reducing redundant capacity. PGES controls expert utilization based on the pool shaped by SCEE, improving stability across both previously encountered and unseen tasks. Extensive experiments show that DIMoE-Adapters outperforms previous state-of-the-art methods across various settings.

Sparse Mixture-of-Experts (MoE) models scale capacity by routing each token to a small subset of experts. However, their routers exhibit a fundamental trade-off: strong load balancing can suppress expert specialization, while aggressive diversity often causes routing collapse. We propose Hi-MoE, a grouped MoE framework that decomposes routing control into two coupled levels: (i) inter-group balancing that enforces fair traffic across expert groups, and (ii) intra-group specialization that promotes complementary expert behaviors while preventing within-group collapse. Our analysis provides a principled explanation of how our hierarchical objectives reshape the router, thereby promoting stable specialization and mitigating collapse. We observe consistent improvements over recent sparse-routing and grouped-MoE baselines across NLP and vision benchmarks, and confirm robustness via scaling studies (model size, expert count) and targeted ablations. In large-scale pre-training on 58B tokens, Hi-MoE-7B achieves a 5.6% perplexity reduction and a 40% improvement in expert balance over OLMoE-7B across diverse evaluation domains.

Mixture-of-Experts (MoE) language models route each token to a small subset of experts, but whether the routes selected by a trained top-$k$ router are good ones is rarely evaluated directly. Holding the model fixed, we compare each standard route against sampled equal-compute alternatives for the same token and score each by the next-token probability it assigns to the realized token in a verified reasoning trajectory. The result is sharply token-conditional: the standard router is well-aligned with route utility on confident tokens but uninformative on the fragile tokens that drive hard reasoning, where lower-loss equal-compute routes consistently exist inside the frozen model but are not selected. The same pattern holds across Qwen3-30B-A3B, GPT-OSS-20B, DeepSeek-V2-Lite, and OLMoE-1B-7B, and follows structurally from how standard top-$k$ training evaluates routing decisions: the language modeling loss scores only the executed route, and load balancing depends only on aggregate routing statistics. A minimal router-only update to the final-layer router, leaving every expert and every other router frozen, is sufficient to shift pass@K on AIME 2024+2025 and HMMT 2025 for both Qwen3-30B-A3B and GPT-OSS-20B, suggesting that at least part of the failure reflects router-reachable misallocation rather than expert capacity alone.

Mixture-of-Experts (MoE) architectures achieve scalable capacity through sparse routing, yet the geometric structure of expert specialization remains poorly understood. We introduce a unified Jacobian-PCA-Grassmann framework for analyzing MoE layers in both function space and representation space. Across pretrained MoE Transformers (Mistral, Qwen), we find a consistent structural asymmetry: experts exhibit strong functional decorrelation (consistently low, near-zero cross-expert Jacobian alignment) while their routed representations occupy distinct but partially overlapping subspaces. This indicates that functional decorrelation and representation overlap coexist rather than coincide in MoE specialization. Controlled routing experiments further indicate that routing sparsity appears to be a key factor shaping this geometry: top-k routing induces sharper functional separation and larger subspace divergence, whereas fully soft routing yields more entangled expert structure. Together, these results suggest a geometric interpretation in which MoE layers may be viewed as implementing locally decorrelated operators over overlapping submanifolds on a shared representation manifold, and provide a general diagnostic framework for studying conditional computation in modern Transformer architectures.

Generating high-performance GPU kernels remains challenging due to the need for both correctness and hardware-aware optimization. While large language models (LLMs) show promise in code generation, they often fail to produce kernels that are both correct and efficient. We propose Kernel Foundry, a diagnosis-driven evolutionary framework for automatic GPU kernel optimization. Our method combines expert-guided, retrieval-augmented initialization with a multi-island evolutionary search, where candidate kernels are iteratively refined using structured diagnostic feedback. A centralized experience library accumulates reusable optimization knowledge to guide subsequent evolution, while explicit mechanisms prevent cheating behaviors that bypass kernel-level computation. Experiments on KernelBench show that our method consistently improves both correctness and performance over strong baselines, achieving up to 100% correctness on Level~2.

Modern Mixture-of-Experts (MoE) architectures allocate expert capacity through a rigid per-layer rule: each transformer layer owns a separate expert set. This convention couples depth scaling with linear expert-parameter growth and assumes that every layer needs isolated expert capacity. However, recent analyses and our routing probe challenge this allocation rule: replacing a deeper layer's learned top-k router with uniform random routing drops downstream accuracy by only 1.0-1.6 points across multiple production MoE models. Motivated by this redundancy, we propose UniPool, an MoE architecture that treats expert capacity as a global architectural budget by replacing per-layer expert ownership with a single shared pool accessed by independent per-layer routers. To enable stable and balanced training under sharing, we introduce a pool-level auxiliary loss that balances expert utilization across the entire pool, and adopt NormRouter to provide sparse and scale-stable routing into the shared expert pool. Across five LLaMA-architecture model scales (182M, 469M, 650M, 830M, and 978M parameters) trained on 30B tokens from the Pile, UniPool consistently improves validation loss and perplexity over the matched vanilla MoE baselines. Across these scales, UniPool reduces validation loss by up to 0.0386 relative to vanilla MoE. Beyond raw loss improvement, our results identify pool size as an explicit depth-scaling hyperparameter: reduced-pool UniPool variants using only 41.6%-66.7% of the vanilla expert-parameter budget match or outperform layer-wise MoE at the tested scales. This shows that, under a shared-pool design, expert parameters need not grow linearly with depth; they can grow sublinearly while remaining more efficient and effective than vanilla MoE. Further analysis shows that UniPool's benefits compose with finer-grained expert decomposition.

Large language models are typically deployed as monolithic systems, requiring the full model even when applications need only a narrow subset of capabilities, e.g., code, math, or domain-specific knowledge. Mixture-of-Experts (MoEs) seemingly offer a potential alternative by activating only a subset of experts per input, but in practice, restricting inference to a subset of experts for a given domain leads to severe performance degradation. This limits their practicality in memory-constrained settings, especially as models grow larger and sparser. We introduce EMO, an MoE designed for modularity-the independent use and composition of expert subsets-without requiring human-defined priors. Our key idea is to encourage tokens from similar domains to rely on similar experts. Since tokens within a document often share a domain, EMO restricts them to select experts from a shared pool, while allowing different documents to use different pools. This simple constraint enables coherent expert groupings to emerge during pretraining using document boundaries alone. We pretrain a 1B-active, 14B-total EMO on 1T tokens. As a full model, it matches standard MoE performance. Crucially, it enables selective expert use: retaining only 25% (12.5%) of experts incurs just a 1% (3%) absolute drop, whereas standard MoEs break under the same setting. We further find that expert subsets in EMO specialize at semantic levels (e.g., domains such as math or code), in contrast to the low-level syntactic specialization observed in standard MoEs. Altogether, our results demonstrate a path toward modular, memory-efficient deployment of large, sparse models and open new opportunities for composable architectures.

We introduce E = T*H/(O+B), a dimensionless control parameter that predicts whether Mixture-of-Experts (MoE) models will develop a healthy expert ecology or collapse into dead experts. E combines four hyperparameters -- routing temperature T, routing entropy weight H, oracle weight O, and balance weight B -- into a single quantity. Through 12 controlled experiments (8 vision, 4 language) totaling over 11,000 training epochs, we establish that E >= 0.5 alone is sufficient to guarantee zero dead experts, removing the necessity for handcrafted load-balancing auxiliary losses. We validate this cross-modally on CIFAR-10, CIFAR-100, TinyImageNet-200, WikiText-2, and WikiText-103. Six additional findings emerge: (1) dead experts can resuscitate -- triggered by balance loss driving router re-exploration; (2) ortho toxicity is dataset-dependent, not universal; (3) task complexity shifts the critical E threshold; (4) model overfitting is decoupled from expert ecological health; (5) three-tier MoE spontaneously collapses into a two-tier functional structure; (6) ecological structure is temperature-invariant across a 50x range. We propose that E serves as a unified diagnostic for MoE training, analogous to the Reynolds number in fluid dynamics.

Real-world image degradation is often unknown, spatially non-uniform, and compositional, requiring all-in-one restoration models to adapt a single set of weights to diverse local corruption patterns without test-time degradation labels. Existing methods typically modulate a shared backbone with global prompts or degradation descriptors, or route features through predefined expert pools. However, compact global conditioning can bottleneck localized degradation evidence, while static expert routing may produce homogeneous updates or rely on unstable sparse assignments. We propose \textbf{Continuous Expert Assembly} (CEA), a token-wise dynamic parameterization framework for all-in-one image restoration. CEA employs a lightweight \textbf{Cross-Attention Hyper-Adapter} to probe intermediate spatial features and synthesize instance-conditioned low-rank routing bases and residual directions. Each spatial token then assembles its own residual update via dense signed dot-product affinities over the generated rank-wise components, avoiding external prompts, static expert banks, and discrete Top- selection. The resulting assembly rule also admits a linear-attention perspective, making its dense token-wise routing behavior transparent. Experiments on AIO-3, AIO-5, and CDD-11 show that CEA improves average restoration quality over strong prompt-, descriptor-, and expert-based baselines, with the clearest gains on spatially varying and compositional degradations, while maintaining favorable parameter, FLOP, and runtime efficiency.

Large-scale vision-language mixture-of-experts (VL-MoE) models provide strong multimodal capability, but efficient deployment on memory-constrained platforms remains difficult. Existing MoE offloading systems are largely designed for text-centric workloads and become much less effective for visual-heavy inputs, where large numbers of visual tokens induce broader and less predictable expert accesses. We present VisMMoE, a VL-MoE offloading system built on a single systems insight: pruning redundant visual tokens can improve offloading not only by reducing computation, but also by reshaping expert demand. We refer to this effect as \textit{visual-expert affinity}: token pruning makes expert accesses more concentrated within layers and more stable across layers, producing a smaller and more predictable expert working set. Guided by this insight, VisMMoE combines affinity-aware token compression, lookahead expert prediction, and cache/pipeline orchestration to improve expert locality and prefetch effectiveness under tight memory budgets. We implement VisMMoE on multiple frameworks and evaluate it on representative VL-MoE models and benchmarks. VisMMoE improves end-to-end inference performance by up to 2.68x and 1.61x, respectively, over strong baselines for today's VL-MoE deployments while maintaining competitive accuracy.

Anomaly detection in medical images is a challenging task, since anomalies are not typically available during training. Recent methods leverage a single pretext task coupled with a large-scale pre-trained model to reach state-of-the-art performance. Instead, we propose to learn multiple self-supervised and pseudo-labeling tasks from scratch, using a joint model based on Mixture-of-Experts (MoE). By carefully integrating multiple proxy tasks, the joint model effectively learns a robust representation of normal anatomical structures, so that anomaly scores can be derived based on how well the multi-task learner (MTL) solves each task during inference. We perform comprehensive experiments on BMAD, a recent benchmark that comprises a broad range of medical image modalities. The empirical results indicate that our multi-task learner is an effective anomaly detector, outperforming all state-of-the-art competitors on BMAD. Moreover, our model produces interpretable anomaly maps, potentially helping physicians in providing more accurate diagnoses.

We present ZAYA1-8B, a reasoning-focused mixture-of-experts (MoE) model with 700M active and 8B total parameters, built on Zyphra's MoE++ architecture. ZAYA1-8B's core pretraining, midtraining, and supervised fine-tuning (SFT) were performed on a full-stack AMD compute, networking, and software platform. With under 1B active parameters, ZAYA1-8B matches or exceeds DeepSeek-R1-0528 on several challenging mathematics and coding benchmarks, and remains competitive with substantially larger open-weight reasoning models. ZAYA1-8B was trained from scratch for reasoning, with reasoning data included from pretraining onward using an answer-preserving trimming scheme. Post-training uses a four-stage RL cascade: reasoning warmup on math and puzzles; a 400-task RLVE-Gym curriculum; math and code RL with test-time compute traces and synthetic code environments built from competitive-programming references; and behavioral RL for chat and instruction following. We also introduce Markovian RSA, a test-time compute method that recursively aggregates parallel reasoning traces while carrying forward only bounded-length reasoning tails between rounds. In TTC evaluation, Markovian RSA raises ZAYA1-8B to 91.9% on AIME'25 and 89.6% on HMMT'25 while carrying forward only a 4K-token tail, narrowing the gap to much larger reasoning models including Gemini-2.5 Pro, DeepSeek-V3.2, and GPT-5-High.

Resource-efficient machine learning increasingly uses sparse Mixture-of-Experts (MoE) architectures, where the gate acts as both a learning component and a routing interface controlling computation, communication, and accuracy. Motivated by finite-rate interpretations of MoE gating, we treat the gate as a stochastic channel and use $I(X;T)$ to quantify the routing information available to the selected expert. To make the associated information quantities tractable beyond synthetic examples, we develop a finite-bank MNIST construction using pretrained CNN experts and a discrete, data-dependent selection rule. Since the selected model belongs to a finite candidate set, the algorithmic mutual information $I(S;W)$ admits a closed-form discrete-entropy estimator from the empirical posterior $q(W|S)$. Sweeping a data-dependence parameter $α$, we observe that $\widehat I(S;W)$ monotonically tracks the generalization gap, while the Xu-Raginsky bound exhibits the expected looseness. We also compare with a uniform union-bound baseline and introduce an empirical estimator of $I(X;T)$ together with a Blahut-Arimoto procedure for tracing an accuracy-rate curve over the expert bank. The proposed framework provides a practical tool for analyzing resource-aware MoE inference systems and for interpreting $I(X;T)$ and $D(R_g)$ as design proxies for efficient expert routing.

Frontier models increasingly adopt Mixture-of-Experts (MoE) architectures to achieve large-model performance at reduced cost. However, training MoE models on HPC platforms is hindered by large memory footprints, frequent large-scale communication across heterogeneous networks, and severe workload imbalance. To characterize these challenges, we develop a mathematical model that quantifies memory, compute, and communication requirements for MoE configurations under various parallelization schemes, verified through micro-benchmarking, code instrumentation, and hardware profiling. Our analysis identifies performance bottlenecks: all-to-all latency at scale from expert parallelism, insufficient compute-communication overlap, low GPU utilization from imbalanced skinny GEMMs, and the absence of platform-aware hybrid parallelization strategies. To address these, we introduce Piper, a framework that leverages resource modeling to identify efficient training strategies for MoE models on target HPC platforms, applying pipeline parallelism with optimized schedules. Piper achieves 2-3.5X higher MFU than state-of-the-art frameworks such as X-MoE, and a novel all-to-all algorithm delivers 1.2-9X bandwidth over vendor implementation.

Mixture-of-experts (MoE) models enable scalable transformer architectures by activating only a subset of experts per token. Recent evidence suggests that performance improves with increasingly granular experts, i.e., many small experts instead of a few large ones. However, this regime substantially increases routing cost, which can dominate computation. We introduce adaptive inverted-index routing for MoE (AIR-MoE), an inverted-index-inspired routing architecture based on vector quantization (VQ). In a first stage, AIR-MoE performs coarse shortlisting by assigning tokens to VQ codewords to construct a candidate set of experts. In a second stage, fine scoring computes exact routing scores restricted to this shortlist. This two-stage procedure approximates true top-k routing while avoiding full expert scoring and, in contrast to prior work, imposing no structural constraints on expert parameters. AIR-MoE serves as a drop-in replacement for standard routers and requires no modifications to the model architecture or loss function. We further provide a lower bound on the mass recall achieved by AIR-MoE that yields insights into its inner workings. Empirically, we demonstrate that AIR-MoE achieves improved performance compared to existing routing approaches in granular MoE settings.

Deep neural network (DNN) inference at the edge demands simultaneous improvements in accuracy, computational efficiency, and energy consumption. Approximate computing and Mixture-of-Experts (MoE) architectures have each been studied as independent routes towards efficient inference, the former by replacing exact arithmetic with low-power approximate multipliers, the latter by routing inputs through specialized expert sub-networks to enable conditional computation. However, their interaction remains entirely unexplored. This paper presents AxMoE, the first study of the impact of approximate multiplication on MoE DNN architectures. We evaluate three MoE variants: Hard MoE, Soft MoE, and Cluster MoE against dense baselines across three CNN architectures (ResNet-20, VGG11_bn, VGG19_bn) on CIFAR-100 and a Vision Transformer (ViT-Small) on Tiny ImageNet-200 dataset, using eight 8-bit signed multipliers (including one exact baseline) from the EvoApproxLib library. Results show that, without retraining, the Dense baseline is the most resilient topology across all CNN architectures, whereas on ViT-Small, all topologies degrade at comparable rates regardless of routing strategy. After approximate-aware retraining, recovery varies substantially across architectures, topologies, and multipliers. ResNet-20 achieves full recovery across the entire multiplier range, whereas VGG architectures recover at moderate multipliers but fail irreversibly at aggressive ones for all topologies except Cluster MoE on VGG11_bn; on ViT-Small, Hard MoE outperforms Dense under aggressive approximation at equal normalized inference cost. These results pave the way for future approximate MoE hardware-software co-design strategies.

In deep reinforcement learning (DRL), an agent is trained from a stream of experience. In a continual learning setting, such agents can suffer from plasticity loss: their ability to learn new skills from new experiences diminishes over training. Recently, Mixture-of-Experts (MoE) networks have been reported to enable scaling laws and facilitate the learning of diverse skills. However, in continual reinforcement learning settings, their performance can degenerate as learning proceeds, indicating a loss of plasticity. To address this, building on Neural Tangent Kernel (NTK) theory, we formalize the plasticity loss in MoE policies as a loss of spectral plasticity. We then derive a tractable proxy for spectral plasticity, one expressible in terms of individual expert feature matrices. Leveraging this proxy, we introduce SPHERE, a practical Parseval penalty tailored for MoE-based policies that alleviates the loss of spectral plasticity. On MetaWorld and HumanoidBench, SPHERE improves average success under continual RL by 133% and 50% over an unregularized MoE baseline, while maintaining higher spectral plasticity throughout training.

Dialogue State Tracking (DST) requires precise extraction of structured information from multi-domain conversations, a task where Large Language Models (LLMs) struggle despite their impressive general capabilities. We present GEM (Graph-Enhanced Mixture-of-Experts), a novel framework that combines language models and graph-structured dialogue understanding with ReAct agent-based reasoning for superior DST performance. Our approach dynamically routes between specialized experts: a Graph Neural Network that captures dialogue structure and turn-level dependencies, and a finetuned T5-Small encoder-decoder for sequence modeling, coordinated by an intelligent router. For complex value generation tasks, we integrate ReAct agents that perform structured reasoning over dialogue context. On MultiWOZ 2.2, GEM achieves 65.19% Joint Goal Accuracy, substantially outperforming end-to-end LLM approaches (best: 38.43%) and surpassing state-of-the-art (SOTA) methods including TOATOD (63.79%), D3ST (58.70%), and Diable (56.48%). Our graph-enhanced mixture-of-experts architecture with ReAct integration demonstrates that combining structured dialogue representation with dynamic expert routing and agent-based reasoning provides a powerful paradigm for dialogue state tracking, achieving superior accuracy while maintaining computational efficiency through selective expert activation.

Multimodal visual object tracking can be divided into to several kinds of tasks (e.g. RGB and RGB+X tracking), based on the input modality. Existing methods often train separate models for each modality or rely on pretrained models to adapt to new modalities, which limits efficiency, scalability, and usability. Thus, we introduce OneTrackerV2, a unified multi-modal tracking framework that enables end-to-end training for any modality. We propose Meta Merger to embed multi-modal information into a unified space, allowing flexible modality fusion and robustness. We further introduce Dual Mixture-of-Experts (DMoE): T-MoE models spatio-temporal relations for tracking, while M-MoE embeds multi-modal knowledge, disentangling cross-modal dependencies and reducing feature conflicts. With a shared architecture, unified parameters, and a single end-to-end training, OneTrackerV2 achieves state-of-the-art performance across five RGB and RGB+X tracking tasks and 12 benchmarks, while maintaining high inference efficiency. Notably, even after model compression, OneTrackerV2 retains strong performance. Moreover, OneTrackerV2 demonstrates remarkable robustness under modality-missing scenarios.