Mixture-of-experts (MoE) architectures scale large language models efficiently, but they demand massive GPU memory. To cope with such demand, models are commonly compressed to reduce their memory footprint. Residual sparsification is a representative compression technique that decomposes each projection matrix of an expert into a shared base matrix and per-expert residual matrix, and then compresses the residuals. Existing sparsification methods compress each residual matrix independently by minimizing its compression error, thereby minimizing the error of each projection matrix. However, our analysis shows that this objective is misaligned with preserving model accuracy after compression. In an expert, the final output is produced through computations coupled across multiple projections and hidden representations. Therefore, even small errors in individual matrices can propagate through hidden representations and projection interactions, leading to large expert output errors and accuracy degradation. To address this misalignment, we propose PARSER, a new residual sparsification method that shifts the compression objective from minimizing isolated matrix errors to preserving the expert output error. PARSER achieves this by introducing output importance, which measures the actual contribution to the expert output error. Our experiments show that, compared with existing methods, PARSER narrows the accuracy gap to the uncompressed model by 1.41$\times$ on Qwen and 1.44$\times$ on DeepSeek, while achieving the same peak memory reduction.
Xiaoyang Lu, Belthangady Akash Vi Narayana Pai, Xian-He Suncs.AR cs.LG
Mixture-of-Experts (MoE) models enable efficient scaling of large language model (LLM) inference but suffer from substantial data-movement overhead when deployed on neural processing unit (NPU)-based systems. Near-Data Processing (NDP) provides a promising way to mitigate this bottleneck via cooperative NPU-NDP execution. However, existing NPU-NDP MoE systems do not fully account for hardware heterogeneity, dynamic expert-level concurrency, and temporal expert reuse during batched inference. This paper presents DynaNDE, a dynamic near-data expert scheduling framework that exploits NPU-NDP collaboration to accelerate batched MoE inference. DynaNDE introduces an analytical performance model that captures hardware heterogeneity, data-movement costs, and communication-computation overlap in cooperative NPU-NDP execution. Guided by this model, DynaNDE determines per-layer expert scheduling across the NPU and NDP while accounting for expert-level concurrency. DynaNDE also incorporates a reuse-aware runtime that avoids redundant parameter movement when experts reside in NPU memory. Experimental results show that DynaNDE achieves substantial throughput improvements over the state-of-the-art NPU-NDP MoE serving framework, with average speedups of 2.6$\times$ and 2.2$\times$ for the prefill and decoding stages, respectively.
As large language models (LLMs) continue to scale, parameter-efficient fine-tuning (PEFT) has become a practical alternative to full-parameter adaptation. Prompt tuning is effective, but existing approaches either use flat prompt structures or hierarchical structures with fixed prompt composition, limiting adaptive prompt specialization. To address this limitation, we propose HiVe, a prompt tuning framework that models prompts at multiple levels and enables input-dependent specialization. HiVe constructs a prompt hierarchy by leveraging inter-task relationships during training, and employs a vertical mixture-of-experts (V-MoE) mechanism at inference time to compose prompts up to the level of specialization required for each input. Experiments show that HiVe consistently outperforms strong prompt tuning baselines across diverse tasks.
Single-token autoregressive decode on CPUs is bound by memory bandwidth, not arithmetic: a modern CPU sustains roughly 1 TFLOP/s of compute but only about 50 GB/s from main memory, and each generated token must stream every active weight once. This report argues that the most effective response is to co-design the model architecture and the inference runtime together. It presents cflow, a CPU-first streaming engine, alongside a family of pipeline-native transformer architectures whose inter-layer dependency graphs are constructed to permit a vertical, stage-major execution schedule. cflow stores weights as L2-sized tiles in compute-consumption order, reads only the top-k experts of each mixture-of-experts layer, fuses projections, and executes a delay-aware schedule from per-model dependency parameters. Across five architectures trained on TinyStories, one (arch2_4_combined) achieves a 2.00x reduction in critical-path weight bandwidth (9.00 to 4.50 MB/token) within 0.24 perplexity of the best candidate, and the tile layout incurs 7.29x fewer L1-data read misses than a row-major baseline. On a 30.9-billion-parameter pipeline-native MoE, cflow decodes at 5.94 tokens/s (tok/s) on a 32-vCPU Ice Lake server, ahead of llama.cpp (4.75) and the vLLM CPU backend (1.65) on comparably sized dense models. Realizing the expert-delay window as asynchronous I/O overlap on a disk-resident expert tier yields a further net win of up to 1.68x, matching the overlap model within 1%. Measurement refutes one of the eight design claims and leaves a second inconclusive; both are reported in full, with the conditions under which they would hold.
Mixture-of-Experts (MoE) models scale capacity for strong quality while keeping per-token compute bounded through sparse expert activation. Yet low-latency MoE serving is increasingly challenging, because it spans two inference phases with fundamentally different bottlenecks: prefill is dominated by token-wise expert computation, whereas decode is constrained by memory traffic from the batch-wise activated expert set. However, existing training-free acceleration methods optimize only a single resource proxy, either the experts each token executes or the experts a batch activates, and either discard the excluded experts' contribution or leave it only implicitly approximated. In this paper, we propose ExFold, a unified training-free expert-folding framework for jointly accelerating MoE prefill and decode. ExFold casts both prefill and decode as one budgeted output-approximation problem: execute only a phase-specific constrained expert set while projecting the contribution of budget-excluded experts onto retained experts using calibrated scalar projectors. Motivated by the observation that many expert outputs are directionally aligned but differ in magnitude, ExFold calibrates a pairwise scalar-projector matrix on unlabeled data and uses it at inference time to fold excluded expert contributions into retained experts. Under this view, prefill acceleration becomes token-level Top-K folding, and decode acceleration becomes batch-level expert-pool folding. The two phases differ only in how retained experts are selected, while excluded contributions are recovered by one shared folding mechanism. We implement ExFold as a plug-and-play plugin in vLLM, with a lightweight expert-folding CUDA kernel, delivering up to 1.41x TTFT and 2.45x TPOT speedups while retaining about 99% of the original average quality.
Chain-of-thought (CoT) prompting improves LLM reasoning by decomposing complex problems into intermediate steps, but its sequential nature increases decoding latency and memory usage. Mixture-of-Experts (MoE) models scale capacity through sparse expert activation, yet their full expert weights often exceed GPU memory and require costly GPU-CPU transfers. Existing runtimes treat all tokens uniformly, overlooking a key structural property of CoT traces: consecutive reasoning stages exhibit coherent and predictable expert activation patterns. Ignoring this stage-level regularity leads to inefficient caching and unnecessary data movement. We propose SAEM, a stage-aware MoE inference runtime that detects reasoning stage boundaries and exploits stage-level activation coherence to guide expert placement. SAEM combines stage-aware caching, expert-aligned token repacking, and in-situ CPU execution to reduce data transfer and kernel fragmentation. On mathematical and scientific reasoning workloads, SAEM achieves an average 1.33x throughput improvement over the strongest state-of-the-art caching and offloading baselines under constrained GPU memory, rising to 1.54x when calibration data matches the workload, demonstrating the effectiveness of stage-aware, locality-driven MoE inference for CoT reasoning.
Sparse mixture-of-experts (MoE) language models reduce arithmetic by activating only a small subset of experts per token, yet deployment still requires storing and moving the full expert bank. We present ExactMoE, an inference design that applies symmetric group-128 four-bit weight quantization only to routed experts, stores those experts in kernel-native MARLIN form in pinned host memory, and executes all selected experts through a configurable GPU-resident slot cache and fused grouped MoE kernels. The router, attention, embeddings, normalization layers, and language-model head remain in BF16. "Exact" refers to complete expert availability and an unchanged top-k routing procedure: no expert is pruned, substituted, or forced to execute on the CPU. It does not imply numerical identity with the BF16 model. On OLMoE-1B-7B-0924-Instruct, evaluated on a single NVIDIA L4, a 16-slot configuration reduces peak reserved GPU memory from 14.168 to 1.836 GiB (87.04%) while retaining 81.85% of BF16 decode throughput. A fully resident 64-slot configuration reaches 31.923 tokens/s versus 21.662 tokens/s for BF16 while reserving 4.061 GiB. Across 12,450 zero-shot multiple-choice questions, ExactMoE obtains 70.3534% normalized accuracy versus 70.8996% for BF16, retaining 99.23% of the baseline accuracy. In a matched 16-token ablation, fused grouped execution is 1.97x as fast as a sequential W4 reference. These results identify a practical memory-transfer-throughput frontier for complete-expert MoE inference.
Dynamic Mixture-of-Experts Serving allocates k replica GPUs among m experts as workloads change. At each round, the online algorithm sees the current workload, chooses integral replica counts, and pays bottleneck service cost plus replica movement. It does not know future workloads. Huang, Lou, and Xiao gave an O(sqrt(log k))-competitive randomized algorithm for this problem. We prove a deterministic O(1)-competitive algorithm. For every number of experts and every k>=1, the algorithm satisfies ALG_det <= 10 C_PB OPT + (5 C_PB + 8) k + 16, where C_PB is the absolute constant from Chasing Positive Bodies at resource augmentation one and covering sparsity two. Consequently, CR_det(k)<=10 C_PB for every k>=1, so CR_det(k)=Theta(1). The multiplicative factor does not depend on the number of experts, replica budget, horizon, or workload values. Thus randomization is not needed for the asymptotic guarantee. The proof has two layers. A finite tangent envelope, summable positive resets, and a nonexpansive balanced projection reduce reciprocal-max service costs to a deterministic exact-budget fractional path. A new deterministic rounding theorem converts every such path to integral allocations with service distortion three and movement bounded by the fractional movement plus 6k. The complete reduction, rounding theorem, causal composition, and quantified main theorem are machine-checked in Lean 4 relative to the positive-body result as the sole scientific source premise. The theorem concerns the allocation model above. It does not include network topology, shared-edge congestion, or routing decisions.
Financial volatility is regime dependent, yet incorporating regime information into neural networks can also destabilize training. This paper asks where such information should enter a neural cross-sectional volatility forecasting model. We study five-day realized-volatility forecasts for 1,027 U.S. equities using a rolling walk-forward evaluation framework in which information, model capacity, hyperparameter tuning, and random seeds are matched across architectures. We propose RG-ResMoE, a regime-gated residual mixture-of-experts architecture in which regime information is used only for expert routing rather than for direct forecasting. The base predictor models volatility from stock features, while a gating network uses regime state variables to route residual corrections. RG-ResMoE consistently outperforms a capacity-matched MLP in both forecasting accuracy and training stability in the main U.S. study. Similar gains are observed on an independent Japanese panel. The integration pathway is decisive: appending the same regime variables directly to the forecasting input degrades both predictive performance and training stability, whereas restricting them to the routing gate improves accuracy and Value-at-Risk calibration. Hard routing consistently underperforms soft routing. The results suggest that, in compact neural volatility forecasting models, the primary value of mixture-of-experts models lies less in increasing model capacity than in controlling how nonstationary regime information influences prediction.
Large language models (LLMs) have achieved remarkable breakthroughs across various applications. However, their architectures remain inefficient in pretraining due to two main limitations: (i) self-attention lacks an explicit inductive bias for locality, leading to redundant modeling of sequence-internal local information; (ii) mixture-of-experts (MoE) implicitly couples knowledge storage with computational pathways, hindering flexible access to sequence-external global knowledge. To overcome these limitations, we propose LoKiFormer, a novel LLM architecture that augments the standard decoder with two dedicated modules: 1) Local Fusion Attention (LFA), which incorporates a convolutional fusion to attention, explicitly capturing local patterns and allowing the attention to operate on more informative representations; 2) Knowledge Memory Module (KMM), which introduces a parametric key-value memory that explicitly stores global knowledge in addressable slots, decoupling storage from computation and enabling direct knowledge retrieval. Together, these modules enable LoKiFormer to achieve more efficient and effective integration of information at both levels. Experimental results show that LoKiFormer converges 1.33x faster in pre-training than baseline models, underscoring its superiority over existing LLM architectures.
Alish Kanani, Layan Badawi, Umit Y. Ograscs.AR cs.AI cs.LG
Mixture-of-Experts (MoE) models are attractive for edge deployment because they provide high model capacity while activating only a small subset of parameters per token, improving compute efficiency. However, MoE inference at the edge is fundamentally limited by memory. Expert parameters are large and often reside in off-chip memory due to capacity, cost, and power constraints, putting expert loading to the critical path. We present APEX: Adaptive Expert Prefetching, a predictive resource management framework that overlaps expert loading with useful computation. APEX introduces a lightweight prefetch router that predicts candidate experts before the attention block to dynamically fetch additional experts using a learned confidence model. This adaptive strategy achieves over 99% overlap accuracy, significantly outperforming fixed top-k prefetching techniques. APEX supports two execution modes: a correctness-preserving mode that guarantees exact routing semantics, and a stall-free mode that eliminates residual stalls by operating on available experts with negligible impact on application accuracy. Across multiple MoE models, the correctness-preserving mode reduces per-token latency by up to 26% and improves energy-delay product (EDP) by up to 41% over state-of-the-art baselines, while the stall-free mode provides additional efficiency gains with negligible impact on application accuracy. These results establish adaptive, confidence-driven expert prefetching as an effective approach for efficient MoE inference on edge systems.
Junghwan Lim, Joon Son Chung, Sungmin Lee +24cs.AI
We introduce Motif 3, a decoder-only Mixture-of-Experts language model with 314 billion total parameters and 13.2 billion activated per token. Each sparse MoE layer contains 384 routed experts, with eight selected per token. This fine-grained sparsity provides substantial expert capacity while limiting computation. Motif 3 is built around Grouped Differential Latent Attention (GDLA), which integrates grouped differential attention with the compressed key-value representation of Multi-head Latent Attention. The architecture further incorporates modified manifold-constrained hyper-connections, Expert Specific PolyNorm activations, and multi-token prediction to improve optimization stability, expert specialization, and inference efficiency. We pretrain Motif 3 on approximately 12.5 trillion tokens spanning web documents, STEM, code, mathematics, multilingual content, and domain-specialized corpora. Expert-balancing and numerical-stabilization techniques support stable training at scale, while selective MXFP8 computation and communication, memory-efficient fused kernels, and window-aware context parallelism enable training with context lengths up to 256K tokens. Our post-training pipeline combines general supervised fine-tuning, six specialist teachers trained with reinforcement learning, a software-engineering teacher trained with supervised fine-tuning, and Multi-teacher On-Policy Distillation. The resulting unified model consolidates complementary capabilities in reasoning, coding, tool use, professional work, long-context understanding, calibrated abstention, and instruction following. Across a broad evaluation suite, Motif 3 demonstrates competitive performance against leading open weight models, including strong results on long-horizon agentic tasks, mathematical reasoning, scientific knowledge, and hallucination-sensitive evaluation.
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.
Mixture-of-Experts (MoE) models have outgrown accelerator memory, and offloading expert weights to host memory is now standard. This makes expert cache management an attractive lever: a policy that raised the hit rate would cut expert traffic per token. Evaluating that is a measurement problem, and we find the measurement fragile. With a trace-driven, event-atomic simulator over three MoE models (40, 64, 128 experts), we isolate three evaluation axes that change conclusions, not just numbers. Replay semantics: under a fused-event traffic contract, an inconsistent per-access replay inflates recency-based policies by 27-29% while leaving frequency-based and static ones within 4%, inverting the policy ranking. Workload contamination: probe sets using one instruction template per category produce verbatim-identical generation prefixes; a matched-pair rendering intervention moves the measured early-window effect by 19.4-31.9 points and reverses which workloads look most cache-friendly. Operating regimes: normalized miss fractions do not transfer across models, so the per-step expert union relative to per-layer capacity must be reported -- yet permuting only the temporal order of an identical event stream moves the offline-optimal gap from 44.9% to 30.8%, so it is not sufficient. Corrected, a stable gap to the offline optimum remains (44.2-45.9% over 13 frozen workload compositions). A forced-admission oracle attributes 84.3-96.6% of it to knowing which resident expert is used furthest in the future. A causal next-use predictor, used as an eviction rule, recovers -11.4% of the gap; it picks an optimal victim 3.4% of the time, against 2.4% for a random resident block and 20.6-22.1% for LRU and LFRU. Our position is narrow: in our evaluated settings a large offline-optimal gap substantially overstates the gains recovered by representative lightweight causal mechanisms.
Sangwoo Ha, Hyunwoo Seo, Yurim Jo +2cs.AR cs.CL cs.LG
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.
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.
Speculative decoding verifies a tree of draft tokens in one target-model forward pass. For a mixture-of-experts (MoE) target, however, parallel verification can activate the union of the experts selected by all tree nodes, even though only a small subset of those nodes reaches the accepted output. Token count, activated-expert union size, and expert-weight traffic are therefore distinct cost measures: reducing the token workload need not shrink the expert union proportionally, and under offloading, transfer traffic also depends on cache residency. We introduce AcceptMoE, a verifier-side expert selector that combines target-router scores with offline-estimated commitment probabilities and automatically adjusts the number of eligible experts for each verification block, eliminating the need for a user-specified expert budget. Under offloading, AcceptMoE conditions expert eligibility on cache residency instead of predicting natural routes and prefetching the corresponding expert weights. Although constraining target-expert eligibility changes the model distribution, across 12 model-task pairs spanning three MoE targets and four benchmarks, AcceptMoE's mean accuracy is 0.27 percentage points lower than that of EAGLE-3 speculative decoding with natural routing. Served with SGLang at batch size one, it reaches 1.290 times the throughput of this baseline with all expert weights in GPU memory, and 2.06 times under physical expert offloading, while reducing host-to-device traffic by 73.6 percent to 77.1 percent.
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.
DiffusionGemma Team, Adrien Ali Taïga, James Assiene +41cs.CL cs.AI
We introduce DiffusionGemma, an experimental open-weight language model that uses discrete diffusion to generate text at exceptionally high speed. Rather than decoding one token at a time, DiffusionGemma iteratively refines blocks of 256 tokens in parallel, avoiding the sequential decoding bottleneck of conventional autoregressive (AR) large language models. Instead of training from scratch, we obtain DiffusionGemma by fine-tuning the mixture-of-experts Gemma 4 model with 3.8B activated and 25.2B total parameters. Our compute-efficient two-stage training pipeline uses fewer than 10% of the starting AR model's total training token budget. The first stage uses supervised fine-tuning to teach bidirectional denoising, while the second stage combines reinforcement learning with sampler distillation to jointly improve generation quality and inference efficiency. DiffusionGemma establishes a new Pareto frontier for the trade-off between generation speed and model capability. Averaged across our full evaluation suite, it generates around 20 tokens per forward pass and achieves roughly 1,500 output tokens per second on a single NVIDIA H100 GPU, which is substantially faster than AR models even with state-of-the-art speculative decoding. DiffusionGemma also retains the starting model's support for thinking mode, multimodal inputs, and long contexts. Despite diffusion fine-tuning, it remains capable of AR generation with only minor performance degradation, suggesting a path toward hybrid diffusion-AR decoding.
Mixture-of-Experts (MoE) variants of Low-Rank Adaptation (LoRA) route every token to a fixed number of experts $k$. Tokens differ in how uncertain the model is about them, so a single k over-spends on easy tokens and under-serves hard ones. We observe that the router's output distribution is already a per-token uncertainty signal: peaked mass indicates confidence, while a flat distribution indicates ambiguity. We introduce CARE (Confidence-Adaptive Routing of Experts), which admits experts in a nucleus fashion. Experts are activated in decreasing router weight until their cumulative mass reaches a threshold, with a small extension when the admitted experts disagree. A budget thermostat calibrates the threshold so that the average number of active experts matches any target. CARE is a drop-in, single-forward-pass rule with no extra parameters. Across eight commonsense benchmarks on LLaMA-3.1-8B and Qwen2.5-7B, as well as math, code, and knowledge tasks, CARE improves over fixed top-k MoE-LoRA at matched compute and matches the fixed-k=4 baseline while activating fewer experts. The same confidence and disagreement signals also improve out-of-distribution detection over MSP, entropy, and multi-pass proxies. We support the design with nucleus fidelity, budget optimality, and an epistemic reading of disagreement, and we release code.
The discovery of scaling laws has motivated training neural networks on ever increasing quantities of data. This is typically done with a constant decoupled weight decay which causes the network weights to shrink steadily over the course of training. Taking inspiration from the Robbins--Monro conditions, we propose to scale weight decay by the fraction of the peak learning rate $η/η_{\max}$. We prove that this scaled weight decay preserves the asymptotic stationarity guarantees of the corresponding unregularized methods for both stochastic gradient descent and the non-Euclidean spectral optimizer Muon, thereby avoiding the additional asymptotic bias introduced by constant decoupled weight decay. This retains the stability benefits of weight decay without changing the asymptotic optimization target. Using a steady-state analysis, we explain why under standard weight decay the weight norm shrinks steadily as training proceeds, whereas under scaled weight decay it settles to a roughly constant value. When applied to the training of mixture-of-experts models, Muon with scaled weight decay (Muon-SW) consistently outpaces Muon with identical hyperparameters, reaching the same validation loss $\mathbf{30\%}$ faster at our largest scale across models from $72 - 930$ million parameters trained at $\sim 600$ tokens per active parameter. If this trend continues to hold, the method promises to substantially accelerate the pre-training of frontier models while requiring only a few lines of code to implement.
A stochastic Gumbel-Top-K router defines, for every token of a mixture-of-experts (MoE) model, a routing law: a distribution over ordered expert lists and mixture weights. We ask which 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, 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 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.
Implicit neural representations (INRs) for time-varying volumetric data are typically trained using dense sampling over spatiotemporal coordinates, where each observation corresponds to a single point in space and time. This coordinate-wise formulation requires extensive sampling during optimization, leading to high computational cost and inefficient use of temporal structure. In this work, we revisit this design choice and show that dense spatiotemporal sampling is not necessary for learning time-varying fields. Instead, we represent the data as a collection of spatially indexed time series and train INRs using sequence-level supervision over each spatial location, rather than coordinate-wise scalar samples. This reformulation eliminates the need for dense spatiotemporal sampling and instead learns each spatial location from its full temporal evolution in a structured manner. We demonstrate that this representation is compatible with a range of existing INR architectures and consistently improves reconstruction quality, while significantly reducing training cost. Furthermore, we show that this formulation can be combined with mixture-of-experts architectures, and that our MoE instantiation further improves reconstruction quality compared to both the base reformulation and existing MoE-based INR methods, providing a stronger capacity allocation under heterogeneous temporal dynamics.
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 - the dense backbone, the experts, and the router - differ enough in size and gradient statistics that they should not receive the same state. 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, 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 matches it with twenty times the state, and Adafactor, which shares the factored estimator but drops momentum, plateaus 40 points behind. The tiers buy memory at no cost to accuracy; the accuracy comes from keeping momentum, which a uniform optimizer shares too. Sweeping the baselines' learning rates narrows but does not close the gap: the best tuned AdamW reaches 118.5, tuned Adafactor 139.7. Where optimizer state lives, these results suggest, matters at least as much as how much of it there is.
Rollout generation is a major bottleneck in Reinforcement Learning (RL) for Mixture-of-Experts (MoE) Large Language Models, motivating low-precision rollout acceleration such as FP8. As an emerging low-precision format, NVFP4 combines fine-grained scaling for accuracy preservation with native W4A4 FP4 GEMMs for higher throughput than FP8. However, we find that directly applying NVFP4 to MoE RL rollout is impractical. NVFP4 rollout with BF16 training collapses after roughly 150 steps, accompanied by rapidly growing rollout-trainer log-probability gaps. Through training-inference error analysis and controlled ablations, we identify activation error, rather than weight error, as the dominant source of FP4 RL instability: weights can be synchronized and aligned by a shared quantization-dequantization path, whereas activations are recomputed online and error is amplified by the coarse E2M1 grid. Therefore, to stabilize NVFP4 RL for MoE, we propose QUantization-error Alignment across Dual Sides (QUADS). On the trainer side, we introduce Asymmetric Quantization-Aware Training fake-quantizing weights while keeping activations unquantized for better alignment. On the rollout side, Residual Activation Compensation corrects high-error activation channels while preserving native W4A4 GEMMs. In our MoE RL experiments on several benchmarks, QUADS achieves BF16-level accuracy, improves average pass@1 by 21.49 points over naive NVFP4 RL, and delivers ~16% higher rollout throughput than FP8.
Multimodal Knowledge Graph Completion (MKGC) requires inferring missing entities from structural, textual, and visual cues. Existing diffusion-based MKGC methods usually denoise directly on raw multimodal features. Such a design forces the denoiser to simultaneously perform relation-dependent cue selection, cross-modal semantic alignment, and structure-aware entity generation, which introduces noisy and semantically inconsistent conditions for diffusion and consequently leads to suboptimal completion performance. To address this limitation, we propose MGDT: MLLM-Guided Diffusion Transformer with Relation-Adaptive Mixture-of-Experts (MGDT), a novel MKGC framework built on an align-then-diffuse paradigm. MGDT first employs a Relation-Adaptive Semantic Routing Mixture-of-Experts (RASR-MoE) module to select relation-relevant multimodal semantic transformation paths and suppress irrelevant modality interference. MGDT then uses a frozen Multimodal Large Language Model (MLLM) as a semantic anchor to align the routed multimodal representations into a unified latent space and reduce cross-modal semantic heterogeneity. Finally, a Knowledge Graph Diffusion Transformer (KGDT) performs graph-conditioned denoising generation in the aligned space to produce the missing entity representation. Experiments on three benchmark datasets show that MGDT consistently outperforms strong baselines.
Sparse Mixture-of-Experts (MoE) models have become an important approach for scaling Large Language Models (LLMs), but their inference efficiency depends strongly on expert activation patterns. Speculative decoding (SD) accelerates autoregressive generation by verifying multiple draft tokens in parallel, yet existing draft selection strategies primarily optimize acceptance likelihood. In large-scale MoE models, however, selecting draft tokens also determines the union of experts activated during verification. We observe that confidence-driven SD can introduce \textit{expert scattering}: high-probability draft tokens may route to disjoint experts, increasing expert-weight memory traffic and reducing the speedup from speculation. Motivated by this observation, we revisit draft-tree selection under the non-uniform memory-cost structure of MoE inference. We propose \textsc{EcoSpec}, a cost-aware speculative decoding framework that incorporates predicted marginal expert activation cost into draft selection. With a lightweight expert predictor and a dynamic expert buffer, \textsc{EcoSpec} favors draft paths that preserve high acceptance likelihood while reusing experts already covered by the current verification set, without modifying the target-model verification rule. We evaluate \textsc{EcoSpec} on three large-scale MoE models, including DeepSeek-V3.1 (671B), Qwen3-235B-A22B, and GPT-OSS-120B, across reasoning, coding, question-answering, and dialogue benchmarks. \textsc{EcoSpec} consistently reduces active expert footprints and improves end-to-end decoding speed, achieving up to $1.62\times$ speedup. These results show that accounting for expert activation cost is important for efficient speculative decoding in large-scale MoE models.