Mixture-of-Experts Models

Semi-supervised adaptation of vision foundation models (VFMs) commonly freezes the pretrained backbone and updates lightweight modules such as LoRA. However, pseudo-labels have mixed reliability, and a single LoRA adapter must absorb reliable, ambiguous, and noisy gradients in the same low-rank space. This can make VFM adaptation sensitive to pseudo-label noise. We propose \textbf{TriNoL}, a \textbf{Tri}ple-expert learning framework from \textbf{No}isy \textbf{L}abels for semi-supervised VFM adaptation. TriNoL routes unlabeled samples into three confidence regions and assigns them to three LoRA experts: a Positive Expert for high-confidence pseudo-labels, an Alignment Expert for medium-confidence ambiguous samples, and a Negative Expert for low-confidence noisy samples. The VFM backbone remains frozen, and only the LoRA experts and classifier head are updated. By separating different pseudo-label reliability regions into specialized adaptation paths, TriNoL improves robustness to noisy supervision while keeping the training cost low.

Underwater images often suffer from diverse and coexisting degradations, including color distortion, scattering haze, texture attenuation, and uneven illumination. These degradations vary across regions and may coexist locally, making conventional uniform restoration difficult to adapt to different degradation patterns. To address this problem, we propose Coexisting and Region-wise Degradation for Underwater Image Enhancement (\textbf{CoRe-UIE}), a degradation-oriented expert collaboration framework. CoRe-UIE combines a content-preserving shared expert with four shared-backbone routed experts for color correction, scattering suppression, texture recovery, and illumination protection. The routed experts share the same architecture but have independent parameters, and are assigned to different regions through input-derived degradation cues and region-adaptive Top-$k$ routing. We further introduce a Hilbert--Schmidt Independence Criterion (HSIC)-based representation constraint to reduce statistical dependence among expert features and alleviate redundant expert responses. Experiments on UIEB, LSUI, and U45 demonstrate that CoRe-UIE achieves competitive quantitative performance and visually balanced enhancement under diverse underwater degradation conditions.

PTQTP decomposes LLM weight matrices into two ternary (trit) planes with two free per-group scales. Tying the scales to a fixed ratio of three collapses the decomposition into a single uniform nine-level quantizer, a known balanced-ternary identity. To our knowledge, at the time of writing, this work is the first to impose that identity as a constraint inside PTQTP's solver. The two trit planes then fold losslessly into one 4-bit code plane that we make the persistent serving representation: disk bytes, expert-cache bytes, and kernel input are the same 4.0625-bits/weight blocks, consumed in one integer dot pass. For this conjunction (ratio-3 nine-level code, CPU-SIMD kernels, SSD expert streaming, identical persistent bytes) we likewise found no precedent. We apply this to the routed experts of DeepSeek-V4-Flash-0731, a 284B-A13B mixture-of-experts model, quantizing in one shot from the released MXFP4 expert weights and streaming experts from SSD on a 64 GB laptop. Against a 4.5-bit Q4_K baseline, measured one process per fixture with an expert-lossless anchor arm as reference control, the tied model matches the official serving API on 5/5 fixtures at step 0 (Q4_K: 4/5) and 12/14 captured continuation steps (11/14), scores 86 vs. 84 on a 100-item MMLU subset, decodes 6.7% faster in decode phase, and ships 9% smaller files: no detected fidelity difference at these small evaluation sizes, and every fixture-level difference between the arms traces to a single measured near-tie cell. The tied fit nevertheless shows higher weight-reconstruction error and worse perplexity, a measured dissociation between proxy metrics and reference fidelity. A cumulative trunk-ternarization ladder and bitwise-pinned aarch64/x86-64 kernels complete the report. All code, formats, and evaluation artifacts are open source in the fucina inference stack.

Sparse Mixture-of-Experts (MoE) routers commonly use the same scores both to select experts and to weight their already-computed outputs. We study whether these two roles, dispatch and aggregation, should be coupled. On pretrained OLMoE-1B-7B, we keep selected Top-8 expert IDs, expert computation, and total selected router mass fixed and change only within-set aggregation. A structured oracle improves full-horizon cross-entropy by 0.0160 +/- 0.0039 across three seeds; the router's top-scored expert is the counterfactual-best vertex only 17.2% of the time, with router-utility Spearman 0.030. We therefore train Fixed-Dispatch Adaptive Aggregation (FDAA), a 301K-parameter post-compute head optimized directly with the language-modeling objective while freezing the backbone, router, and experts. On OLMoE, FDAA improves fresh WikiText-103 test by Delta CE = -0.1523 +/- 0.0031 across three seeds, and mixed-domain training gives robust gains on WikiText-103, C4, and held-out Penn Treebank under frozen confirmatory evaluation. We also replicate the fixed-dispatch audit on DeepSeek-V2-Lite, which uses Top-6 routed experts plus shared experts. Best-vertex headroom remains significant on WikiText and C4, while router Top1 identifies the best selected expert in only 12.5% and 16.7% of audited examples. In a one-seed mixed-domain replication, FDAA improves locked WikiText and PTB, while C4 is statistically neutral. These results support a cross-architecture distinction between expert selection and expert commitment.

Timely crop-disease identification is critical to food security. Multi-crop recognition suits Mixture-of-Experts (MoE), but conventional soft-routing MoE learns crop assignment freely end-to-end, letting a few experts dominate (expert collapse) with no semantic correspondence to crops, and facing high retraining costs, unstable rejection of non-target inputs, and a saturated accuracy ceiling. We shift the objective from accuracy toward a trade-off among deployment cost, scaling flexibility, and rejection stability, using deterministic hard routing. We propose AdapterMoE: a RouterHead classifies the crop and rejects non-target crops via a Maximum Softmax Probability threshold, with a dual-gate Energy+KNN out-of-distribution module catching distribution-shifted inputs; five per-crop Adapters atop a frozen EfficientNet-B0 backbone discriminate diseases, each calibrated via Temperature Scaling. Because experts are hard-isolated at the data level, the design avoids expert collapse and exposes an add_crop interface for local, per-crop updates instead of full retraining. On PlantVillage (5 crops, 26 classes), across a fair five-system comparison, AdapterMoE attains accuracy statistically indistinguishable from the best baselines (Macro-F1 within a 0.24-point band) while cutting training cost to about 9% of full-network baselines, expanding to a new crop in

Mixture-of-Experts models increase parameter capacity while keeping the computation activated by each token bounded, but their architectural evolution cannot be explained by a chronological list of model releases alone. This technical survey synthesizes primary papers, official technical reports, and prior surveys to organize modern Mixture-of-Experts systems along five coupled dimensions: expert granularity, expert topology, routing freedom, the scope of load balancing, and execution structure. We describe eight architectural milestones as a dependency graph with six mainline developments and two orthogonal branches, rather than as eight successive generations. We then analyze individual systems through four control planes: Expert Topology, Routing, Balance, and Expert Parallelism. These planes specify which experts exist, which experts process each token, how aggregate load is controlled, and how selected computation is mapped onto physical devices. The framework connects algorithmic choices such as Top-k routing, shared experts, fine-grained experts, and dynamic expert composition with systems concerns including token dispatch, device placement, all-to-all communication, and communication-computation overlap. We conclude with equal-budget pretraining experiments, quality and systems metrics, and open research questions. The main trend is a shift from merely activating more sparse parameters toward decoupling semantic routing, computational budgets, and physical execution.

Sparse mixture-of-experts (MoE) layers expand recommendation capacity through conditional computation, yet a trained checkpoint still stores and routes over its full expert bank. We study a deployment problem: convert that checkpoint to a smaller standard MoE under an explicit expert budget, without adding a compression-specific online module. To address this, we introduce UniMoMo, a post-training compression framework formulated as a constrained graph coarsening problem. Rather than relying on parameter distance, UniMoMo groups experts based on their functional similarity, using an unlabeled calibration set to measure how similarly experts respond to shared recommendation states. To prevent performance degradation, we introduce a layer-adaptive protection mechanism that restricts the merging of high-traffic experts based on their routing exposure. Across Amazon Beauty, KuaiRec, and TenRec with 2, 4, and 6 MoE blocks, the final four-expert checkpoints obtain source-relative five-run mean NDCG@10 ratios of 99.92%--102.30% and measured A100 speedups of 1.28$\times$--1.63$\times$. An aggressive two-expert, top-1 operating point obtains ratios of 98.36%--104.24% and speedups of 1.47$\times$--2.21$\times$. These endpoint results evaluate the complete conversion-and-adaptation workflow and show that a trained recommendation MoE can be exported at multiple serving budgets.

Multivariate Time Series Classification (MTSC) demands models that can effectively capture complex temporal patterns across multiple scales while remaining computationally efficient. However, existing approaches generally struggle to reconcile fine-grained representation learning, especially under class imbalance and real-world constraints. In this paper, we present FreSH, a Frequency-Segmented Hierarchical Multi-Expert Framework designed to address these challenges. FreSH introduces a new perspective for MTSC by enabling adaptive, multi-scale analysis of temporal signals, allowing different aspects of the data to be modeled in a complementary and coordinated manner. By combining localized specialization with holistic context modeling, FreSH achieves strong representational capacity without incurring excessive computational overhead. An adaptive fusion strategy further enhances flexibility, enabling the model to dynamically emphasize the most informative components of the input. In addition, we incorporate a more robust optimization objective that improves learning stability across varying sample difficulties and class distributions. Extensive evaluations on 30 UEA benchmark datasets and real-world vibration data demonstrate that FreSH consistently outperforms state-of-the-art methods in classification accuracy, while substantially reducing model size and efficiency.

Large mixture-of-experts (MoE) language models with 26--120 billion parameters exceed the memory capacity of consumer devices through three simultaneous pressures: resident weight matrices, key-value (KV) cache state that grows linearly with context, and dozens of expert sublayers that must be paged on demand. We present RotaryQuant, a three-axis compression system that addresses all three. Mixed-precision weight quantization assigns bit-widths by architectural role: 4-bit for dense layers, 2-bit for routed experts, and 8-bit for the shared expert whose high activation kurtosis resists aggressive compression. LRU expert offloading pages non-resident experts to disk under genuine memory pressure. The novel axis is IsoQuant, a KV cache compression method that applies a Walsh--Hadamard transform followed by block-diagonal SO(4) rotations to isotropize activation distributions before 3-bit scalar quantization, requiring $O(d \log d)$ operations and 256 stored parameters per head versus $O(d^2)$ and 16{,}384 for dense rotation methods. A fused four-kernel Metal GPU pipeline performs attention directly on packed 3-bit tensors without materializing full-precision KV state---a different execution model, not just a quantization scheme. The combined system fits Gemma 4-26B-A4B and Qwen3-30B-A3B within a 16,GB budget and Nemotron-H 120B within 32,GB, running interactively at 9--19 tok/s with near-zero perplexity degradation ($Δ$PPL $\leq +0.0012$) and 100% retrieval accuracy at 32K context.

Safety alignment in multilingual models is uneven: a model that reliably refuses a harmful request in English will often comply with the same request in a lower-resource language. We trace this gap mechanistically in sarvam, an Indic-multilingual mixture-of-experts reasoning model, and find it is not a failure to detect harm. Harm is encoded as an internal direction that is nearly language-invariant in mid-network (English-vs-Indic cosine ${\approx}0.9$ at $L11$), and steering that direction upstream causally controls refusal. But the detection direction is orthogonal to the change that actually writes the refusal, which is late and assembled over the course of generation rather than read off in a single forward pass. We attribute the write to a specific, localizable circuit, a mixture-of-experts writer held in check by an attention opposer and price every way of intervening on it: damping the opposer is cheap and effective, amplifying the writer is a cost wall, and surgical edits to the responsible heads do nothing. The circuit's organization, and the gradient method that exposes it, recur in a second, unrelated MoE model, while the lever's strength is architecture-specific. The result is a cost-measured map of where a multilingual safety repair can land, and what it costs

Load Balancing has emerged as a critical problem in expert-parallel distributed inference of Mixture-of-Experts (MoE) models. As routing distributions are typically skewed across experts, devices hosting lighter-loaded experts must idle to wait for the heaviest during expert computing, leading to inefficiency. Existing load-balancing approaches primarily rely on expert replication or migration within each layer, which introduce additional overhead and limit their flexibility and scalability. To address this problem, we propose EasyBalance, a cross-layer load balancing strategy that requires no modifications to the expert-device mapping, enabling instant adaptability and incurring essentially no additional overhead. Our key insights are that (1) experts of other layers can be viewed as naturally redundant for the current layer, and (2) cross-layer MoE workloads can be jointly executed to mitigate their individual imbalance. Based on these observations, EasyBalance greedily schedules a subset of cross-layer workloads to run at each MoE step and defers the remaining workloads for future balancing opportunities, effectively leveraging cross-layer imbalance mitigation. Extensive experiments across models, tasks, and configurations demonstrate that EasyBalance consistently accelerates distributed MoE inference, reducing GPU idling by mostly over 40%. Code is available at https://github.com/yize-wu/EasyInfra.

Most language-model access controls regulate behavior while leaving the same computation available to every request. We study a different systems question: can trusted authorization determine which newly trained parameters are reachable by the forward pass? Policy-Masked Private Experts freezes a pretrained sparse Mixture-of-Experts (MoE) model, trains a disjoint expert branch, and selects the public or private pool before top-k routing. The resulting claim is narrow but testable: under the declared trusted computing base (TCB), an unauthorized request executes no private expert. It does not imply that the public model lacks the same semantic capability. We test this separation between execution control and task utility in Qwen3-30B-A3B and DeepSeek-V2-Lite. Three Qwen BF16 seeds update all 32 private experts while the public fingerprint remains unchanged. Across 64 adversarial scenarios and 96 deny/fail-closed events, unauthorized private execution is zero; independent hooks exactly match 11,616 routed private rows and allow-deny-allow recovery is exact. On two prospectively frozen Qwen benchmarks, the private branch improves exact tool use by 5.0 percentage points (pp) (five versus zero discordances; one-sided Holm p = 0.03125, corresponding two-sided exact p = 0.0625) and 21.3 pp (percentile-bootstrap 95% CI [13.3, 29.3], Holm p = 0.000031). Three arm-blinded model evaluators retain a positive external effect of 18.7 pp (95% CI [9.3, 28.0]). A parameter-matched Lora has similar external utility, but a post-hoc request gate leaves 1,225 adapter calls under deny; the disjoint expert branch leaves none. DeepSeek reproduces the route invariant and gains 27.0 pp. A valid sealed evaluation is near-neutral. These results support auditable, reversible control over a trained parameter path, while showing that useful transfer remains distribution dependent.

Vision-Language-Action (VLA) models have become a powerful paradigm for robot manipulation, but training a single generalist policy for heterogeneous robot embodiments remains an open problem. Existing methods have two main limitations. First, they underuse dynamics priors shared across diverse visual and interaction data, limiting cross-embodiment transfer. Second, they require extensive manual preprocessing to convert embodiment-specific actions into a common format. To overcome these limitations, we propose DyPES-VLA, a cross-embodiment VLA that learns shared Dynamics Priors and Embodiment-Specific control. First, we learn shared dynamics priors by training the vision-language model (VLM) with a future-prediction objective on cross-embodiment data, driving the shared query representation to capture object motion, contact, and interaction-induced scene changes. Second, an embodiment-specific Mixture-of-Experts (MoE) action head translates these shared dynamics priors into executable controls directly in each embodiment's native action space, without manually pre-aligning heterogeneous actions into a common format. This head shares attention layers to capture common temporal action structures, while its embodiment-specific feed-forward experts resolve the unique kinematic constraints and control semantics of distinct embodiments. As a generalist policy, our \ourmethod achieves state-of-the-art performance across simulation and real-world evaluations, reaching 98.0% success on LIBERO, 59.25% on RoboCasa-GR1, and 89.02% on RoboTwin~2.0.

On-device deployment of Large Language Models (LLMs) has become essential for personalized edge applications. A primary bottleneck is external memory access (EMA) in feed-forward network (FFN) layers. Speculative decoding and mixture-of-experts (MoE) are promising solutions. Speculative decoding reduces the number of decoding stages by generating multiple tokens per stage, and MoE minimizes per-stage cost through sparse expert activation. However, there is an incompatibility when combining these two techniques. We propose EdgeXpert, a software-hardware co-designed LLM accelerator that resolves this incompatibility. In the prefill stage, the prompt-wise expert reuse reformulates routing as prompt-level expert reuse rather than independent per-token expert selection. It identifies important tokens using a lightweight encoder, constructs a shared expert set from them, and routes less important tokens with a reduced expert budget to lower expert EMA. In the decode stage, depth-aware expert coalescing exploits the contextual similarity and mutual exclusivity of same-depth candidate tokens. Rather than loading the union of all required channels, EdgeXpert loads only salient channels and applies computational calibration to recover accuracy without additional memory access. Synthesized in Samsung 28nm technology at 800 MHz, EdgeXpert achieves up to 56.3% latency reduction and 44.1% energy reduction compared to prior works, while maintaining near-baseline accuracy.

Mixture-of-experts vision-language models (MoE-VLMs) increase model capacity with sparse expert activation, yet deployment requires storing the full expert pool. Training-free expert merging reduces this burden, and many routing-based methods aggregate routing statistics across all tokens to determine merge compatibility. However, MoE-VLM inference is phase-structured: image-context tokens carry visual content, question tokens specify the query, and answer tokens produce the output, with different counts and routing distributions. Because image-context tokens are far more numerous, global aggregation can overemphasize image-context processing and obscure phase-conditioned expert roles, making experts serving different phases appear interchangeable and degrading model performance. We therefore argue that MoE-VLM expert merging should preserve phase-conditioned expert roles, judging compatibility by how experts serve different phases rather than globally aggregated routing statistics. Based on this view, we propose RoleMerge, a training-free method that constructs each expert's Routing Role Profile (RRP) from phase-normalized routing statistics, capturing its relative phase preference. Guided by expert-phase information loss, RoleMerge merges experts with compatible profiles and their corresponding router entries while preserving answer-decoding expert distinctions. Experiments on three models and multiple benchmarks show that RoleMerge preserves more of the full model's performance than alternative expert-merging methods at matched expert-retention ratios, with relative improvements of up to 9.6 percent in six-task macro-average performance. These results validate phase-conditioned expert roles as a more effective basis than global routing aggregation for MoE-VLM expert merging.

Memory-efficient matrix optimizers such as Sinkhorn gradient descent remove most AdamW optimizer state for dense Transformer matrices, but direct application to Mixture-of-Experts (MoE) training is unreliable. We study this failure in a controlled 110M-parameter nanowhale DeepSeek-style MoE pretraining setting. A SAGE/Sinkhorn hybrid reduces optimizer state from 0.883GB to 0.331GB but degrades evaluation loss to 3.8265, far above the AdamW baselines observed in the same setup (3.58--3.64 across the seeds we study). We show that routed MoE expert matrices are the dominant failure point: their gradients are conditional, temporally varying, and poorly served by stateless Sinkhorn normalization. We propose MESH, a hidden-momentum Sinkhorn update for MoE experts. MESH restores a temporal first-moment signal through the gradient-buffer lifecycle, without storing the expert first moment as optimizer state. MESH is an optional block-preconditioned variant that adds a coarse neuron/block inverse-RMS multiplier. Across ablations, temporal smoothing before matrix normalization is the primary causal ingredient; block/neuron preconditioning can improve the memory-quality frontier, but is not established as universally necessary. In two additional seeds, MESH and MESH-B reduce optimizer-state memory by 62.5% and peak PyTorch CUDA allocation by about 12.6% relative to AdamW, with a modest evaluation-loss gap. Full-state diagnostic variants recover AdamW-like performance in ablations, supporting the conclusion that MoE experts need temporal smoothing, but not necessarily full coordinate-wise AdamW state.

Mixture-of-Experts (MoE) models enable model scaling while maintaining low inference-time compute by activating only a subset of experts per token. However, conventional routing relies on a fixed top-k selection, forcing the model to spend the same compute regardless of how many experts are relevant. We introduce elbow-based routing, a training-free inference-time modification that dynamically adjusts the number of experts on a per-token basis. Our method examines the sorted router probability distribution and identifies an elbow point that separates high- and low-probability experts. We find that most router distributions exhibit clear inflection points suitable for this strategy, and we show both theoretically and empirically that elbow-based routing preserves expert load balance. Experiments on a state-of-the-art MoE model demonstrate an average latency reduction of 5.3% while maintaining accuracy across six benchmarks.

Mixture-of-experts (MoE) networks pursue specialization through learned routers, gates, and load-balancing losses, yet at matched total-parameter budgets learned routers can underperform equal-weight No-Routing baselines. Is the bottleneck the routing algorithm, or the alignment between training-signal granularity and the target categories? We probe the question with SpecDrop, a fixed parameter-free routing scheme: each of $K$ branches receives weight $p_a$ for its assigned category and a small leakage $p_i > 0$ otherwise, merged through a category-independent fixed denominator, with no learned routing parameters and no auxiliary losses; the category label is required at inference. On vision tasks where each image has one superclass label (CIFAR-100 on ResNet-110; ImageNet-1K on ViT-S/16), SpecDrop reaches 79.23% on CIFAR-100 and 79.89% on ImageNet-1K, exceeding parameter-matched baselines that do not use the label (+4.75 over dense on CIFAR-100; +6.53 over the No-Routing+SE control on ImageNet-1K). These gains quantify what category supervision buys when deployed through routing -- not an advantage over label-aware deployments of the baselines: given the same label, masking a dense model's outputs is stronger for accuracy alone (85.2 / 83.7). SpecDrop's contribution is converting the label into trained-in modular structure: 58%/100% branch-category alignment, and masking gains of 0.00 (CIFAR) / +1.06 (ImageNet) -- the output-space restriction is largely internalized during training. On fuzzy partitions, where training units span multiple categories (SlimPajama-6B language modeling with a 30M Transformer; SuperNI instruction tuning over Llama-3.2-1B with LoRA), the routing mechanism reduces to the matched No-Routing controls within seed noise, the null our thesis predicts. Granularity alignment, not algorithm choice, localizes when routing helps. Code: https://github.com/Beryex/SpecDrop

Diffusion language models (dLLMs) offer an alternative to autoregressive (AR) language modeling, yet the scaling behavior of Mixture-of-Experts (MoE) dLLMs remains poorly understood. We systematically characterize how optimization hyperparameters, compute allocation, and architecture scale for MoE dLLMs, identifying quantitative differences from scaling trends previously reported for AR models. Specifically, for optimization, the optimal nominal batch size grows faster, while the optimal learning rate decays more rapidly with compute. For model--data allocation, IsoFLOP analysis reveals a slight data-side tilt: the optimal token budget grows faster than activated model-side computation. For MoE architecture, larger scales increasingly favor larger expert pools at fixed activated capacity, while moderate expert granularity remains consistently effective and the preferred fraction of activated capacity assigned to shared experts remains stable across scales. Guided by these findings, we train LLaDA MoE v2, a 30B-A3B dLLM, from scratch on 23.5T tokens. With approximately 65% as many pretraining tokens as Qwen3, LLaDA MoE v2 approaches Qwen3 on several knowledge, reasoning, and coding benchmarks. After supervised fine-tuning alone, it outperforms SDAR Chat on seven of eight reasoning and coding benchmarks and remains close to Qwen3 on several tasks. These results establish practical scaling laws and design principles for MoE dLLMs.

Parameter-efficient fine-tuning (PEFT) enables efficient adaptation of large language models, but existing MoE-based PEFT methods typically improve capacity by storing multiple full LoRA experts, causing adapter storage to grow linearly with the number of experts and restricting adaptation to a fixed expert pool. We ask whether MoE-based PEFT can produce instance-specific adaptations without explicitly storing a separate LoRA module for each expert. To address this gap, we propose MoEGen, an adaptation framework that shifts MoE-based PEFT from expert selection to expert-conditioned parameter generation. Instead of storing each expert as a full LoRA adapter, MoEGen represents each expert as a small learnable vector, termed an expert code. It routes each input over these vectors and uses their weighted combination to condition a lightweight hypernetwork that generates input-specific low-rank updates. This design decouples expert capacity from adapter storage while enabling instance-conditioned adaptation. Experiments on eight commonsense reasoning benchmarks show consistent improvements over strong static and MoE-based PEFT baselines across three backbones. MoEGen also performs strongly in joint medical and legal-domain adaptation.

Generating natural-language coaching feedback on motor skills can accelerate learning, yet expert coaches are scarce and expensive. Existing reference-based methods require expert demonstrations at both training and inference time, limiting practical deployment. We propose AIDE (Automated Instruction via Distilled Expertise), a framework that exploits expert references only during training and generates feedback from a learner's pose sequence alone at inference. A teacher model first learns to generate feedback from paired learner-expert poses via a frozen language model, producing separate learner tokens and difference tokens that encode the learner-expert difference. A student model then inherits the teacher's encoder and weight initialization, replacing the explicit expert comparison with an auxiliary module that produces complementary tokens from the learner's pose alone. On the ExpertAF dataset, AIDE outperforms reference-free baselines on most metrics and performs comparably to methods requiring expert demonstrations at both training and inference, with LLM-based evaluation supporting these findings.

Neural receivers have been recognized as a promising paradigm for the next-generation (NextG) communications. However, due to the reliance on a static network optimized for specific channel conditions, their generalization capability across diverse scenarios remains a significant challenge. To address this issue, this paper proposes a novel channel-adaptive neural receiver network (CARNet) based on the mixture-of-experts (MoE) framework. The proposed architecture employs multiple expert networks together with an efficient routing mechanism to enable signal detection in various scenarios. The experts are constructed via stacked ResNet blocks and specialize in robust signal detection within specific channel conditions, while the routing mechanism incorporates a lightweight representation learning module, which projects the coarse channel estimate into a low-dimensional latent embedding. The learned embedding characterizes task-relevant channel conditions and provides efficient guidance for accurate expert selection. Link-level simulation experiments demonstrate that the proposed CARNet achieves superior performance across diverse channel conditions.

Brain-computer interfaces (BCIs) have been widely used in motor rehabilitation, disease diagnosis, and other neural engineering scenarios. However, conventional neural signal decoding algorithms often suffer from limited generalizability and high adaptation costs, motivating recent interest in BCI foundation models. Existing approaches still struggle to jointly achieve general transferability, accurate decoding, and efficient downstream adaptation. We present STEAM, a hierarchical transfer framework that reconciles general-purpose representation learning with paradigm-specific specialization in EEG foundation models. The framework is instantiated as a dual-branch spatio-temporal encoder in which a shared soft mixture-of-experts (SSMoE) module aligns the spatial and temporal branches, allowing complementary representations to exchange information through a compact set of soft slots. Across seven downstream datasets and fourteen evaluation settings, STEAM attains the best average rank among the compared methods at a competitive inference cost measured in FLOPs. Building upon the Stage-I general initialization, the hierarchical pre-training strategy further specializes the model to a target paradigm without retraining from scratch, yielding consistent gains in paradigm-specific decoding accuracy.

Mixture-of-experts (MoE) models increase parameter capacity by activating only a small subset of experts for each token. This conditional-computation paradigm has enabled autoregressive language models to scale model capacity without a proportional increase in per-token computation. In diffusion language models (DLMs), however, each denoising forward jointly revisits all token positions despite their sharply different refinement demands, while the default fixed token-choice routing assigns them a uniform expert budget, creating a mismatch between expert computation and refinement demand. We argue that MoE inference in DLMs should therefore be viewed as refinement-aware compute allocation across heterogeneous token refinement states. We propose REFLEX (\textbf{RE}finement-aware \textbf{FLEX}ible expert allocation), a training-free method that keeps the default router unchanged while reorganizing expert computation around the evolving refinement process. Specifically, REFLEX introduces a coarse-to-fine hierarchy for expert-budget allocation that aligns computation with block-relative refinement roles while using the Frontier-Progress Score to resolve active-block priorities. Across multiple widely used benchmarks on two representative MoE-based DLMs, LLaDA-MoE and LLaDA2.0-mini, REFLEX reduces allocated expert computation by 15% on average while preserving or even improving generation quality on most benchmarks relative to default routing. Compared with autoregressive-style variable-expert routing methods, REFLEX also yields a more consistent quality--computation trade-off, further supporting the importance of allocating expert computation according to the heterogeneous refinement demands exposed within each denoising forward.

Retrieval-augmented generation (RAG) improves factuality but adds latency and engineering overhead at serving time. We propose RING (Retrieval-Internalized Generation), a holistic paradigm spanning both architecture and training that injects large-scale external knowledge into a \textit{Mixture-of-Memory Experts} and learns parametric search over this internal memory via reinforcement learning, removing the external retriever entirely. Training proceeds in three stages: continued pre-training injects new corpora into a Knowledge Expert via our novel \textit{Dual Causal Attention}; supervised fine-tuning teaches a ``search-then-answer'' pattern; and reinforcement learning with hierarchical rewards optimizes the routing-and-search policy over the parametric memory. Unlike prior parametric injection methods that pair internal memory with a fixed or rule-based retriever, RING {learns} its retrieval policy directly from task signals. We further frame RING theoretically as a search-free approximation to the classical RAG objective. To evaluate large-scale injection of genuinely {new} knowledge without test-time leakage, we further construct News-2025, a benchmark built from news strictly post-dating the base LLM's pretraining cutoff. RING matches or surpasses both search-based RAG and parametric injection baselines in accuracy and efficiency.

The normalized Transformer (nGPT) realizes hyperspherical representation learning by constraining model parameter vectors and activation vectors to the unit hypersphere. In this paper, we describe a practical training recipe for nGPT and evaluate it on modern hybrid Mamba-2--Transformer Mixture-of-Experts (MoE) models. The recipe introduces Logit Gradient Preconditioning, Logarithmic Learning Rate Decay, GatedAdamW, angular update control, and optional exploration mechanisms. Compared with an unnormalized model of the same hybrid MoE architecture trained with AdamW, the 30B-total-parameter nGPT model reaches the same validation loss using approximately half as many training tokens. The recipe scales across the models considered, which contain up to 30B total parameters.

Modern reranking recipes---billion-scale cross-encoders, mixture-of-experts (MoE) backbones, and distillation against strong teachers---have outpaced the training infrastructure available to most academic groups. Existing Tevatron reranker training relies on the Hugging Face Trainer with DeepSpeed or PyTorch FSDP1, but these backends lack efficient support for large-scale MoE training. We present Tevatron 3.0, which integrates a Megatron-Core training backend into Tevatron while preserving its data pipeline, evaluation workflow, and Hugging Face-compatible checkpoints. We benchmark existing distributed training configurations against the new backend, showing that Megatron matches FSDP reranker quality and training efficiency under comparable data-parallel settings, is up to 22% faster in the recommended single-node configuration, and supports both LoRA and full-parameter fine-tuning. Crucially, expert parallelism enables training a 30B-parameter Qwen3-30B-A3B MoE reranker, which is infeasible with PyTorch FSDP1. Using this framework, we conduct a controlled comparison of MoE versus dense models, LoRA versus full-parameter tuning, and distillation versus contrastive training on BEIR-15 with three first-stage retrievers, and report serving throughput for Hugging Face and vLLM. We find that the MoE reranker matches dense 8B quality while activating less than half as many parameters and achieving substantially higher inference throughput. We will release the framework and trained checkpoints.

Mixture-of-Experts (MoE) models have become a dominant architecture for large-scale AI services, yet deploying them over geo-distributed heterogeneous edge servers remains challenging. When the Top-k activated experts of a token are spread across multiple servers, the optimal routing depends jointly on cross-server link bandwidth, heterogeneous GPU computing capability, GPU-CPU expert loading delay, instantaneous queueing backlog, and replica-level quantization quality loss. Existing distributed inference and MoE serving methods address these factors separately and do not provide a unified framework for online multi-server collaborative routing. In this paper, we propose HetRoute, a heterogeneous-cost-aware collaborative routing framework for distributed edge MoE inference. HetRoute introduces a unified per-assignment cost model that explicitly captures four cost components: cross-server transmission, GPU-CPU offloading, GPU computation with queueing, and quantization-induced quality penalty. Guided by this model, the offline stage determines expert server placement, GPU-CPU residency, and replica precision through a routing-cost-coupled deployment algorithm, while the online stage routes the Top-k activated expert set as a whole by minimizing the bottleneck layer cost via exact enumeration or beam search. Theoretical analysis establishes fallback feasibility, a bound on the number of participating servers, per-layer optimality for small candidate domains, and online computational complexity. Trace-driven evaluation on three MoE models over a heterogeneous 10-server edge testbed shows that HetRoute reduces average inference latency by up to 59.0% and P99 latency by up to 58.0%, cuts cross-server traffic by up to 72.1%, and achieves 2.13x throughput improvement compared with representative baselines, while keeping quality degradation within the configured budget.

Vision-language MoE batches contain different numbers of image and text tokens. Image resolution, image count, tiling, and prompt length all change this token mix. We call the standard token-level Switch auxiliary loss Std-Aux. Std-Aux balances only the mixed load, so large image and text load errors can cancel at one mix. On our main model, the same trained router shows more than a fivefold change in load imbalance across image resolutions. We hold the image and text load profiles fixed and derive the exact load curve as the token mix varies. The image-text load gap controls sensitivity to the token mix. Physical preprocessing can also change the conditional profiles. The fixed-profile law excludes such changes. To design a remedy, we examine the router input structure. Image and text occupy distinct regions, while visual tokens group strongly by source image. The modality boundary motivates separate image and text terms. The image boundary motivates one equal-weight routing instance per image. ReBA, or Relax Within, Balance Across, implements both choices. Across four split backbones, ReBA lowers load on every reported benchmark input while keeping mean task accuracy comparable to Std-Aux. ReBA also lowers average load over the tested range and worst physical load under resolution and tiling shifts. Code is available at https://github.com/ZiangWu-77/ReBA.

Serving Mixture-of-Experts (MoE) large language models across distributed edge servers is bottlenecked by the cross-server expert transmission. The existing approaches mainly focus on how to reach a remote expert faster. However, in this paper, we instead consider whether a given layer, and the layers after it, need to be executed at all. To this end, a communication-aware adaptive-depth framework is proposed in this paper, termed TrimMoE, which couples layer skipping and confidence-based early exit with substitute execution and server-expert selection under a unified quality budget. Specifically, in the offline stage, TrimMoE freezes the backbone, trains the lightweight per-layer exit heads, calibrates the per-layer importance thresholds, and allocates the expert replicas by a skip/exit-aware redundancy benefit. In the online stage, a transition-aware look-ahead anticipates the token movement, so that the depth reduction targets the costliest transmissions, and besides, two feedback rules adapt the delay-quality weights and the exit threshold. Moreover, we prove that the substitution-and-skipping proxy degradation never exceeds the configured budget, and that the early exit is admitted only under a calibrated confidence gate. On a heterogeneous 10-server testbed with Switch-Base-8E, Qwen-MoE-A2.7B, and Mixtral-8x7B, TrimMoE reduces the average latency by up to 62.8%, lowers the cross-server traffic and the remote-execution ratio, and sustains high throughput under load, while keeping the task-quality degradation within a 2% bound.

Reward models are central to learning from human preferences, yet identifying what drives their predictions remains challenging. Recent sparse Mixture-of-Experts (MoE) reward models seek to improve interpretability by routing prompts to specialized experts and characterizing experts through examples with high routing weights. However, routing weights only reveal which prompts an expert $\textit{receives}$, not how it $\textit{judges}$ responses, providing only a partial account of expert behavior. We therefore propose $\textbf{Co}$ntribution-$\textbf{Co}$ntrast ($\textbf{CoCo}$) response-level interpretation, which faithfully characterizes experts' roles using chosen-rejected response pairs with the largest contribution contrasts, jointly capturing routing and preference behavior. Across automatic and human evaluations, CoCo yields more coherent, faithful, and specialized interpretations than router-based, score-based, and sparse autoencoder-based alternatives while maintaining competitive reward modeling accuracy. To the best of our knowledge, this is the first systematic study of interpretation methods for MoE reward models.

Recent byte-level large language models (LLMs) have made tokenizer-free modeling increasingly competitive by grouping bytes into dynamically sized patches. However, existing byte-patch architectures still apply the same dense feed-forward computation to every patch. This uniform computation cannot adapt model capacity to variations in patch semantics and granularity. We address this limitation with EntropyMoE, a Mixture-of-Experts (MoE) architecture designed for dynamic byte patches. EntropyMoE replaces the dense feed-forward modules in the global patch Transformer with Top-K expert layers. Each dynamic patch serves as the basic unit of expert routing, and its byte coverage determines its contribution to workload accounting. The router selects experts directly from patch entropy, using the same granularity signal that underlies dynamic patch construction to organize sparse computation. Patch entropy and length jointly define the feature space for regulating expert specialization. Experiments show that EntropyMoE achieves the lowest held-out bits-per-byte among matched dense and sparse baselines while maintaining comparable downstream accuracy. These results establish patch entropy as an effective routing coordinate for sparse conditional computation and extend Mixture-of-Experts modeling beyond tokenizer-based representations.

Implicit Neural Representations (INRs) encode signals as the weights of a coordinate-based neural network and have recently been proposed as an alternative domain for downstream learning. While promising, classification directly in weight space remains challenging due to the high dimensionality and complex structure of INR parameters. Furthermore, the way discriminative information is distributed across INR weights remains poorly understood. We propose a hierarchical Mixture-of-Experts (HMoE) Transformer that processes INR weights using conditional computation aligned with the structure of the underlying implicit network. Coupled with a meta-learning framework that shapes INR parameters for downstream tasks, our model achieves state-of-the-art accuracy across standard benchmarks, ranging from low-resolution datasets to high-resolution ImageNet-1K. To gain insight into how INRs encode discriminative information, we develop weight-space attribution and pruning methods that identify parameters most relevant for classification. These analyses reveal how class-specific structure emerges within INR layers and support the suitability of MoE architectures for weight-space learning. Our approach advances both the performance and interpretability of weight-space classifiers.

Adapting deep learning models to profound clinical heterogeneity typically relies on parameter-efficient fine-tuning (PEFT) to avoid the severe overfitting associated with full end-to-end network updates. Although PEFT successfully navigates limited data scenarios, it inherently forces the training of a separate, isolated adapter for every specific diagnostic task. Consolidating these isolated adapters into a single generalist network risks negative transfer, as optimization gradients from conflicting visual domains interfere. To address this, we propose MoPET, a mixture-of-experts (MoE) method that uses a learned sparse router to direct each input through a small subset of low-rank PEFT experts injected into a frozen foundation model, sharing capacity across datasets while limiting cross-domain gradient conflict. Through selected evaluations on the MedMNIST benchmark, we first establish that PEFT outperforms full network updates, improving average accuracy from 86.50% to 88.97%. We then show that a single MoPET model consolidates four heterogeneous datasets into one network, improving average accuracy over the best isolated PEFT adapters (93.46% versus 92.83%). Finally, we show that co-training with auxiliary datasets improves accuracy on data-constrained clinical targets, raising average target accuracy over the strongest isolated adapter from 81.58% to 83.58%. Our source code is publicly available at https://github.com/sdoerrich97/mopet.

Ambivalence and hesitancy (A/H) are subtle behavioural states that may be expressed through language, voice, facial activity, and other non-verbal cues. The ABAW11 A/H Video Recognition Challenge asks systems to assign a binary A/H label to each naturalistic interview video. Performance is measured using Macro-F1 so that recognition of both A/H and No-A/H samples receives equal importance. We present CALM-AH, a multimodal ensemble that combines textual, acoustic, visual, and derived behavioural-statistical features. We construct 15 non-empty combinations of these feature branches. For each combination, we select the best of three classifier families using validation binary cross-entropy and optimise its decision threshold for validation Macro-F1. The resulting binary decisions are combined using fixed hard-voting weights transferred from BROTHER. We further introduce Reliability-Gated Multi-Expert Consensus(RG-MEC), an anchor-preserving decision-level ensemble that combines an initial prediction with three complementary correction experts: CALM-AH, AffectGPT, and a GPT-based semantic verifier. The initial system provides the default prediction. Its label is overridden only when all three correction experts unanimously support the same alternative class; otherwise, the anchor prediction is retained. This unanimity-gated design limits the influence of isolated expert errors while permitting bidirectional correction when task-specific, multimodal-affective, and semantic-pragmatic evidence are fully consistent. On the participant-disjoint ABAW11 dataset, CALM-AH achieves a Macro-F1 of 0.7525, and the complete RG-MEC system achieves 0.7771.

Multi-modal object Re-Identification (ReID) aims to retrieve target instances by leveraging complementary information across modalities. However, existing methods suffer from two challenges. First, they often fail to exploit well-aligned and reliable semantic priors, making them vulnerable to background clutter and cross-modal misalignment. On the other hand, they typically rely on holistic feature modeling, overlooking the synergy between global and local representations. To overcome these limitations, we propose a robust multi-modal ReID framework with dual semantic guidance and global-local mutual modulation, which mainly consists of three key components, namely the Text-Semantic Injector (TSI), the Masked Global-Local Modulator (MGLM), and the Hierarchical MoE Fusion (HMF). The TSI enhances semantic awareness by integrating clean and coherent textual features into visual tokens. The MGLM enables part-aware cross-modal interaction through joint guidance from soft masks and global context, improving fine-grained feature alignment. Finally, the HMF adaptively aggregates multi-spectral features under local semantic supervision, yielding discriminative and robust representations. Extensive experiments on three multi-modal ReID benchmarks demonstrate the effectiveness of the proposed method. The code will be made publicly available at https://github.com/zw-absin/DSGM upon acceptance.

Multimodal medical anomaly detection identifies samples deviating from normal patterns, where scarce abnormal cases make normality modeling from normal data practical. In retinal Optical Coherence Tomography (OCT) and OCT Angiography (OCTA) anomaly detection, existing unsupervised methods rely on visual feature distributions, reconstruction residuals, or encoder-decoder discrepancies, making anomaly scores depend on appearance-level deviations, while multimodal normality also contains semantic organization described in normal medical reports. To this end, we propose Report-Guided Mixture-of-Experts (ReMoE), which distills normal report semantics into an image-to-text prior student, builds modality-aware priors, and uses Report-Guided Modality Modulation (RMM) to modulate features through mixture-of-experts routing. Experiments on a private OCT/OCTA dataset with paired normal reports and a public OCTA500-3MM setting using a fixed normal report demonstrate state-of-the-art performance.

Sparse mixture-of-experts (MoE) language models route each token to multiple experts, suggesting a geometric account of their benefit: co-selected experts should contribute distinct representation directions. Existing evidence often conflates route coherence, candidate quality, and candidate-by-context interaction. We distinguish these quantities using an Expert Subspace Separation Index (ESSI), matched-route residuals, and a prefix-controlled $2\times2$ factorial; frozen-route interventions and a controlled Top-$k$ study assess functional value. Three paired contrasts organize the findings. First, across six MoE architectures, expert subspaces overlap substantially, yet actual routes explain token representations better than matched alternatives. Second, across the 39 factorial cells in OLMoE, Mixtral, and DeepSeek, the selected candidate explains more of the residual representation than the strongest unselected rival in every cell, yet the actual prefix narrows this advantage throughout: all interactions are negative, and every 95% confidence interval lies below zero. Third, this geometric narrowing does not imply functional redundancy: adding later experts improves next-token prediction in 24 of 39 frozen-route comparisons, while the other 15 estimates are inconclusive; a controlled training study also favors Top-2 over Top-1 in all three seeds. We call this joint pattern coherent overlap: routing selects token-relevant experts from a shared geometric neighborhood, while useful multi-expert computation persists without disjoint linear coverage. Separating these quantities clarifies why geometric similarity alone cannot determine redundancy or pruning value.

Mathematically equivalent expert-reduction orders can produce observably different sparse-MoE executions. We isolate this effect in native DeepSeek-V4-Flash by freezing local MoE state and varying only aggregation semantics. Four schemes separate operand representation from accumulator precision. At one layer-5 fork, 720 A-mode orders yield 10 continuation basins; 720 B-mode orders form 360 exact structural classes and 11 basins. Under one Chinese prompt, the B classes split into 202 layoffs, 113 hiring, and 45 other continuations. Maximum-L-infinity B-branch selection separates 12, 24, and 36 of 50 prompts by 8, 16, and 32 tokens. Across 192 persistent trajectories per scheme, P32, A, and B change every native-reference route trajectory, while C preserves routes, token sequences, and texts. A separate 192-trajectory C check matches native MoE, post-mHC, next-router, and LM states bitwise. For one controlled B branch, exact post-mHC endpoint reconstruction reproduces the measured downstream trajectory. At the next decode boundary, exact FP64 reconstruction of the branch's full persistent state yields agreement for 301 downstream post-mHC states, 301 persistent-state checkpoints, 301 routes, predictions, and text over seven steps, given the same naturally generated next input. These controls identify post-mHC as an intra-token boundary and full persistent state as a cross-token continuation boundary. Identical tokens need not imply identical autoregressive state: divergence can survive a token boundary and become visible later. These results make expert operand conversion, accumulator precision, and reduction order part of a numerical compatibility contract for sparse-MoE runtimes and hardware backends. They establish controlled causal possibility, not deployment incidence; C's order invariance is limited to evaluated six-term states and schedules.

Vision-language models (VLMs), such as CLIP, are vulnerable to adversarial attacks, posing a serious problem for real-life applications and deployment. Adversarial fine-tuning emerges as a prominent defense method; however, different fine-tuning strategies often produce specialized models with distinct robustness characteristics. Each fine-tuned model in turn thrives in some evaluation settings but falters on others, limiting their defensive capabilities. We refer to these specialized fine-tuned models as robust model experts and propose a collaborative adversarial fine-tuning framework: CARE - Collaborative Adversarial Robustness fine-tuning using Embedding alignment. CARE maintains multiple experts during training, enables knowledge exchange through embedding-space harmonization, and consolidates the learned knowledge into a single unified robust model. Experts benefit from one another while preserving their individual specializations, enabling the final model to inherit complementary robustness properties. In this paper, we demonstrate CARE on two different adversarial fine-tuning strategies with complementary robustness behaviors. Extensive experiments on classic image classification and downstream vision-language tasks display the effectiveness of our approach, with CARE being able to outperform individually learned model experts. The results suggest that collaborative learning across model experts is a promising direction for improving adversarial robustness.

Heterogeneous recommendation feeds present complex challenges that extend beyond those found in highly homogeneous environments (e.g., music-only or video-only closed-ecosystem platforms). In Google Discover, a unified feed integrates diverse content sourced from the decentralized open web, including web articles, long-form and short-form videos, user-generated content (UGC), and beyond. Different content types exhibit distinct feature densities and user interaction patterns. Building a unified ranking model that sustains high performance across such heterogeneity, while avoiding negative transfer or majority bias, remains a significant industrial challenge. This paper presents an end-to-end case study on the industrial-scale multi-task ranking of heterogeneous feeds, grounded in real-world deployment. We introduce HA-MoE, a heterogeneity-adaptive multi-gated mixture-of-experts architecture that incorporates explicit heterogeneity context into both gating networks and expert representations. This approach enables effective specialization without significantly increasing operational overhead. To support reliable deployment, we introduce LENS, a lightweight observability framework that provides interpretable diagnostics of expert specialization and tracks this functional heterogeneity across continuous retraining. We evaluate our method using Dual-Level AUC (DL-AUC), a heterogeneity-aware evaluation metric that combines global ranking performance with cross-segment ranking correctness. Offline evaluations on a large-scale industrial dataset demonstrate consistent improvements over baseline models. Furthermore, online A/B testing confirms gains in feed activity and exploration metrics. Together, offline and online results validate the effectiveness of our approach for managing heterogeneity in industrial-scale recommender systems.

The promise of multimodal fusion lies in combining complementary sources of evidence, yet more evidence does not always yield a better prediction. Recent multimodal models have advanced fusion through richer cross-modal interaction and sample-adaptive fusion. However, the influence assigned to a modality during fusion does not reveal whether that source is unreliable, redundant, or poorly matched to a specialized expert. To address this limitation, we introduce TIER-MoE, a risk-guided subspace mixture-of-experts model that defines sample-specific modality reliability as the prediction loss its unimodal predictor is expected to incur. This risk is learned from out-of-fold predictions generated by models that were not trained on the corresponding sample. TIER-MoE combines the estimated risk with expert-specific subspace compatibility for sparse modality-expert routing, while an always-active shared path preserves multimodal complementarity. We evaluate TIER-MoE on four public multimodal biomedical datasets spanning Alzheimer's disease status, skin-lesion malignancy, and retinal classification. Results demonstrate its superiority over state-of-the-art methods in predictive performance and probability calibration, with consistent improvements in Macro-F1 and Brier score and strong zero-shot generalization to an external cohort.

The use of digital pens for online handwriting trajectory reconstruction is a prevalent method for human-computer interaction. In this study, we focus on a digital pen equipped with sensors where we aim at reconstructing the online handwriting trajectory. This pen enables writing on any surface and preserving the digital trace of handwriting. This type of pen could be used as an aid to learning to write in classroom. In this paper, we propose a new approach learning to finely reconstruct the touching trajectories while precisely analyzing the hovering part in order to position the next touching trace correctly. This relies on a Mixture-Of-Experts (MOE) approach. The first expert is dedicated for the pencil touch, and is named touching expert model. The second one is dedicated for the hovering pen trajectory, and is named hovering expert model. We improve on the learning of each of these experts based on additional context or specific examples. In addition we introduce a novel public benchmark dataset, to enable future research and comparisons in the field of handwriting reconstruction. The results demonstrates a significant enhancement compared to its primary competitors.

Mixture of Experts (MoE) architectures have become key to large language models; however, their typical round-robin (RR) scheduling introduces significant bottlenecks. In this paper, we demonstrate that RR causes a previously-undiscovered exponential incast phenomenon with MoE traffic. We propose an alternative proactive fair scheduling framework tailored for MoE workloads, which effectively prevents fabric oversubscription. We also outline how it can be implemented in NICs. Finally, through extensive simulations with real and synthetic workloads, we demonstrate that this framework consistently eliminates incast, maintains a near-100% link utilization, and reduces Collective Completion Time (CCT).

Glaucoma is a group of eye diseases that damage the optic nerve, often caused by elevated intraocular pressure. It is a leading cause of irreversible vision loss and is typically developed slowly and painlessly, making it difficult to notice until significant damage has occurred. Therefore, early detection is crucial to prevent or slow the progression of vision loss. In recent years, deep learning based uni-modal models have improved the accuracy and efficiency of glaucoma detection, empowering doctors with tools for earlier diagnosis, better monitoring, and timely treatment. Building on this, multi-modal models have emerged, leveraging the strengths of different imaging modalities to learn richer and more robust representations, further enhancing glaucoma detection accuracy. However, multi-modal learning faces challenges such as imbalanced and under-optimized uni-modal representations due to joint learning objectives. To address this, we propose a balanced soft mixture-experts model with three experts and load balancing loss. The performance is measured by AUC, our proposed method surpasses the performance of all uni-modal baselines, conventional multi-modal models, and current stateof- the-art balanced multi-modal models. The proposed model can be generalized to other disease detections such as diabetic retinopathy.

Biomedical image analysis spans diverse modalities and tasks, yet real-world deployment is hindered by severe distribution shifts across scanners, protocols, and patient populations. High-performing models consequently require repeated domain-specific fine-tuning, which is a costly cycle that becomes impractical when labels are scarce or privacy constraints limit data sharing. We propose OPERA (Offline Policy-guided Expert Routing and Adaptation), a multi-agent ensemble framework that addresses this deployment bottleneck by treating expert weight assignment as an offline policy learning problem: a routing policy is learned from a small validation set without gradient updates to any expert agent, then deployed with test-time adaptation to handle distribution shift. OPERA coordinates heterogeneous specialist agents through complementary mechanisms. The expert profiling module learns selection policies offline, enabling informed allocation of expertise. Each agent undergoes confidence calibration through temperature adjustment, ensuring more reliable probabilistic outputs. OPERA also incorporates distribution aware adaptation, where class weights are dynamically adjusted at the batch level using statistics derived from unlabeled test data. Instance level routing assigns each sample to the most suitable expert by leveraging inter model agreement and predictive entropy. We evaluate OPERA on 9 datasets covering fundus photography, chest X-ray, CT, MRI, and multimodal diagnostic benchmarks, comparing against 30+ baselines across classification, segmentation, and multimodal settings. OPERA consistently improves performance and calibration quality, demonstrating that offline policy-guided expert agents coordination is a practical path to deployable biomedical AI without retraining. Code is on \href{https://github.com/HUANGLIZI/OPERA}{GitHub}.

Model merging aims to combine multiple domain-specialized experts trained from a shared foundation model into a single multi-task model. Existing approaches largely focus on improving the merging procedure itself and typically assume experts obtained through full-parameter fine-tuning. In this work, we revisit expert training for model merging. We first show that prompt-based adaptation provides a strong baseline: independently learned prompts can be exploited across tasks while keeping the backbone fixed, avoiding the interference introduced by weight merging. Building on this observation, we introduce Dual-Tuned Experts (DTEs), a two-stage training strategy that first learns prompts and then fine-tunes the vision encoder. This reduces the magnitude of task-specific parameter updates and produces experts with higher merge compatibility. Experiments across multiple CLIP architectures, full fine-tuning, and LoRA experts show that DTEs consistently improve merged performance of standard merging approaches and remain effective even when combining heterogeneous sets of experts.

Large Mixture-of-Experts (MoE) language models are attractive for end-device deployment because only a small subset of experts is active per token, but their routed expert weights often exceed accelerator memory. We target latency-critical single-user settings where routed experts are staged on demand from CPU memory to a GPU or from Flash to a mobile NPU. In this setting, self-speculative decoding faces a new bottleneck: increasing the draft expert set improves accuracy but triggers extra expert loading, while cheap small-footprint drafts have low acceptance; moreover, verifying a multi-token block activates the union of target experts and is no longer close to one target step. We propose DraftExpert, an expansion-aware self-speculative decoding framework for expert-offloaded MoE inference. DraftExpert trains one lightweight accelerator-resident draft expert per layer by self-distilling residual, logit/token, and router-agreement signals from the frozen target MoE. At inference time, it uses a fixed-footprint shared+top-1+draft-expert drafter together with confidence--expansion truncation and target-expert prefetching, while final tokens are still exactly verified by the target model. On DeepSeek-V2-Lite and Moonlight-16B-A3B across CPU-GPU and Flash-NPU offload, DraftExpert improves decode throughput by 1.45x on average, raises draft acceptance to 8487%, and achieves 8688% prefetch hit rates.

A stochastic Gumbel-Top-$K$ router defines, for every token of a mixture-of-experts (MoE) model, a \emph{routing law}: a distribution over ordered expert lists and mixture weights. We ask which \emph{joint} distributions over the routing choices of different tokens are reachable while every individual token's complete routing law is held exactly fixed. We give a two-sided construction, \emph{Hierarchical Copula-Gumbel-Top-$K$} (\CGA{}). Within a group of related tokens, an exchangeable Gaussian copula positively correlates the Gumbel perturbations at each expert coordinate, which can increase within-group expert-set coherence. Across disjoint pairs of groups, a tunable antithetic construction introduces a selectable amount of negative dependence. We prove that both operations leave each token's ordered Top-$K$ sample, mixture weights, and inclusion probabilities identical in distribution to independent routing \emph{at a routing layer conditioned on its pre-routing logits}; conditional expected expert traffic is preserved as a consequence. We characterize the resulting trade-off: positive within-group coupling can only inflate the variance of realized expert loads relative to independent routing, while nonnegative cross-group opposition can only reduce it relative to flat coupling at the same within-group strength. Coherence and load dispersion are thus controlled by two complementary dependence dials on the invariance constraint surface. Because the base model is untouched, the dials can be driven by a small controller over frozen features, trainable with a score-function estimator: the frozen network is evaluated only in the forward direction, and gradients are confined to the controller. An initial small-scale pilot validates the mechanism and the training route, but does not establish task-level fine-tuning gains.

Mixture-of-Experts (MoE) architectures have been widely adopted in large language models, yet parameter-efficient fine-tuning (PEFT) for MoE models remains underexplored. Existing PEFT methods for MoE either ignore router priors with uniform adapters, reducing efficiency and risking forgetting, or rely on static expert selection, limiting per-token capacity and cross-expert feature learning. In this paper, we make the first attempt to fine-tune MoE models with MoE-style low-rank adaptation: our method, entitled MoE$^2$-LoRA, deeply couples the pretrained expert specialization with task-specific adaptivity via a dual-channel Routing-Conditioned Projection (RCP) module, which reuses base router activations to inform LoRA routing. We further introduce a single global LoRA expert pool shared across all layers, enabling model-wide adaptation with emergent layer-wise affinities and balanced expert utilization. MoE$^2$-LoRA simultaneously benefits from the advantages of prior reuse, dynamic adapter routing, and model-wide knowledge sharing. Evaluated on multiple MoE backbones with varying scales and expert granularities, MoE$^2$-LoRA consistently achieves state-of-the-art downstream accuracy while retaining stronger general capabilities.

Federated prompt tuning (FPT) enables collaborative adaptation of vision--language models (VLMs) using lightweight prompts. Existing methods often address heterogeneity and privacy through a split-prompt design under local differential privacy (DP), combining a shared prompt for global transfer with private prompts for local adaptation. However, a single shared prompt may over-smooth diverse transferable knowledge, weakening the balance between personalization and generalization. Multi-expert prompts (MEPs) can better capture this diversity, but enlarge the communicated space, increasing DP noise and communication cost while making robust expert composition more difficult. We propose FedSEPT, a privacy-preserving Fed}erated Subspace-decomposed Expert Prompt Tuning. Specifically, we employ Subspace-decomposed Expert Modeling (SEM) to parameterize multiple prompt experts with shared low-rank factors, a fixed public basis, and private residuals, thereby confining communication and DP perturbation to a compact factor space while enabling direct server aggregation in a common coordinate system. We further design Instance-aware Expert Fusion (IEF), which adaptively combines semantically complementary experts via on-device routing and performs efficient logit-level fusion using cached expert-specific text features. Extensive experiments on 11 heterogeneous benchmarks show that, under the same privacy constraints, FedSEPT achieves a better trade-off between local adaptation and global generalization than strong baselines.

Efficient multimodal inference is increasingly constrained not only by model quality or FLOP count, but also by the cost of preserving, moving, routing, caching, and quantizing multimodal representations under latency, memory, and energy constraints. This paper reviews recent advances in efficient vision-language and multimodal large language models, covering visual token compression, video token management, KV-cache optimization, Mixture-of-Experts (MoE) routing, low-bit quantization, edge deployment, and hardware-aware benchmarking. We argue that these techniques cannot be treated as independent optimizations. Visual token compression alters downstream feature distributions and MoE routing decisions, routing behavior affects expert utilization and quantization sensitivity, quantized router logits influence expert assignment, KV-cache policies determine retained multimodal evidence, and hardware constraints often transform computational savings into memory and communication bottlenecks. We organize the literature around these interactions and identify key design trade-offs, including accuracy versus token budget, static versus adaptive compression, sparse routing efficiency versus expert collapse, and low-bit inference versus modality-specific degradation. Finally, we introduce Temporal Routing Consistency as a diagnostic for video MoE models and highlight open research directions in routing-aware compression, cross-modal cache management, hardware-aware co-design, and unified benchmarking for multimodal edge intelligence.

We consider the problem of learning compositional robot policies end-to-end from expert demonstrations, without any pre-specified notion of task decomposition or hierarchy. We ask whether a VLA trained with a simplified Mixture-of-Experts (MoE) action head can emergently learn to decompose tasks into reusable, interpretable primitives. We find that learned experts are heavily reused across tasks and consistently correspond to qualitatively distinct low-level behaviors, suggesting that the router implicitly learns to perform high-level sequencing while experts serve as compositional primitives. Our MoE matches the task performance of a monolithic baseline while demonstrating meaningful expert specialization, a step toward modular, interpretable robot policies that emerge from data alone.

Full-parameter post-training of trillion-parameter-scale MoE models introduces substantial system-level challenges for large-scale distributed training, including severe memory pressure, non-overlapped communication overhead, and inefficient kernel execution. While most large-scale LLM training systems are built around GPU-based clusters, this report presents an end-to-end optimization practice on the Ascend NPU SuperPOD. Using the DeepSeek-V4 model family as the target workload, we develop a hierarchical optimization framework spanning model-level parallelism, computation-communication orchestration, and low-level kernel execution. The resulting system achieves 34.22% Model FLOPs Utilization (MFU) with a 2.93x improvement over the open-source baseline recipe while maintaining training stability. Building on this optimized infrastructure, we further establish a CPT and SFT workflow for complex Operations Research (OR) tasks. We refer to the integrated framework as SLAI T-Rex. Using DeepSeek-V4-Flash, we develop OR-oriented CPT and SFT data pipelines that combine collected domain resources with solver-verified synthetic optimization documents. The resulting dataset contains 10K high-quality SFT samples spanning four task categories and three problem representations. The specialized model achieves the highest average zero-shot Pass@1 score among the evaluated models, reaching 71.81% and outperforming GPT-5.4-Mini and the base DeepSeek-V4-Flash model by 3.98 and 11.27 percentage points, respectively. Overall, this work demonstrates a full-stack pathway from efficient trillion-parameter model post-training on Ascend infra to domain-specialized Flash models for solver-grounded mathematical modeling, advancing frontier-model systems for complex reasoning.

Mixture-of-Experts (MoE) architectures increase model capacity without proportionally increasing computation cost and have become a key building block for scaling large language models (LLMs) to trillion-parameter regimes. Efficient deployment of these MoE models relies on distributed execution across multiple GPUs, where each MoE layer involves two all-to-all communications: dispatching tokens to expert ranks and returning the expert outputs to their source ranks. Conventional MoE implementations launch this return all-to-all after expert compute completes, exposing communication latency on the critical path and reducing GPU utilization. We present a fine-grained approach that overlaps expert compute with the second all-to-all via tile-level signaling and scheduling. Our producer-consumer co-design combines: (1) a persistent per-rank computation kernel (producer) that covers all local experts on the rank to eliminate repeated kernel launch overhead and prioritizes remote-critical tiles, and (2) a persistent communication kernel (consumer) on a small dedicated partition of streaming multiprocessors (SMs) that issues segment-granular transfers as tiles become ready. Our co-design avoids intrusive changes to the underlying computation operators or communication primitives, making it practical for improving distributed MoE execution efficiency on multi-GPU systems. On a 4-A100 GPU platform, evaluated on three MoE models against four state-of-the-art MoE systems, our approach achieves up to 2.64x end-to-end speedup and 2.74x MoE-layer speedup. Compared with a conventional non-overlap baseline, our approach consistently improves both operator- and MoE-layer-level performance across varying GEMM shapes, router modes, and a broad range of producer/consumer SM partitions, while preserving correctness.

Optimizer state is the largest single line item in the memory budget of mixture-of-experts (MoE) training. On a 6.78B-parameter MoE language model AdamW keeps 50.6 GB of first and second moments to update 12.6 GB of bfloat16 weights. We study SkewAdam, an optimizer built on the observation that the three parameter populations of an MoE differ enough in size and gradient statistics that they should not receive the same state. Those populations are the dense backbone, the experts and the router. SkewAdam keeps float32 momentum plus a factored second moment for the backbone (5% of parameters), a factored second moment alone for the experts (95%) and an exact second moment for the router (<0.01%). The resulting state occupies 1.29 GB or 2.6% of AdamW's and peak training memory falls from 81.4 GB to 31.3 GB, within the budget of a 40 GB accelerator. In a controlled comparison from identical initializations over 82M tokens, SkewAdam reaches validation perplexity 108.4, ahead of AdamW (126.8), Muon (120.2) and Lion (393.7), and settles router load balance to within 1% of its uniform floor. The allocation is not what earns that perplexity. A tier ablation reaches the same value while carrying twenty times the state, so the tiers buy memory rather than accuracy. Same-platform runs separate what does earn it. Removing momentum costs 31 perplexity points (tuned Adafactor, 139.7) and replacing the factored second moment and its update clipping with a full second moment costs 10 (tuned AdamW, 118.5), so neither tuned baseline reaches the untuned tiered policy. Where optimizer state lives, these results suggest, matters at least as much as how much of it there is.

Electroencephalographic (EEG) abnormalities arise from dynamic changes in neural synchrony across spatial and temporal scales, yet many computational approaches reduce these dynamics to static features. We present a Spatial Multi-Expert Graph Transformer that models each EEG recording as a sequence of dynamic functional connectivity graphs. Time-resolved connectivity is estimated using the weighted Phase Lag Index (wPLI), and hierarchical graph encoding aggregates information from electrode to regional and global levels. A multi-expert transformer architecture enables subtype-aware reasoning, with a gating mechanism adaptively fusing expert outputs for global abnormality prediction. Experiments on the TUAB dataset show competitive abnormal EEG detection performance and demonstrate the potential of dynamic graph modeling with adaptive expert fusion for interpretable, subtype-aware spatial--temporal analysis.

Although mixture-of-experts, MoE, models have been increasingly adopted to scale large language models with moderate computation cost, it remains challenging to deploy MoE inference over resource-constrained and bandwidth-limited edge infrastructures. Existing distributed MoE serving methods mainly rely on exact expert placement, caching, replication, or communication scheduling, while overlooking the functional similarity among experts, which provides an opportunity to reduce cross-server token transmission. Therefore, this paper introduces a similarity-aware expert allocation and distributed deployment framework, dubbed OrderMoE, which aims to accelerate edge MoE inference while balancing inference latency, communication overhead, server workload, and inference quality. OrderMoE first constructs an expert similarity model based on router-induced logits representations and partitions experts in each MoE layer into multiple similarity groups. Then, it develops a similarity-aware expert grouping and deployment strategy to improve local similarity coverage across edge servers. Since reducing remote expert invocation and preserving exact inference quality are conflicting objectives, OrderMoE further designs a quality-aware and trajectory-aware runtime server-expert selection algorithm to decide whether a token should invoke its remote target expert or use a feasible local substitute expert. Experimental results on a real distributed edge testbed show that OrderMoE significantly reduces average latency, tail latency, cross-server traffic, and remote expert invocation ratio, while introducing only small and controllable inference quality degradation.

Time series forecasting (TSF) is vital to many applications, yet existing models often struggle to capture the heterogeneous long-range global patterns and short-range local variations in multivariate time series. While some approaches partially model these dependencies, they often do not jointly exploit temporal and feature-wise information. To address this challenge, we propose HyBDM, a multi-scale hybrid model that decomposes temporal dynamics into global patterns and local variations, which are modeled by two specialized experts. The Global Patterns Expert employs an enhanced BiConv-Mamba module that integrates bidirectional convolutions, an M-SSM layer, a forgetting mechanism, and a GDD-MLP module for cross-channel modeling. The Local Variations Expert uses a Local Window Transformer (LWT) to perform efficient locality-aware attention with reduced computational complexity. In addition, a Multi-Scale Patcher and a Long-Short Router enable multi-resolution representations and adaptive fusion of the two experts. Experiments on six benchmark datasets show that HyBDM outperforms state-of-the-art methods in both forecasting accuracy and computational efficiency, demonstrating its effectiveness in bridging global-local dependencies for multivariate TSF.

The strongest open-weight coding models are mixture-of-experts (MoE) networks: most of their size comes from large pools of "expert" subnetworks, of which only a few act on any token. That pool is why these models do not fit on the machines most developers own, yet for a user who only wants coding help, most experts encode abilities that will never be invoked. We ask how many experts can be removed, and which, by pruning two recent open-weight MoE models from different families (Qwen3.6-35B-A3B and Gemma-4-26B-A4B) under five selection strategies, judged the way a user would: by whether the model still writes correct code. Half the experts can be removed from either model with no statistically detectable loss on the primary code benchmark, and the damage lands almost entirely on abilities outside coding, the intended trade. But the winning strategy flips between the two models, so a recipe validated on one family cannot be assumed to work on another. We further show that perplexity, the metric much of the pruning literature leans on, can rate a broken model above an intact one; that a lightweight fine-tune recovers about half of what aggressive pruning loses; and that against quantizing the full model to the same memory, pruning wins only where quantization would have to drop below 3 bits per weight. Five attempts to overturn that crossover, with failure criteria fixed in advance (better calibration, guarded selection, causal expert importance, failure attribution, and an agentic evaluation letting each model repair its failures from execution feedback), all leave it standing; the last shows single-shot benchmarks overstate compression penalties broadly, as one repair turn erases the 2-bit quantization penalty entirely. Expert pruning works, but it demands per-model validation on the task the model will actually serve.