Mixture-of-Experts Models

We propose S3 (Specialization, Selection, Sparsification), a framework that rethinks multimodal learning through a structural perspective. Instead of encoding all signals into a fixed embedding, S3 decomposes multimodal inputs into semantic experts and selectively routes them for each task. Specialization forms concept-level experts in a shared latent space, Selection adapts routing for task-specific needs, and Sparsification prunes low-utility paths to yield compact, information-minimal representations. Across four MultiBench benchmarks, S3 improves accuracy and shows a consistent reverse U-shaped sparsity-performance trend, with peak performance at intermediate sparsity. These results suggest that structuring multimodal representations as selectable semantic components provides a practical and principled alternative to contrastive learning or InfoMax-driven approaches.

Encoder-decoder imbalance and the reliance on large input volumes make many 3D brain tumor segmentation models both compute-heavy and brittle. We present M\textsuperscript{4}Fuse, a lightweight network that prioritizes discriminative brain tumor cues over exhaustive appearance reconstruction. Our method balances encoder and decoder capacity and replaces depth expansion with a synergistic design: it propagates long-range context with linear complexity via a grouped state space mixer, denoises and aligns skip features using a cross-scale dual-stage gating bridge, and absorbs cross-site acquisition shifts with a sample-level mixture-of-experts. On the BraTS2019 and BraTS2021 benchmarks, M\textsuperscript{4}Fuse outperforms other lightweight excellent methods in both parameter count and performance. Even at a challenging input resolution of $64\times128\times128$ (half that of existing excellent models), M\textsuperscript{4}Fuse reduces parameters by 62.63% and improves average performance by 0.09%. Ablations of key components validate the method's exceptional parameter-to-accuracy efficiency and robustness across diverse data centers.

Production LLM workloads increasingly serve discriminative tasks, such as classification, recommendation, and verification, whose answers are read from the logits of a single prefill pass with no autoregressive decoding. Serving these prefill-only workloads on mixture-of-experts (MoE) models is bottlenecked not by compute but by the distributed execution required to fit the model: existing parallel strategies (tensor, expert, and pipeline parallelism) trade memory pressure for redundant computation, communication, and synchronization, severely degrading MoE prefill serving efficiency. We observe that these overheads stem from coupling expert placement with synchronous activation routing -- a design inherited from the decoding era. The long, compute-bound forward passes of large-batch prefill open a per-layer window wide enough to stream expert weights in the background, replacing per-layer activation AllToAll with asynchronous weight AllGather fully overlapped with computation. We propose MoE-Prefill, a prefill-only serving system whose backend, AsyncEP (Asynchronous Expert Parallelism), gathers experts by weight rather than routing them by activation, and whose frontend co-enforces a physically-derived saturation threshold through prefix-aware routing and true-FLOPs load tracking. On Qwen3-235B-A22B across four hardware/precision configurations, MoE-Prefill delivers 1.35-1.37x throughput over the strongest distributed baseline on real-world workloads and up to 1.59x on long-context synthetic workloads, sustaining 29.8-36.2% per-GPU model FLOPs utilization.

Sparse MoE routing fails at domain transitions, where the current token belongs to one distribution and the next to another. In a controlled experiment (4 experts, 5 seeds), standard affinity routing assigns only 0.006 +/- 0.001 probability to the correct expert at the transition. Three lightweight gate modifications raise this to 0.748 +/- 0.002 (124x), cutting experts needed for 99% coverage from infeasible to a small constant: temporal memory (beta), a per-expert LIF membrane potential accumulating routing context across tokens; precision-weighted gating (Pi), a per-expert inverse variance of recent prediction error, yielding 31x contrast between reliable and unreliable experts; and anticipatory routing, a next-state predictor conditioned on the beta-accumulated hidden state. The mechanisms draw from Friston's Free Energy Principle and use LIF dynamics from spiking neural networks. An ablation across all 2^3 subsets reveals a super-additive beta x Ant interaction: anticipation alone gives nothing (+0.000 +/- 0.001); beta alone gives modest gain (+0.295 +/- 0.013); combined they close 75% of the oracle gap (+0.741 +/- 0.002, exceeding the sum by +0.446 +/- 0.014). This is structural: a stateless predictor cannot detect approaching transitions because pre-transition tokens are distributionally identical to within-domain tokens. In a character-level MoE LM (5 seeds), beta-routing reduces transition-step BPC from 6.56 +/- 0.01 (Standard) to 4.01 +/- 0.15 (beta-MoE); the beta + Ant gate places 0.86 +/- 0.02 probability on the correct domain expert before that domain appears in input, vs 0.42 +/- 0.12 for Standard MoE. Reference implementations (~200 lines each): https://github.com/russellwmy/affinity-is-not-enough

Safety alignment is critical for the responsible deployment of large language models (LLMs). As Mixture-of-Experts (MoE) architectures are increasingly adopted to scale model capacity, understanding their safety robustness becomes essential. Existing adversarial attacks, however, have notable limitations. Prompt-based jailbreaks rely on heuristic search and transfer poorly, model intervention methods require privileged access to internal representations, and optimization-based input attacks remain output-centric and are fundamentally limited to MoE models due to the non-differentiable routing mechanism. In this paper, we present RouteHijack, a routing-aware jailbreak for MoE LLMs. Our key insight is that safety behavior is concentrated in a small subset of experts, creating an opportunity to steer model behavior by influencing routing decisions through input optimization. Building on this observation, RouteHijack first performs response-driven expert localization to identify safety-critical and harmful experts by contrasting activations under safe refusals and harmful completions. It then constructs adversarial suffixes with a routing-aware objective that suppresses safety experts, promotes harmful experts, and prevents early-stage refusal during generation. At inference time, the optimized suffix is appended to a malicious prompt, requiring only input access. Across seven MoE LLMs, RouteHijack achieves a 69.3% average attack success rate (ASR), outperforming prior optimization-based attack by $3.2\times$. RouteHijack also transfers zero-shot across five sibling MoE variants, raising average ASR from 27.7% to 61.2%, and further generalizes to three MoE-based VLMs, increasing average ASR from 2.47% to 38.7%. These findings expose a fundamental vulnerability in sparse expert architectures and highlight the need for defenses beyond output-level alignment.

Leveraging continuous solar energy harvesting at high efficiency, space data centers are envisioned as a promising platform for executing energy-intensive large language models (LLMs). Recognizing this advantage, space and AI conglomerates (e.g., SpaceX, Google) are actively investing in this vision. One key challenge, however, is the efficient distributed deployment of a large-scale LLM in a satellite network due to the limited onboard computing and communication resources. This gives rise to a placement problem that involves partitioning and mapping model components to satellites such that the fundamentally different model architecture and network topology can be reconciled to ensure low-latency token generation. To address this problem, we present the Space Network of Mixture-of-Experts (SpaceMoE) framework targeting the distributed execution of a popular mixture-of-experts (MoE) model in space. The proposed placement strategies are two-level: (1) layer placement, which assigns MoE layers to satellite subnets; and (2) intra-layer expert placement, which assigns individual experts to satellites associated with the same layer/subnet. For layer placement, we exploit the ring-like communication pattern of autoregressive inference to partition the satellite constellation along the orbiting direction into subnets arranged on a ring, each hosting one MoE layer. Based on this architecture, we formulate and solve an optimization problem for intra-layer expert placement to map experts with heterogeneous activation probabilities onto satellites. The derived strategy reveals an intuitive principle: a frequently activated expert should be mapped to a satellite on a routing path with low expected latency. Experiments over a thousand-satellite constellation show that SpaceMoE achieves at least a threefold latency reduction compared with conventional random and ablation-based placement strategies.

Tree-based speculative decoding accelerates autoregressive generation by verifying multiple draft candidates in parallel, but this advantage weakens for sparse Mixture-of-Experts (MoE) models. As the draft tree grows, different branches activate different experts, expanding the union of activated experts and substantially increasing target-side verification cost. We propose EVICT, a training-free, hyperparameter-free, and lossless adaptive verification method for MoE speculative decoding. EVICT makes every verified token count by truncating the draft tree before target verification and retaining only the cost-effective prefix. It leverages fine-grained drafter signals to estimate candidate benefit, combines them with offline-profiled verification cost, and remains highly compatible with the high-performance graph-based serving framework SGLang. Extensive experiments on diverse MoE backbones and benchmarks show that EVICT achieves up to 2.35x speedup over autoregressive decoding and an average 1.21x speedup over the state-of-the-art baseline EAGLE-3, while significantly reducing unnecessary expert activations during verification.

Mixture-of-experts (MoE) architectures have turned LLM serving into a cluster-scale workload in which communication consumes a considerable portion of LLM serving runtime. This has prompted industry to invest heavily in expensive high-bandwidth scale-up networks. We question whether such costly infrastructure is strictly necessary. We present the first systematic cross-layer analysis of network cost-effectiveness for MoE LLM serving, comparing four representative XPU (e.g., GPU/TPU) topologies (scale-up, scale-out, 3D torus, and 3D full-mesh). We find that lower-cost switchless topologies are more cost-effective than the scale-up topology across all serving scenarios explored, improving cost-effectiveness by 20.6-56.2%. In particular, the 3D full-mesh topology is Pareto-optimal in terms of the performance-cost tradeoff. We also find that current scale-up link bandwidths are over-provisioned: reducing the link bandwidth improves throughput per cost by up to 27%. A forward-looking analysis of upcoming GPU generations indicates that the cost-performance advantage of switchless networks will likely persist.

Mixture-of-Experts (MoE) models offer high capacity with efficient inference cost by activating a small subset of expert models per input. However, deploying MoE models requires all experts to reside in memory, creating a gap between the resource used by activated experts and the provisioned resources. This underutilization is further pronounced in multi-tenant scenarios. In this paper, we propose FaaSMoE, a multi-tenant MoE serving architecture built on Function-as-a-Service (FaaS) platforms. FaaSMoE decouples the control and execution planes of MoE by deploying experts as stateless FaaS functions, enabling on-demand and scale-to-zero expert invocation across tenants. FaaSMoE further supports configurable expert granularity within functions, trading off per-expert elasticity for reduced invocation overhead. We implement a prototype with an open-source edge-oriented FaaS platform and evaluate it using Qwen1.5-moe-2.7B under multi-tenant workloads. Compared to a full-model baseline, FaaSMoE uses less than one third of the resources, demonstrating a practical and resource-efficient path towards scalable MoE serving in a multi-tenant environment.

LoRA-MoE has emerged as an effective paradigm for parameter-efficient fine-tuning, combining the low training cost of LoRA with the increased adaptation capacity of Mixture-of-Experts (MoE). However, existing LoRA-MoE frameworks typically adopt a fixed and uniform expert configuration across heterogeneous Transformer modules (\eg, attention query/key projections and MLP gating networks), ignoring their distinct functional roles and capacity requirements. This design leads to localized over-provisioning, redundant trainable parameters, and unnecessary optimizer-state overhead. Moreover, prior methods enforce load balancing among experts throughout training. Although beneficial in the early stage, this constraint becomes restrictive once routing patterns stabilize, limiting expert specialization on downstream tasks. In this paper, we propose DMEP, a novel LoRA-MoE fine-tuning framework based on Dynamic Module-wise Expert Pruning. DMEP tracks expert utilization during training and physically removes low-utility experts on a per-module basis, yielding a more compact expert structure tailored to different modules. The pruned model then continues training without the load-balancing constraint, freeing the remaining experts to focus entirely on the downstream task and develop specialized expertise. By jointly adapting module-wise expert capacity and eliminating unnecessary balancing, DMEP improves both parameter efficiency and training efficiency. Extensive experiments on multiple reasoning benchmarks show that DMEP reduces trainable parameters by 35%--43% and improves training throughput by about 10%, while maintaining or surpassing the downstream reasoning accuracy of uniform LoRA-MoE baselines.

The optimal kernel configuration for Mixture-of-Experts (MoE) inference depends on both batch size and the expert routing distribution, yet production systems dispatch from batch size alone, leaving 10-70% of kernel throughput unrealized. We present RaMP, a routing-aware dispatch framework. A performance-region analysis derives, from hardware constants alone, when each optimization helps, correctly predicting all 8 tested architectures, including 3 unseen. A four-parameter wave cost model selects the fastest configuration from the runtime expert histogram, achieving 0.93% mean regret versus exhaustive search, fitted from just 10-24 minutes of one-time profiling per model. Because the model depends only on CTA grid geometry, it is kernel-agnostic: applied to Alpha-MoE, it delivers 1.14x with no source modification. Paired with a co-designed CuTe DSL kernel exposing 134-268 polymorphic configurations, RaMP delivers 1.22x kernel speedup over static dispatch and 1.30x end-to-end speedup in vLLM serving over Triton, 1.41x over DeepGEMM, and 1.13x over FlashInfer CUTLASS.

We present Marco-MoE, a suite of fully open multilingual sparse Mixture-of-Experts (MoE) models. Marco-MoE features a highly sparse design in which only around 5% of the total parameters are activated per input token. This extreme sparsity, combined with upcycling from dense models, enables efficient pre-training on 5T tokens. Our models surpass similarly-sized competitors on English and multilingual benchmarks, achieving a best-in-class performance-to-compute ratio. We further post-train these models to create Marco-MoE-\textsc{Instruct} variants, which surpass the performance of competing models possessing $3$--$14\times$ more activated parameters. Our analysis reveals that Marco-MoE learns structured expert activation patterns shared across related languages, while maintaining highly specialized utilization for linguistically isolated ones. We further show that Marco-MoE allows for scalable language expansion without the interference typical of dense models. To support the community, we disclose our full training datasets, recipes, and model weights.

Diffusion models have achieved success in high-fidelity data synthesis, yet their capacity for more complex, structured reasoning like text following tasks remains constrained. While advances in language models have leveraged strategies such as latent reasoning and recursion to enhance text understanding capabilities, extending these to multimodal text-to-image generation tasks is challenging due to the continuous and non-discrete nature of visual tokens. To tackle this problem, we draw inspiration from modular human cognition and propose a recursive, sparse mixture-of-experts framework integrated into conventional diffusion models. Our approach introduces a recursive component within joint attention layers that iteratively refines visual tokens over multiple latent steps while efficiently sharing parameters via sparse selection of neural modules. At each step, a gating network is devised to dynamically select specialized neural modules, conditioned on the current visual tokens, the diffusion timestep, and the conditioning information. Comprehensive evaluation on class-conditioned ImageNet image generation tasks and additional studies on the GenEval and DPG benchmark demonstrate the superiority of the proposed method in enhancing model image generation performance.

We present Move-Then-Operate, a Vision language action framework that explicitly decouples robotic manipulation into two distinct behavioral phases: coarse relocation (move) and contact-critical interaction (operate). Unlike monolithic policies that conflate these heterogeneous regimes, our architecture employs a dual-expert policy routed by a learnable phase selector, introducing a structural inductive bias that isolates phase-specific dynamics. Phase labels are automatically generated via an MLLM-based pipeline conditioned on lightweight contextual cues such as end-effector velocity and subtask decomposition to ensure alignment with human motor patterns. Evaluated on the RoboTwin2 benchmark, our method achieves an average success rate of $68.9\%$, outperforming the monolithic $π_0$ baseline by $24\%$. It matches or exceeds models trained on $10\times$ more data and reaches peak performance in $40\%$ fewer training steps, demonstrating that architectural disentanglement of move and operate phases is a highly effective and efficient strategy for mastering high-precision manipulation.

Most recent state-of-the-art (SOTA) large language models (LLMs) use Mixture-of-Experts (MoE) architectures to scale model capacity without proportional per-token compute, enabling higher-quality outputs at manageable serving costs. However, MoE inference at scale is fundamentally bottlenecked by expert load imbalance and inefficient token routing, especially in multi-node deployments where tokens are not guaranteed to be routed to local experts, resulting in significant inter-node all-to-all communication overhead. To systematically characterize these challenges, we profile SOTA open-source MoE models, including Llama 4 Maverick, DeepSeek V3-671B, and Qwen3-230B-A22B, on various datasets and collected over 100k real expert activation traces. Upon studying the expert activation patterns, we uncover various persistent properties across all the frontier MoE models: variable expert load imbalance, domain-specific expert activation where expert popularity shifts across task families (code, math, chat, general), and a strong correlation between prefill and decode expert activations. Motivated by these findings, we propose workload-aware micro-batch grouping and an expert placement strategy to maximize token locality to the destination expert, thereby reducing inter-node communication. Across models and datasets, these optimizations help reduce all2all communication data up to 20, resulting in lower MoE decode latency and better accelerator utilization.

Large Language Models (LLMs) based on Mixture-of-Experts (MoE) are pivotal in industrial applications for their ability to scale performance efficiently. However, standard MoEs enforce uniform expert sizes,creating a rigidity that fails to align computational costs with varying token-level complexity. While heterogeneous expert architectures attempt to address this by diversifying expert sizes, they often suffer from significant system-level challenges, specifically unbalanced GPU utilization and inefficient parameter utilization, which hinder practical deployment. To bridge the gap between theoretical heterogeneity and robust industrial application, we propose Mixture of Heterogeneous Grouped Experts (MoHGE) which introduces a two-level routing mechanism to enable flexible, resource-aware expert combinations. To optimize inference efficiency, we propose a Group-Wise Auxiliary Loss, which dynamically steers tokens to the most parameter-efficient expert groups based on task difficulty. To address the critical deployment challenge of GPU load balancing, we introduce an All-size Group-decoupling Allocation strategy coupled with an Intra-Group Experts Auxiliary Loss. These mechanisms collectively ensure uniform computation distribution across GPUs. Extensive evaluations demonstrate that MoHGE matches the performance of MoE architectures while reducing the total parameters by approximately 20% and maintaining balanced GPU utilization. Our work establishes a scalable paradigm for resource-efficient MoE design, offering a practical solution for optimizing inference costs in real-world scenarios. The code is publicly available at https://github.com/UnicomAI/MoHGE.

Despite MoE models leading many benchmarks, supervised fine-tuning (SFT) for the MoE architectures remains difficult because its router layers are fragile. Methods such as DenseMixer and ESFT mitigate router collapse with dense mixing or auxiliary load-balancing losses, but these introduce noisy gradients that often degrade performance. In preliminary experiments, we systematically pruned experts and observed that while certain super experts are activated far more frequently, discarding less used experts still leads to notable performance degradation. This suggests that even rarely activated experts encode non-trivial knowledge useful for downstream tasks. Motivated by this, we propose an auxiliary-loss-free MoE SFT framework that combines bias-driven sparsification with always-active gated condenser experts. Rather than enforcing balanced activation across all experts, our method encourages task-relevant experts to remain active while pushing long-tailed experts toward inactivity. The condenser experts provide a persistent, learnable pathway that alleviates gradient starvation and facilitates consolidation of information that would otherwise remain fragmented across sparsely activated experts. Analysis further suggest that this design better preserves long-tailed expert information under sparse routing. Experiments on large-scale MoE models demonstrate that our approach outperforms state-of-the-art SFT baselines such as DenseMixer and ESFT, achieving average gain of 2.5%+ on both mathematical reasoning and commonsenseQA benchmarks.

Continual Model Merging (CMM) sequentially integrates task-specific models into a unified architecture without intensive retraining. However, existing CMM methods are hindered by a fundamental saturation-redundancy dilemma: backbone-centric approaches face parameter saturation and representation interference within fixed capacities, whereas Mixture-of-Experts (MoE) variants resort to indiscriminate expansion, incurring expert redundancy and a routing bottleneck reliant on additional data-driven optimization. To resolve these challenges, we propose MADE-IT (Manifold-Aware Dynamic Expert Evolution and Implicit rouTing), an adaptive CMM method that orchestrates expert management and activation by grounding intrinsic expert representations in manifold geometry. We introduce a projection-based subspace affinity metric coupled with a distribution-aware adaptive threshold mechanism to guide autonomous expert evolution, harmonizing diversity with architectural parsimony. Furthermore, to bypass parameterized gating networks, we design a data-free and training-free implicit routing mechanism that activates experts via feature-subspace alignment. Extensive experiments demonstrate that MADE-IT consistently outperforms strong baselines in accuracy and robustness across long-horizon and shuffled task sequences, while significantly pruning redundant experts, particularly within generic modules and early layers.

Current model training approaches incorporate user information directly into shared weights, making individual data removal computationally infeasible without retraining. This paper presents a three-layer architecture that decouples personal data from shared weights by combining a static base model, composable domain-expert LoRA adapters that shape behavior without imparting user data, and per-user proxy artefacts whose deletion constitutes deterministic unlearning. Evaluation on Phi-3.5-mini and Llama-3.1-8B confirms per-user differentiation in which personal data influences outputs while remaining isolated, verified by a return to baseline after proxy removal (KL divergence of approximately 0.21 nats, 82-89% verification pass rate) and near-zero cross-user contamination. Because user-specific information never enters shared weights, the architecture mitigates model inversion, membership inference, and training-data extraction against shared model components by construction. The approach converts machine unlearning from an intractable weight-editing problem into a deterministic deletion operation that preserves personalization alongside privacy-enhancing guarantees and is compatible with differentially private stochastic gradient descent (DP-SGD) for privacy-preserving shared model improvement.

Nowadays, visual data forgery detection plays an increasingly important role in social and economic security with the rapid development of generative models. Existing face forgery detectors still can't achieve satisfactory performance because of poor generalization ability across datasets. The key factor that led to this phenomenon is the lack of suitable metrics: the commonly used cross-dataset AUC metric fails to reveal an important issue where detection scores may shift significantly across data domains. To explicitly evaluate cross-domain score comparability, we propose \textbf{Cross-AUC}, an evaluation metric that can compute AUC across dataset pairs by contrasting real samples from one dataset with fake samples from another (and vice versa). It is interesting to find that evaluating representative detectors under the Cross-AUC metric reveals substantial performance drops, exposing an overlooked robustness problem. Besides, we also propose the novel framework \textbf{S}emantic \textbf{F}ine-grained \textbf{A}lignment and \textbf{M}ixture-of-Experts (\textbf{SFAM}), consisting of a patch-level image-text alignment module that enhances CLIP's sensitivity to manipulation artifacts, and the facial region mixture-of-experts module, which routes features from different facial regions to specialized experts for region-aware forgery analysis. Extensive qualitative and quantitative experiments on the public datasets prove that the proposed method achieves superior performance compared with the state-of-the-art methods with various suitable metrics.

Recent progress in deep learning has been driven by increasingly large-scale models, but the resulting computational cost has become a critical bottleneck. Sparse Mixture of Experts (MoE) offers an effective solution by activating only a small subset of experts for each input, achieving high scalability without sacrificing inference speed. Although effective, sparse MoE training exhibits characteristic optimization difficulties. Because the router receives informative gradients only through the experts selected in the forward pass, it suffers from gradient blocking and obtains little information from unselected routes. This limited, highly localized feedback makes it difficult for the router to learn appropriate expert-selection scores and often leads to unstable routing dynamics, such as fluctuating expert assignments during training. To address this issue, we propose TGR-MoE: Teacher-Guided Routing for Sparse Vision Mixture-of-Experts, a simple yet effective method that stabilizes router learning using supervision derived from a pretrained dense teacher model. TGR-MoE constructs a teacher router from the teacher's intermediate representations and uses its routing outputs as pseudo-supervision for the student router, suppressing frequent routing fluctuations during training and enabling knowledge-guided expert selection from the early stages of training. Extensive experiments on ImageNet-1K and CIFAR-100 demonstrate that TGR consistently improves both accuracy and routing consistency, while maintaining stable training even under highly sparse configurations.

Person-Job Fit (PJF) is a critical component for online recruitment. Existing approaches face several challenges, particularly in handling low-quality job descriptions and similar candidate-job pairs, which impair model performance. To address these challenges, this paper proposes a large language model (LLM) based method with two novel techniques: (1) LLM-based data augmentation, which polishes and rewrites low-quality job descriptions by leveraging chain-of-thought (COT) prompts, and (2) category-aware Mixture of Experts (MoE) that assists in identifying similar candidate-job pairs. This MoE module incorporates category embeddings to dynamically assign weights to the experts and learns more distinguishable patterns for similar candidate-job pairs. We perform offline evaluations and online A/B tests on our recruitment platform. Our method relatively surpasses existing methods by 2.40% in AUC and 7.46% in GAUC, and boosts click-through conversion rate (CTCVR) by 19.4% in online tests, saving millions of CNY in external headhunting expenses.

Mixture-of-experts models provide a flexible framework for learning complex probabilistic input-output relationships by combining multiple expert models through an input-dependent gating mechanism. These models have become increasingly prominent in modern machine learning, yet their theoretical properties in the Bayesian framework remain largely unexplored. In this paper, we study Bayesian mixture-of-experts models, focusing on the ubiquitous softmax-based gating mechanism. Specifically, we investigate the asymptotic behavior of the posterior distribution for three fundamental statistical tasks: density estimation, parameter estimation, and model selection. First, we establish posterior contraction rates for density estimation, both in the regimes with a fixed, known number of experts and with a random learnable number of experts. We then analyze parameter estimation and derive convergence guarantees based on tailored Voronoi-type losses, which account for the complex identifiability structure of mixture-of-experts models. Finally, we propose and analyze two complementary strategies for selecting the number of experts. Taken together, these results provide one of the first systematic theoretical analyses of Bayesian mixture-of-experts models with softmax gating, and yield several theory-grounded insights for practical model design.

Mixture-of-Experts models, now popular for scaling capacity at fixed inference speed, switch experts at nearly every token. Once a model outgrows available GPU memory, this churn can render optimizations like offloading and pre-fetching ineffective. We make the case that the options framework in reinforcement learning is a perfect match to tackle this problem, and argue for temporally extended mixture-of-experts layers. Building on the option-critic framework with deliberation costs, we add a controller to each layer that learns when to switch expert sets and which to load. By applying this to gpt-oss-20b with low-rank adapters and a self-distillation reward, our method reduces switch rates from over 50% to below 5% while retaining up to 90% of base-model accuracy on MATH, MMLU, and MMMLU. This shows that even existing pre-trained models can be converted to temporally extended MoEs with lightweight training, with the deliberation cost allowing model trainers to trade off switching rates against capability. We hope this opens a principled path, grounded in the options framework, for memory-efficient serving and continual learning in ever-growing MoE models.

Robotic parkour provides a compelling benchmark for advancing locomotion over highly challenging terrain, including large discontinuities such as elevated steps. Recent approaches have demonstrated impressive capabilities, including dynamic climbing and jumping, but typically rely on sequential multilayer perceptron (MLP) architectures with densely activated layers. In contrast, sparsely gated mixture-of-experts (MoE) architectures have emerged in the large language model domain as an effective paradigm for improving scalability and performance by activating only a subset of parameters at inference time. In this work, we investigate the application of sparsely gated MoE architectures to vision-based robotic parkour. We compare control policies based on standard MLPs and MoE architectures under a controlled setting where the number of active parameters at inference time is matched. Experimental results on a real Unitree Go2 quadruped robot demonstrate clear performance gains, with the MoE policy achieving double the number of successful trials in traversing large obstacles compared to a standard MLP baseline. We further show that achieving comparable performance with a standard MLP requires scaling its parameter count to match that of the total MoE model, resulting in a 14.3% increase in computation time. These results highlight that sparsely gated MoE architectures provide a favorable trade-off between performance and computational efficiency, enabling improved scaling of control policies for vision-based robotic parkour. An anonymized link to the codebase is https://osf.io/v2kqj/files/github?view_only=7977dee10c0a44769184498eaba72e44.

Mixture-of-Experts (MoE) has become the dominant architecture for scaling large language models: frontier models routinely decouple total parameters from per-token computation through sparse expert routing. Scaling laws show that under fixed active computation, model quality scales predictably with total parameters, and MoEs realize this by increasing expert count. However, training large MoEs is expensive, as memory requirements and inter-device communication both scale with total parameter count. We propose expert upcycling, a method for progressively expanding MoE capacity by increasing the number of experts during continued pre-training (CPT). Given a trained E-expert model, the upcycling operator constructs an mE-expert model through expert duplication and router extension while holding top-K routing fixed, preserving per-token inference cost. Duplication provides a warm initialization: the expanded model inherits the source checkpoint's learned representations, starting from a substantially lower loss than random initialization. Subsequent CPT then breaks the symmetry among duplicated experts to drive specialization. We formalize the upcycling operator and develop a theoretical framework decomposing the quality gap into a capacity term and an initialization term. We further introduce utility-based expert selection, which uses gradient-based importance scores to guide non-uniform duplication, more than tripling gap closure when CPT is limited. In our 7B-13B total parameter experiments, the upcycled model matches the fixed-size baseline on validation loss while saving 32% of GPU hours. Comprehensive ablations across model scales, activation ratios, MoE architectures, and training budgets yield a practical recipe for deploying expert upcycling, establishing it as a principled, compute-efficient alternative to training large MoE models from scratch.

Human perception generalizes well across different domains, but most vision models struggle beyond their training data. This gap motivates multi-dataset learning, where a single model is trained on diverse datasets to improve robustness under domain shifts. However, unified training remains challenging due to inconsistencies in data distributions and label semantics. Mixture-of-Experts (MoE) models provide a scalable solution by routing inputs to specialized subnetworks (experts). Yet, existing MoEs often fail to specialize effectively, as their load-balancing mechanisms enforce uniform input distribution across experts. This fairness conflicts with domain-aware routing, causing experts to learn redundant representations, and reducing performance especially on rare or out-of-distribution domains. We propose GEM (Global Expert Mapping), a planner-compiler framework that replaces the learned router with a global scheduler. Our planner, based on linear programming relaxation, computes a fractional assignment of datasets to experts, while the compiler applies hierarchical rounding to convert this soft plan into a deterministic, capacity-aware mapping. Unlike prior MoEs, GEM avoids balancing loss, resolves the conflict between fairness and specialization, and produces interpretable routing. Experiments show that GEM-DINO achieves state-of-the-art performance on the UODB benchmark, with notable gains on underrepresented datasets and solves task interference in few-shot adaptation scenarios.

Apple Neural Engine (ANE) is a dedicated neural processing unit (NPU) present in every Apple Silicon chip. Mixture-of-Experts (MoE) LLMs improve inference efficiency via sparse activation but are challenging for NPUs in three ways: expert routing is unpredictable and introduces dynamic tensor shapes that conflict with the shape-specific constraints of NPUs; several irregular operators, e.g., top-k, scatter/gather, etc., are not NPU-friendly; and launching many small expert kernels incurs substantial dispatch and synchronization overhead. NPUs are designed to offload AI compute from CPU and GPU; our goal is to enable such offloading for MoE inference, particularly during prefill, where long-context workloads consume substantial system resources. This paper presents NPUMoE, a runtime inference engine that accelerates MoE execution on Apple Silicon by offloading dense, static computation to NPU, while preserving a CPU/GPU fallback path for dynamic operations. NPUMoE uses offline calibration to estimate expert capacity and popularity that drives three key techniques: (1) Static tiers for expert capacity to address dynamic expert routing (2) Grouped expert execution to mitigate NPU concurrency limits (3) Load-aware expert compute graph residency to reduce CPU-NPU synchronization overhead. Experiments on Apple M-series devices using three representative MoE LLMs and four long-context workloads show that NPUMoE consistently outperforms baselines, reducing latency by 1.32x-5.55x, improving energy efficiency by 1.81x-7.37x, and reducing CPU-cycle usage by 1.78x-5.54x through effective NPU offloading.

Extending a fully post-trained language model with new domain capabilities is fundamentally limited by monolithic training paradigms: retraining from scratch is expensive and scales poorly, while continued training often degrades existing capabilities. We present BAR (Branch-Adapt-Route), which trains independent domain experts, each through its own mid-training, supervised finetuning, and reinforcement learning pipeline, and composes them via a Mixture-of-Experts architecture with lightweight router training. Unlike retraining approaches that mix all domains and require full reprocessing for any update (with cost scaling quadratically), BAR enables updating individual experts independently with linear cost scaling and no degradation to existing domains. At the 7B scale, with experts for math, code, tool use, and safety, BAR achieves an overall score of 49.1 (averaged across 7 evaluation categories), matching or exceeding re-training baselines (47.8 without mid-training, 50.5 with). We further show that modular training provides a structural advantage: by isolating each domain, it avoids the catastrophic forgetting that occurs when late-stage RL degrades capabilities from earlier training stages, while significantly reducing the cost and complexity of updating or adding a domain. Together, these results suggest that decoupled, expert-based training is a scalable alternative to monolithic retraining for extending language models.

An LLM's residual stream is both state and instruction: it encodes the current context and determines the next transformation. We introduce a parameter-free decomposition for Mixture-of-Experts models that splits each layer's hidden state into a control signal that causally drives routing and an orthogonal content channel invisible to the router. Across six MoE architectures, we find that models preserve surface-level features (language, token identity, position) in the content channel, while the control signal encodes an abstract function that rotates from layer to layer. Because each routing decision is low-bandwidth, this hand-off forces compositional specialization across layers. While individual experts remain polysemantic, expert paths become monosemantic, clustering tokens by semantic function across languages and surface forms. The same token (e.g., ":") follows distinct trajectories depending on whether it serves as a type annotation, an introductory colon, or a time separator. Our decomposition identifies the source of this structure: clusters in the control subspace are substantially more monosemantic than those in the full representation. As a result, the natural unit of interpretability in MoEs is not the expert but the trajectory.

Binge eating disorder (BED) is the most prevalent eating disorder. However, current diagnostic frameworks remain largely grounded in symptom-based criteria rather than underlying biological mechanisms, thereby limiting early detection and the development of biologically-informed interventions. Emerging studies have begun to investigate the neurobiological signatures of BED, yet their findings are often difficult to generalize due to the reliance on hypothesis-driven parametric models, single-modality analyses, and limited data diversity. Therefore, there is a critical need for advanced data-driven frameworks capable of modeling multimodal data to uncover generalizable and biologically meaningful signatures of BED. In this study, we propose the Interpretable Modality-Aware Mixture-of-Experts (IMA-MoE), a novel architecture designed to integrate heterogeneous neuroimaging, behavioral, hormonal, and demographic measures within a unified predictive framework. By encoding each measure as a distinct token, IMA-MoE enables flexible modeling of cross-modal dependencies while preserving modality-specific characteristics. We further introduce a token-importance mechanism to enhance interpretability by quantifying the contribution of each measure to model predictions. Evaluated on the large-scale Adolescent Brain Cognitive Development (ABCD) dataset, IMA-MoE demonstrates superior performance in differentiating BED from healthy controls compared with baseline methods, while revealing sex-specific predictive patterns, with hormonal measures contributing more prominently to prediction in females. Collectively, these findings highlight the promise of interpretable, data-driven multimodal modeling in advancing biologically-informed characterization of BED and facilitating more precise and personalized interventions in neuropsychiatric disorders.

We present a Mixture-of-Experts-based foundation model applied to the GlueX DIRC detector at Jefferson Lab, demonstrating its utility as a unified framework for fast simulation, particle identification, and hit-level noise filtering of Cherenkov photons. By leveraging a single shared transformer backbone across all tasks, the approach eliminates the fragmentation of task-specific pipelines while maintaining competitive-and in several cases superior-performance relative to established methods. The model operates directly on low-level detector inputs, performing hit-by-hit autoregressive generation over split spatial and temporal vocabularies with continuous kinematic conditioning, and supports class-conditional generation of pions and kaons through its Mixture-of-Experts architecture. We benchmark against the standard geometrical reconstruction and prior deep learning methods across the full kinematic phase space of the GlueX DIRC, demonstrating that the foundation model framework transfers effectively to this detector without architectural modification. This work positions the foundation model as a practical and scalable alternative to the suite of task-specific models currently proposed for GlueX DIRC analysis.

Large language models (LLMs) often exhibit undesirable behaviors, such as safety violations and hallucinations. Although inference-time steering offers a cost-effective way to adjust model behavior without updating its parameters, existing methods often fail to be simultaneously effective, utility-preserving, and training-efficient due to their rigid, one-size-fits-all designs and limited adaptability. In this work, we present FineSteer, a novel steering framework that decomposes inference-time steering into two complementary stages: conditional steering and fine-grained vector synthesis, allowing fine-grained control over when and how to steer internal representations. In the first stage, we introduce a Subspace-guided Conditional Steering (SCS) mechanism that preserves model utility by avoiding unnecessary steering. In the second stage, we propose a Mixture-of-Steering-Experts (MoSE) mechanism that captures the multimodal nature of desired steering behaviors and generates query-specific steering vectors for improved effectiveness. Through tailored designs in both SCS and MoSE, FineSteer maintains robust performance on general queries while adaptively optimizing steering vectors for targeted inputs in a training-efficient manner. Extensive experiments on safety and truthfulness benchmarks show that FineSteer outperforms state-of-the-art methods in overall performance, achieving stronger steering performance with minimal utility loss. Code is available at https://github.com/YukinoAsuna/FineSteer

Educational assistants should spend more computation only when the task needs it. This paper rewrites our earlier draft around the system that was actually implemented and archived in the repository: a sample-level 1B to 7B cascade for the shared-8 EduBench benchmark. The final system, Pangu-ACE, uses a 1B tutor-router to produce a draft answer plus routing signals, then either accepts the draft or escalates the sample to a 7B specialist prompt. We also correct a major offline evaluation bug: earlier summaries over-credited some open-form outputs that only satisfied superficial format checks. After CPU-side rescoring from saved prediction JSONL, the full Chinese test archive (7013 samples) shows that cascade_final improves deterministic quality from 0.457 to 0.538 and format validity from 0.707 to 0.866 over the legacy rule_v2 system while accepting 19.7% of requests directly at 1B. Routing is strongly task dependent: IP is accepted by 1B 78.0% of the time, while QG and EC still escalate almost always. The current archived deployment does not yet show latency gains, so the defensible efficiency story is routing selectivity rather than wall-clock speedup. We also package a reproducible artifact-first paper workflow and clarify the remaining external-baseline gap: GPT-5.4 re-judging is implemented locally, but the configured provider endpoint and key are invalid, so final sampled-baseline alignment with GPT-5.4 remains pending infrastructure repair.

Mixture-of-Experts (MoE) models have become the dominant architecture for large-scale language models, yet on-premises serving remains fundamentally memory-bound as batching turns sparse per-token compute into dense memory activation. Memory-centric architectures (PIM, NMP) improve bandwidth but leave compute underutilized under MoE's low arithmetic intensity at high batch sizes. Speculative decoding (SD) trades idle compute for fewer target invocations, yet verification must load experts even for rejected tokens, severely limiting its benefit in MoE especially at low batch sizes. We propose ELMoE-3D, a hybrid-bonding (HB)-based HW-SW co-designed framework that unifies cache-based acceleration and speculative decoding to offer overall speedup across batch sizes. We identify two intrinsic elasticity axes of MoE-expert and bit-and jointly scale them to construct Elastic Self-Speculative Decoding (Elastic-SD), which serves as both an expert cache and a strongly aligned self-draft model accelerated by high HB bandwidth. Our LSB-augmented bit-sliced architecture exploits inherent redundancy in bit-slice representations to natively support bit-nested execution. On our 3D-stacked hardware, ELMoE-3D achieves an average $6.6\times$ speedup and $4.4\times$ energy efficiency gain over naive MoE serving on xPU across batch sizes 1-16, and delivers $2.2\times$ speedup and $1.4\times$ energy efficiency gain over the best-performing prior accelerator baseline.

Expert specialization is fundamental to Mixture-of-Experts (MoE) model success, yet existing metrics (cosine similarity, routing entropy) lack theoretical grounding and yield inconsistent conclusions under reparameterization. We present an information-geometric framework providing the first rigorous characterization of MoE specialization dynamics. Our key insight is that expert routing distributions evolve on the probability simplex equipped with the Fisher information metric, enabling formal analysis via Riemannian geometry. We prove that standard heuristic metrics violate parameterization invariance (Theorem 1), establish that specialization corresponds to geodesic flow with quantified approximation bounds (Theorem 2), and derive a failure predictor with theoretical threshold justification (Theorem 3). The framework introduces two principled metrics: Fisher Specialization Index (FSI) achieving r=0.91+/-0.02 correlation with downstream performance, and Fisher Heterogeneity Score (FHS) predicting training failure at 10% completion with AUC=0.89+/-0.03 -- outperforming validation-loss-based early stopping by 23% while requiring 40x fewer compute cycles. We validate intervention protocols achieving 87% recovery rate when FHS>1 is detected. Comprehensive experiments across language modeling (WikiText-103, C4), vision MoE (ImageNet), and scaling studies (8-64 experts, 125M-2.7B parameters) validate our theoretical predictions.

Empowering Large Multimodal Models (LMMs) with image generation often leads to catastrophic forgetting in understanding tasks due to severe gradient conflicts. While existing paradigms like Mixture-of-Transformers (MoT) mitigate this conflict through structural isolation, they fundamentally sever cross-modal synergy and suffer from capacity fragmentation. In this work, we present Symbiotic-MoE, a unified pre-training framework that resolves task interference within a native multimodal Mixture-of-Experts (MoE) Transformers architecture with zero-parameter overhead. We first identify that standard MoE tuning leads to routing collapse, where generative gradients dominate expert utilization. To address this, we introduce Modality-Aware Expert Disentanglement, which partitions experts into task-specific groups while utilizing shared experts as a multimodal semantic bridge. Crucially, this design allows shared experts to absorb fine-grained visual semantics from generative tasks to enrich textual representations. To optimize this, we propose a Progressive Training Strategy featuring differential learning rates and early-stage gradient shielding. This mechanism not only shields pre-trained knowledge from early volatility but eventually transforms generative signals into constructive feedback for understanding. Extensive experiments demonstrate that Symbiotic-MoE achieves rapid generative convergence while unlocking cross-modal synergy, boosting inherent understanding with remarkable gains on MMLU and OCRBench.

Mixture-of-Experts (MoE) models exhibit striking performance disparities across languages, yet the internal mechanisms driving these gaps remain poorly understood. In this work, we conduct a systematic analysis of expert routing patterns in MoE models, revealing a phenomenon we term Language Routing Isolation, in which high- and low-resource languages tend to activate largely disjoint expert sets. Through layer-stratified analysis, we further show that routing patterns exhibit a layer-wise convergence-divergence pattern across model depth. Building on these findings, we propose RISE (Routing Isolation-guided Subnetwork Enhancement), a framework that exploits routing isolation to identify and adapt language-specific expert subnetworks. RISE applies a tripartite selection strategy, using specificity scores to identify language-specific experts in shallow and deep layers and overlap scores to select universal experts in middle layers. By training only the selected subnetwork while freezing all other parameters, RISE substantially improves low-resource language performance while preserving capabilities in other languages. Experiments on 10 languages demonstrate that RISE achieves target-language F1 gains of up to 10.85% with minimal cross-lingual degradation.

Large Language Models (LLMs) have demonstrated advanced capabilities but often suffer from factual inaccuracies (hallucinations) and systematic biases. These issues, sometimes amplified in specific architectures like Mixture-of-Experts (MoE) which motivate our work, pose risks for reliable deployment. To address these challenges, we propose the Council Mode, a multi-agent consensus framework. Our approach dispatches queries to multiple heterogeneous frontier LLMs in parallel and synthesizes their outputs using a dedicated consensus model. The pipeline consists of three phases: an intelligent triage for query complexity, parallel generation across diverse models, and a structured synthesis that identifies agreement, disagreement, and unique findings. In our evaluation, conducted under controlled no-web settings, the Council Mode achieved a 41.7% relative reduction in hallucination rates on a 1,200-sample HaluEval subset and a 7.5-point improvement on TruthfulQA compared to the top-performing individual model. On our curated MDR-500 multi-domain reasoning benchmark, the Council Mode achieved a Quality Score of 95.4%, representing a 9.2-point improvement over the best individual model. The framework also exhibited lower measured bias variance under our rubric-based evaluation protocol. We provide a cost-effectiveness analysis showing that the framework incurs a 4.2x token-cost overhead, making it most suitable for accuracy-prioritized applications where the cost of errors exceeds the added inference cost. These findings suggest that structured multi-agent consensus is a promising direction for enhancing the reliability and factual grounding of LLM-generated content.

Actor-level stance detection aims to determine an author expressed position toward specific geopolitical actors mentioned or implicated in a text. Although transformer-based models have achieved relatively good performance in stance classification, they typically rely on unified representations that may not sufficiently capture heterogeneous linguistic signals, such as contrastive discourse structures, framing cues, and salient lexical indicators. This motivates the need for adaptive architectures that explicitly model diverse stance-expressive patterns. In this paper, we propose StanceMoE, a context-enhanced Mixture-of-Experts (MoE) architecture built upon a fine-tuned BERT encoder for actor-level stance detection. Our model integrates six expert modules designed to capture complementary linguistic signals, including global semantic orientation, salient lexical cues, clause-level focus, phrase-level patterns, framing indicators, and contrast-driven discourse shifts. A context-aware gating mechanism dynamically weights expert contributions, enabling adaptive routing based on input characteristics. Experiments are conducted on the StanceNakba 2026 Subtask A dataset, comprising 1,401 annotated English texts where the target actor is implicit in the text. StanceMoE achieves a macro-F1 score of 94.26%, outperforming traditional baselines, and alternative BERT-based variants.

Mixture-of-Experts (MoE) layers increase model capacity by activating only a small subset of experts per token, and typically rely on a learned router to map hidden states to expert assignments. In this work, we ask whether a dedicated learned router is strictly necessary for MoE routing. We propose Self-Routing, a parameter-free routing mechanism that uses a designated subspace of the token hidden state directly as expert logits, eliminating the router projection entirely while leaving the rest of the MoE layer unchanged. We evaluate Self-Routing on language modeling across different expert counts and model scales, and on ImageNet-1K classification by comparing it against a standard learned router, random-routing baselines, and dense non-MoE baselines. Our results show that Self-Routing remains competitive with the learned-router baseline while removing all dedicated routing parameters, and yields more balanced expert utilization, with about 17 % higher average normalized routing entropy and no explicit load-balancing loss. On ImageNet-1K with DeiT-S/16, Self-Routing also slightly improves over the corresponding learned-router MoE. These findings suggest that effective MoE routing can emerge from the hidden representation itself without requiring a separate learned router module.

Mixture-of-Experts (MoE) models substantially improve performance by increasing the capacity of dense architectures. However, directly training MoE models requires considerable computational resources and introduces extra overhead in parameter storage and deployment. Therefore, it is critical to develop an approach that leverages the multi-expert capability of MoE to enhance performance while incurring minimal additional cost. To this end, we propose a novel pre-training approach, termed ExFusion, which improves the efficiency of Transformer training through multi-expert fusion. Specifically, during the initialization phase, ExFusion upcycles the feed-forward network (FFN) of the Transformer into a multi-expert configuration, where each expert is assigned a weight for later parameter fusion. During training, these weights allow multiple experts to be fused into a single unified expert equivalent to the original FFN, which is subsequently used for forward computation. As a result, ExFusion introduces multi-expert characteristics into the training process while incurring only marginal computational cost compared to standard dense training. After training, the learned weights are used to integrate multi-experts into a single unified expert, thereby eliminating additional overhead in storage and deployment. Extensive experiments on a variety of computer vision and natural language processing tasks demonstrate the effectiveness of the proposed method.

Sparse Mixture-of-Experts (SMoE) architectures are increasingly used to scale large language models efficiently, delivering strong accuracy under fixed compute budgets. However, SMoE models often suffer from severe load imbalance across experts, where a small subset of experts receives most tokens while others are underutilized. Prior work has focused mainly on training-time solutions such as routing regularization or auxiliary losses, leaving inference-time behavior, which is critical for deployment, less explored. We present a systematic analysis of expert routing during inference and identify three findings: (i) load imbalance persists and worsens with larger batch sizes, (ii) selection frequency does not reliably reflect expert importance, and (iii) overall expert workload and importance can be estimated using a small calibration set. These insights motivate inference-time mechanisms that rebalance workloads without retraining or router modification. We propose Replicate-and-Quantize (R&Q), a training-free and near-lossless framework for dynamic workload rebalancing. In each layer, heavy-hitter experts are replicated to increase parallel capacity, while less critical experts and replicas are quantized to remain within the original memory budget. We also introduce a Load-Imbalance Score (LIS) to measure routing skew by comparing heavy-hitter load to an equal allocation baseline. Experiments across representative SMoE models and benchmarks show up to 1.4x reduction in imbalance with accuracy maintained within +/-0.6%, enabling more predictable and efficient inference.

Mixture-of-Experts (MoE) architectures are evolving towards finer granularity to improve parameter efficiency. However, existing MoE designs face an inherent trade-off between the granularity of expert specialization and hardware execution efficiency. We propose OmniMoE, a system-algorithm co-designed framework that pushes expert granularity to its logical extreme. OmniMoE introduces vector-level Atomic Experts, enabling scalable routing and execution within a single MoE layer, while retaining a shared dense MLP branch for general-purpose processing. Although this atomic design maximizes capacity, it poses severe challenges for routing complexity and memory access. To address these, OmniMoE adopts a system-algorithm co-design: (i) a Cartesian Product Router that decomposes the massive index space to reduce routing complexity from O(N) to O(sqrt(N)); and (ii) Expert-Centric Scheduling that inverts the execution order to turn scattered, memory-bound lookups into efficient dense matrix operations. Validated on seven benchmarks, OmniMoE (with 1.7B active parameters) achieves 50.9% zero-shot accuracy across seven benchmarks, outperforming coarse-grained (e.g., DeepSeekMoE) and fine-grained (e.g., PEER) baselines. Crucially, OmniMoE reduces inference latency from 73ms to 6.7ms (a 10.9-fold speedup) compared to PEER, demonstrating that massive-scale fine-grained MoE can be fast and accurate. Our code is open-sourced at https://github.com/flash-algo/omni-moe.

Temporal causal representation learning methods assume that causal mechanisms switch instantaneously between discrete domains, yet real-world systems often exhibit continuous mechanism transitions. For example, a vehicle's dynamics evolve gradually through a turning maneuver, and human gait shifts smoothly from walking to running. We formalize this setting by modeling transitional mechanisms as convex combinations of finitely many atomic mechanisms, governed by time-varying mixing coefficients. Our theoretical contributions establish that both the latent causal variables and the continuous mixing trajectory are jointly identifiable. We further propose TRACE, a Mixture-of-Experts framework where each expert learns one atomic mechanism during training, enabling recovery of mechanism trajectories at test time. This formulation generalizes to intermediate mechanism states never observed during training. Experiments on synthetic and real-world data demonstrate that TRACE recovers mixing trajectories with up to 0.99 correlation, substantially outperforming discrete-switching baselines.

While Mixture-of-Experts (MoE) scales capacity via conditional computation, Transformers lack a native primitive for knowledge lookup, forcing them to inefficiently simulate retrieval through computation. To address this, we introduce conditional memory as a complementary sparsity axis, instantiated via Engram, a module that modernizes classic $N$-gram embedding for O(1) lookup. By formulating the Sparsity Allocation problem, we uncover a U-shaped scaling law that optimizes the trade-off between neural computation (MoE) and static memory (Engram). Guided by this law, we scale Engram to 27B parameters, achieving superior performance over a strictly iso-parameter and iso-FLOPs MoE baseline. Most notably, while the memory module is expected to aid knowledge retrieval (e.g., MMLU +3.4; CMMLU +4.0), we observe even larger gains in general reasoning (e.g., BBH +5.0; ARC-Challenge +3.7) and code/math domains~(HumanEval +3.0; MATH +2.4). Mechanistic analyses reveal that Engram relieves the backbone's early layers from static reconstruction, effectively deepening the network for complex reasoning. Furthermore, by delegating local dependencies to lookups, it frees up attention capacity for global context, substantially boosting long-context retrieval (e.g., Multi-Query NIAH: 84.2 to 97.0). Finally, Engram establishes infrastructure-aware efficiency: its deterministic addressing enables runtime prefetching from host memory, incurring negligible overhead. We envision conditional memory as an indispensable modeling primitive for next-generation sparse models.

Mixture-of-Experts (MoE) has become a prominent paradigm for scaling Large Language Models (LLMs). Parameter-efficient fine-tuning methods, such as LoRA, are widely adopted to adapt pretrained MoE LLMs to downstream tasks. However, existing approaches typically assign identical LoRA ranks to all expert modules, ignoring the heterogeneous specialization of pretrained experts. This uniform allocation leads to a resource mismatch: task-relevant experts are under-provisioned, while less relevant ones receive redundant parameters. To address this, we propose DR-LoRA, a Dynamic Rank LoRA framework for fine-tuning pretrained MoE models. Specifically, DR-LoRA initializes all expert LoRA modules with a small active rank and uses an expert saliency score, which combines routing frequency and gradient-based rank importance, to identify which experts would benefit most from additional capacity. It then periodically expands the active ranks of the task-critical expert LoRA, progressively constructing a heterogeneous rank distribution tailored to the target task. Experiments on three MoE models across six tasks show that DR-LoRA consistently outperforms LoRA and other strong baselines, demonstrating that task-adaptive heterogeneous rank allocation is an effective strategy to improve active capacity utilization in MoE fine-tuning.

Generalist vision--language--action (VLA) policies are typically trained on heterogeneous mixtures of robot demonstrations spanning diverse embodiments, action spaces, and observation configurations. Modeling such heterogeneity with a shared dense action module can induce negative transfer, particularly when action spaces or visual observations differ across data sources. We address this issue with HiMoE-VLA, a VLA framework built around a Hierarchical Mixture-of-Experts (HiMoE) action module. HiMoE uses Action-Space MoE layers at the input/output boundaries to specialize computation for distinct action spaces, Heterogeneity-Balancing MoE layers in neighboring layers to provide balanced capacity for residual variation in observations, scenes, and embodiments, and dense Transformer blocks in the middle to integrate shared representations. Two auxiliary objectives further guide this hierarchy: a contrastive Action-Space Regularization objective for boundary specialization and a load-balancing objective for stable expert utilization. HiMoE-VLA reaches 3.98 on CALVIN, 98.0% on LIBERO, and 75.0% and 63.7% average success on real xArm7 and ALOHA tasks; under controlled heterogeneous co-training, it turns the negative transfer observed in strong baselines into positive transfer. The code and models are publicly available at https://github.com/ZhiyingDu/HiMoE-VLA.

Enabling lifelong learning in LLMs demands resolving the stability-plasticity dilemma (i.e., models must incorporate new knowledge without overwriting prior representations) while maintaining scalability under bounded parameter growth. Existing PEFT methods fail to satisfy this triad; shared-parameter approaches suffer from catastrophic interference, while task-isolated expansions preclude knowledge transfer and scale linearly. We propose Mixtures of SubExperts (MoSEs), a modular and sparse framework that factorizes model capacity into reusable, compositional primitives. MoSEs augment transformer layers with lightweight SubExperts and a learned sub-routing function that dynamically selects and composes a sparse subset of modules conditioned on task inputs. This induces a structured decomposition of the parameter space where knowledge is localized yet accessible, mitigating interference while preserving reuse. Specifically, MoSEs balance the dilemma via three pillars: (i) stability by isolating knowledge within sparsely activated modules, (ii) plasticity through routing-driven recombination and selective expansion, and (iii) scalability via sublinear growth in effective capacity. Notably, the routing mechanism enables compositional generalization, allowing new tasks to be represented as combinations of previously acquired sub-functions. We empirically validate MoSEs on TRACE and SuperNI, showing reduced forgetting, improved forward transfer, and better parameter efficiency over strong PEFT baselines. MoSEs establish a new Pareto frontier, achieving state-of-the-art performance while maintaining strict parameter budgets. Our results suggest that modular sparsity and compositional routing are key inductive biases for building foundation models that continually learn without saturation.

Mixture-of-Experts (MoE) models have shown strong potential in scaling language models efficiently by activating only a small subset of experts per input. However, their widespread deployment remains limited due to the high memory overhead associated with storing all expert parameters, particularly as the number of experts increases. To address this challenge, prior works have explored expert dropping and merging strategies, yet they often suffer from performance drop at high compression ratios. In this paper, we introduce PuzzleMoE, a training-free MoE compression method that achieves both high accuracy and efficient inference through two key innovations: First, PuzzleMoE performs sparse expert merging by identifying element-wise weight redundancy and specialization. It uses a dual-mask to capture both shared and expert-specific parameters. Second, to avoid the overhead of storing binary masks and signs, PuzzleMoE introduces a bit-packed encoding scheme that reuses underutilized exponent bits, enabling efficient MoE inference on GPUs. Extensive experiments demonstrate that PuzzleMoE can compress MoE models by up to 50% while maintaining accuracy across various tasks. Specifically, it outperforms prior MoE compression methods by up to 16.7% on MMLU at 50% compression ratio, and achieves up to 1.28\times inference speedup.

Dementia disorders such as Alzheimer's disease (AD) and frontotemporal dementia (FTD) exhibit overlapping electrophysiological signatures in EEG that challenge accurate diagnosis. Existing EEG-based methods are limited by full-band frequency analysis, which hinders the precise differentiation of dementia subtypes and severity stages. To address this limitation, we propose a Variational Mixture of Graph Neural Experts (VMoGE) framework that integrates multi-band EEG analysis with variational graph neural networks and a mixture-of-experts architecture. Each expert specializes in a specific EEG frequency band and models brain connectivity using a Gaussian Markov random field prior, while a variational gating mechanism adaptively integrates the expert outputs. This design enables the model to learn frequency-specific brain network representations while modeling latent uncertainty through variational inference. Experimental results on two EEG dementia datasets show that VMoGE achieves strong performance, with an AUC of 0.89 for HC vs. AD classification in the main comparison and competitive results across dementia subtyping and CDR staging tasks. Clinically, VMoGE offers three key translational values: the expert gating weights correlate with MMSE scores and CDR severity; slow-wave $δ$- and $θ$-band contributions are associated with AD-related EEG slowing and disease progression; and spatially localized activation maps reveal posterior $θ$- and $α$-band alterations and region-specific $β$-band changes, providing neurophysiologically interpretable markers aligned with known AD neuropathology.

Visual navigation policy is widely regarded as a promising direction, as it mimics humans by using egocentric visual observations for navigation. However, optical information of visual observations is difficult to be explicitly modeled like LiDAR point clouds or depth maps, which subsequently requires intelligent models and large-scale data. To this end, we propose to leverage the intelligence of the Vision-Language-Action (VLA) model to learn diverse navigation capabilities from synthetic expert data in a teacher-student manner. Specifically, we implement the VLA model, MM-Nav, as a multi-view VLA (with 360 observations) based on pretrained large language models and visual foundation models. For large-scale navigation data, we collect expert data from three reinforcement learning (RL) experts trained with privileged depth information in three challenging tailor-made environments for different navigation capabilities: reaching, squeezing, and avoiding. We iteratively train our VLA model using data collected online from RL experts, where the training ratio is dynamically balanced based on performance on individual capabilities. Through extensive experiments in synthetic environments, we demonstrate that our model achieves strong generalization capability. Moreover, we find that our student VLA model outperforms the RL teachers, demonstrating the synergistic effect of integrating multiple capabilities. Extensive real-world experiments further confirm the effectiveness of our method.

Encoder-Decoder architectures are widely used in deep learning-based Deformable Image Registration (DIR), where the encoder extracts multi-scale features and the decoder predicts deformation fields by recovering spatial locations. However, current methods lack specialized extraction of features (that are useful for registration) and predict deformation jointly and homogeneously in all three directions. In this paper, we propose a novel expert-guided DIR network with Mixture of Experts (MoE) mechanism applied in both encoder and decoder, named SHMoAReg. Specifically, we incorporate Mixture of Attention heads (MoA) into encoder layers, while Spatial Heterogeneous Mixture of Experts (SHMoE) into the decoder layers. The MoA enhances the specialization of feature extraction by dynamically selecting the optimal combination of attention heads for each image token. Meanwhile, the SHMoE predicts deformation fields heterogeneously in three directions for each voxel using experts with varying kernel sizes. Extensive experiments conducted on two publicly available datasets show consistent improvements over various methods, with a notable increase from 60.58% to 65.58% in Dice score for the abdominal CT dataset. Furthermore, SHMoAReg enhances model interpretability by differentiating experts' utilities across/within different resolution layers. To the best of our knowledge, we are the first to introduce MoE mechanism into DIR tasks.

Discrete optimization is ubiquitous in science and engineering. The vast array of existing discrete optimization problems, coupled with the continuous emergence of new ones, necessitates off-the-shelf optimizers capable of generating high-quality solutions for a large variety of optimization problems. This article introduces MEGO, a novel general-purpose neural optimizer for binary optimization under the black-box setting, intended for broad applicability across diverse binary optimization problem classes with minimal problem-specific customization. MEGO comprises a mixture-of-experts trained without domain knowledge. When presented with a new problem instance to solve, it employs a routing policy to dynamically activate the most relevant expert models to generate high-quality solutions. The strong generalization capability of MEGO is demonstrated on six problem classes from different disciplines, including classic problems and real-world applications. Trained solely on classic problems, MEGO effectively generalizes to unseen and complex real-world problem classes, significantly outperforming widely-used general-purpose optimizers in both solution quality and efficiency. Furthermore, MEGO provides a computational approach for quantifying similarity between optimization problems and classifying them, which is fundamentally different from the conventional analysis-based problem classification.

Mixtures of experts (MoE) are conditional mixture models in which both the mixing proportions and the component densities depend on the predictors, and are widely used for regression, classification and model-based clustering of heterogeneous data. Fitting MoE by maximum likelihood becomes unstable, and sometimes infeasible, when the predictors are numerous or correlated. We propose a regularized maximum likelihood framework for simultaneous parameter estimation and feature selection in MoE whose experts belong to the generalized linear model family, covering Gaussian, Poisson and multinomial responses within a single formulation. Sparsity is induced in both the gating network and the experts through $\ell_1$ penalties, and the penalized log-likelihood is maximized by a proximal Newton-EM algorithm whose M-step reduces to weighted Lasso problems with closed-form coordinate-ascent updates. Unlike existing penalized MoE procedures, the algorithm requires neither a local quadratic approximation of the penalty nor any matrix inversion, it returns exactly sparse estimates without thresholding, and a proximal Newton-type variant guarantees a monotone increase of the penalized objective at every iteration. On simulated data and five real data sets, the method recovers the actual sparsity support and delivers prediction and clustering accuracy that is competitive with, and often better than, state-of-the-art regularized MoE. The source codes of our developed algorithms and their documentation are publicly available on Github at https://github.com/nv-thin/GLM-RMoE.

Mixture-of-Experts (MoE) models have become mainstream for scaling language models to hundreds of billions of expert parameters. Despite sparse expert activation, existing inference engines keep all experts GPU-resident, crowding out the key-value cache in large-batch, long-output offline workloads. We present FluxMoE, which decouples experts from physical GPU residency and adapts their footprint to available memory through a new \emph{expert paging} abstraction. FluxMoE combines PagedTensor for transparent remapping, a bandwidth-balanced hierarchy spanning losslessly compressed GPU memory and host DRAM, and a budget-aware residency planner. Unlike CPU-GPU co-inference and whole-layer offloading, FluxMoE streams weights on demand while keeping expert computation on GPUs. We implement FluxMoE atop vLLM and evaluate it on three MoE models. For GLM-4.5 on 8$\times$H20 GPUs, FluxMoE delivers up to 7.2$\times$ vLLM's throughput and 79.0% lower average Time-Per-Output-Token (TPOT), without measurable model-quality loss using lossless compression. For Mixtral-8$\times$7B-Instruct on 2$\times$L40S GPUs, where weight-resident vLLM cannot fit, FluxMoE delivers 4.3$\times$ KTransformers's throughput and 29.1% lower average TPOT.

Real-world time series exhibit complex and evolving dynamics, making accurate forecasting extremely challenging. Recent multi-modal forecasting methods leverage textual information such as news reports to improve prediction, but most rely on token-level fusion that mixes temporal patches with language tokens in a shared embedding space. However, such fusion can be ill-suited when high-quality time-text pairs are scarce and when time series exhibit substantial variation in characteristics, thus complicating cross-modal alignment. In parallel, mixture-of-experts (MoE) architectures have proven effective for both time series modeling and multi-modal learning, yet many existing MoE-based modality integration methods still depend on token-level fusion. To address this, we propose Expert Modulation, a new mechanism for multi-modal time series prediction that conditions both routing and expert computation on textual signals, enabling direct and efficient cross-modal control over expert behavior. Through theoretical analysis and experiments, our proposed method demonstrates strong improvements in multi-modal time series prediction. The current code implementation is available at https://github.com/BruceZhangReve/MoME

Looped Transformers increase effective depth by iterating a shared block of layers, but most evaluations compare at fixed model size, conflating architectural advantage with extra FLOPs. We study looping on Mixture-of-Experts Transformers while closely matching per-token FLOPs, total non-embedding parameters, and KV cache. Through a series of ablations, we arrive at a recipe we call SMELT (Sparse MoE Transformer, middle layers Loop Twice), which loops the middle half of layers twice while matching the unlooped Baseline on all three budgets. We scale SMELT across four sizes up to 54B non-embedding parameters and fit a separate Chinchilla-style scaling law for each architecture. SMELT's loss drops faster with compute, saving 6.8--18.0% of training FLOPs on the compute-optimal frontier. The advantage transfers to downstream benchmarks beyond what validation loss predicts, is largest on Code, and grows with sample length and the number of in-context examples. Mechanistic analysis shows that the second visit reduces the attention sink and redirects mass toward content-relevant tokens, an inductive bias that may underlie the observed performance gains. These results show that looping can improve Transformers even under budget matching, offering a practical recipe that turns depth reuse into measurable gains.

Mixture-of-Experts (MoE) models activate few experts per token, yet their full expert sets can exceed GPU memory and require repeated weight transfers during decoding. We formulate expert-cache management as a model-side algorithmic problem and propose cache-aware post-training that jointly adapts the MoE backbone and lightweight auxiliary routers while preserving the native inference-time Top-K rule. The update-only Temporal Router learns same-layer retention across tokens without proactive loading. The full Spatio-Temporal Router adds a Spatio Router that uses the causal predecessor's hidden state to refine the temporal cache before target-layer access. We evaluate both modes on Qwen3 and GPT-OSS across GSM8K, MATH, and CommonsenseQA. Temporal Router consistently improves hit rate and reduces expert-weight traffic over matched LM-only baselines. On Qwen3, the full mode improves adjusted hit rate by 1.15--18.03 points and reduces traffic by 4.6--53.3% relative to the strongest evaluated prefetching baseline; GPT-OSS results are competitive but task-dependent. Auxiliary-only training preserves baseline accuracy but yields modest coverage gains; joint post-training achieves substantially higher coverage. Sensitivity analyses distinguish the effects of cache capacity, refinement budget, and cache-loss weight on coverage, traffic, and quality.