Latent Prototype Routing: Achieving Near-Perfect Load Balancing in Mixture-of-Experts

Mixture-of-Experts (MoE) architectures have emerged as a key strategy for scaling large language models (LLMs) efficiently. However, current MoE systems suffer from severe load imbalance, where only a small subset of experts is consistently activated during training and inference, leading to significant underutilization of model capacity and computational resources. In this work, we revisit expert routing through a clustering perspective and propose Latent Prototype Routing (LPR), a novel routing framework that generalizes existing approaches while promoting balanced expert utilization without compromising downstream performance. Extensive experiments across multiple open-source MoE models -- including DeepSeek-V3, Qwen3-MoE, and Mixtral -- demonstrate that LPR reduces the Gini coefficient of expert load from 0.70 to 0.035 on average, improves the min-max expert load ratio from 1e-6 to 0.70, achieving near-perfect load balancing.

Who Does What in Deep Learning? Multidimensional Game-Theoretic Attribution of Function of Neural Units

Neural networks now generate text, images, and speech with billions of parameters, producing a need to know how each neural unit contributes to these high-dimensional outputs. Existing explainable-AI methods, such as SHAP, attribute importance to inputs, but cannot quantify the contributions of neural units across thousands of output pixels, tokens, or logits. Here we close that gap with Multiperturbation Shapley-value Analysis (MSA), a model-agnostic game-theoretic framework. By systematically lesioning combinations of units, MSA yields Shapley Modes, unit-wise contribution maps that share the exact dimensionality of the model's output. We apply MSA across scales, from multi-layer perceptrons to the 56-billion-parameter Mixtral-8x7B and Generative Adversarial Networks (GAN). The approach demonstrates how regularisation concentrates computation in a few hubs, exposes language-specific experts inside the LLM, and reveals an inverted pixel-generation hierarchy in GANs. Together, these results showcase MSA as a powerful approach for interpreting, editing, and compressing deep neural networks.

MoE-GPS: Guidlines for Prediction Strategy for Dynamic Expert Duplication in MoE Load Balancing

In multi-GPU Mixture-of-Experts (MoE) network, experts are distributed across different GPUs, which creates load imbalance as each expert processes different number of tokens. Recent works improve MoE inference load balance by dynamically duplicating popular experts to more GPUs to process excessive tokens, which requires predicting the distribution before routing. In this paper, we discuss the tradeoff of prediction strategies, accuracies, overhead, and end-to-end system performance. We propose MoE-GPS, a framework that guides the selection of the optimal predictor design under various system configurations, by quantifying the performance impact to system-level model runtime. Specifically, we advocate for Distribution-Only Prediction, a prediction strategy that only predicts overall token distribution which significantly reduces overhead compared to the traditional Token-to-Expert Prediction. On Mixtral 8x7B MMLU dataset, MoE-GPS suggests Distribution-Only Prediction which improves end-to-end inference performance by more than 23% compared with Token-to-Expert Prediction.

A Dataset for Addressing Patient's Information Needs related to Clinical Course of Hospitalization

Patients have distinct information needs about their hospitalization that can be addressed using clinical evidence from electronic health records (EHRs). While artificial intelligence (AI) systems show promise in meeting these needs, robust datasets are needed to evaluate the factual accuracy and relevance of AI-generated responses. To our knowledge, no existing dataset captures patient information needs in the context of their EHRs. We introduce ArchEHR-QA, an expert-annotated dataset based on real-world patient cases from intensive care unit and emergency department settings. The cases comprise questions posed by patients to public health forums, clinician-interpreted counterparts, relevant clinical note excerpts with sentence-level relevance annotations, and clinician-authored answers. To establish benchmarks for grounded EHR question answering (QA), we evaluated three open-weight large language models (LLMs)--Llama 4, Llama 3, and Mixtral--across three prompting strategies: generating (1) answers with citations to clinical note sentences, (2) answers before citations, and (3) answers from filtered citations. We assessed performance on two dimensions: Factuality (overlap between cited note sentences and ground truth) and Relevance (textual and semantic similarity between system and reference answers). The final dataset contains 134 patient cases. The answer-first prompting approach consistently performed best, with Llama 4 achieving the highest scores. Manual error analysis supported these findings and revealed common issues such as omitted key clinical evidence and contradictory or hallucinated content. Overall, ArchEHR-QA provides a strong benchmark for developing and evaluating patient-centered EHR QA systems, underscoring the need for further progress toward generating factual and relevant responses in clinical contexts.

Decoding Knowledge Attribution in Mixture-of-Experts: A Framework of Basic-Refinement Collaboration and Efficiency Analysis

The interpretability of Mixture-of-Experts (MoE) models, especially those with heterogeneous designs, remains underexplored. Existing attribution methods for dense models fail to capture dynamic routing-expert interactions in sparse MoE architectures. To address this issue, we propose a cross-level attribution algorithm to analyze sparse MoE architectures (Qwen 1.5-MoE, OLMoE, Mixtral-8x7B) against dense models (Qwen 1.5-7B, Llama-7B, Mixtral-7B). Results show MoE models achieve 37% higher per-layer efficiency via a "mid-activation, late-amplification" pattern: early layers screen experts, while late layers refine knowledge collaboratively. Ablation studies reveal a "basic-refinement" framework--shared experts handle general tasks (entity recognition), while routed experts specialize in domain-specific processing (geographic attributes). Semantic-driven routing is evidenced by strong correlations between attention heads and experts (r=0.68), enabling task-aware coordination. Notably, architectural depth dictates robustness: deep Qwen 1.5-MoE mitigates expert failures (e.g., 43% MRR drop in geographic tasks when blocking top-10 experts) through shared expert redundancy, whereas shallow OLMoE suffers severe degradation (76% drop). Task sensitivity further guides design: core-sensitive tasks (geography) require concentrated expertise, while distributed-tolerant tasks (object attributes) leverage broader participation. These insights advance MoE interpretability, offering principles to balance efficiency, specialization, and robustness.

From prosthetic memory to prosthetic denial: Auditing whether large language models are prone to mass atrocity denialism

The proliferation of large language models (LLMs) can influence how historical narratives are disseminated and perceived. This study explores the implications of LLMs' responses on the representation of mass atrocity memory, examining whether generative AI systems contribute to prosthetic memory, i.e., mediated experiences of historical events, or to what we term "prosthetic denial," the AI-mediated erasure or distortion of atrocity memories. We argue that LLMs function as interfaces that can elicit prosthetic memories and, therefore, act as experiential sites for memory transmission, but also introduce risks of denialism, particularly when their outputs align with contested or revisionist narratives. To empirically assess these risks, we conducted a comparative audit of five LLMs (Claude, GPT, Llama, Mixtral, and Gemini) across four historical case studies: the Holodomor, the Holocaust, the Cambodian Genocide, and the genocide against the Tutsis in Rwanda. Each model was prompted with questions addressing common denialist claims in English and an alternative language relevant to each case (Ukrainian, German, Khmer, and French). Our findings reveal that while LLMs generally produce accurate responses for widely documented events like the Holocaust, significant inconsistencies and susceptibility to denialist framings are observed for more underrepresented cases like the Cambodian Genocide. The disparities highlight the influence of training data availability and the probabilistic nature of LLM responses on memory integrity. We conclude that while LLMs extend the concept of prosthetic memory, their unmoderated use risks reinforcing historical denialism, raising ethical concerns for (digital) memory preservation, and potentially challenging the advantageous role of technology associated with the original values of prosthetic memory.

FloE: On-the-Fly MoE Inference on Memory-constrained GPU

With the widespread adoption of Mixture-of-Experts (MoE) models, there is a growing demand for efficient inference on memory-constrained devices. While offloading expert parameters to CPU memory and loading activated experts on demand has emerged as a potential solution, the large size of activated experts overburdens the limited PCIe bandwidth, hindering the effectiveness in latency-sensitive scenarios. To mitigate this, we propose FloE, an on-the-fly MoE inference system on memory-constrained GPUs. FloE is built on the insight that there exists substantial untapped redundancy within sparsely activated experts. It employs various compression techniques on the expert's internal parameter matrices to reduce the data movement load, combined with low-cost sparse prediction, achieving perceptible inference acceleration in wall-clock time on resource-constrained devices. Empirically, FloE achieves a 9.3x compression of parameters per expert in Mixtral-8x7B; enables deployment on a GPU with only 11GB VRAM, reducing the memory footprint by up to 8.5x; and delivers a 48.7x inference speedup compared to DeepSpeed-MII on a single GeForce RTX 3090 - all with only a 4.4$\%$ - 7.6$\%$ average performance degradation.

QoS-Efficient Serving of Multiple Mixture-of-Expert LLMs Using Partial Runtime Reconfiguration

The deployment of mixture-of-experts (MoE) large language models (LLMs) presents significant challenges due to their high memory demands. These challenges become even more pronounced in multi-tenant environments, where shared resources must accommodate multiple models, limiting the effectiveness of conventional virtualization techniques. This paper addresses the problem of efficiently serving multiple fine-tuned MoE-LLMs on a single-GPU. We propose a serving system that employs \textit{similarity-based expert consolidation} to reduce the overall memory footprint by sharing similar experts across models. To ensure output quality, we introduce \textit{runtime partial reconfiguration}, dynamically replacing non-expert layers when processing requests from different models. As a result, our approach achieves a competitive output quality while maintaining throughput comparable to serving a single model while incurring a negligible increase in time-to-first-token (TTFT). Experiments on a server with a single NVIDIA A100 GPU (80GB) using Mixtral-8x7B models demonstrate an 85\% average reduction in turnaround time compared to NVIDIA's multi-instance GPU (MIG). Furthermore, experiments on Google's Switch Transformer Base-8 model with up to four variants demonstrate the scalability and resilience of our approach in maintaining output quality compared to other model merging baselines, highlighting its effectiveness.

FloE: On-the-Fly MoE Inference

With the widespread adoption of Mixture-of-Experts (MoE) models, there is a growing demand for efficient inference on memory-constrained devices. While offloading expert parameters to CPU memory and loading activated experts on demand has emerged as a potential solution, the large size of activated experts overburdens the limited PCIe bandwidth, hindering the effectiveness in latency-sensitive scenarios. To mitigate this, we propose FloE, an on-the-fly MoE inference system on memory-constrained GPUs. FloE is built on the insight that there exists substantial untapped redundancy within sparsely activated experts. It employs various compression techniques on the expert's internal parameter matrices to reduce the data movement load, combined with low-cost sparse prediction, achieving perceptible inference acceleration in wall-clock time on resource-constrained devices. Empirically, FloE achieves a 9.3x compression of parameters per expert in Mixtral-8x7B; enables deployment on a GPU with only 11GB VRAM, reducing the memory footprint by up to 8.5x; and delivers a 48.7x inference speedup compared to DeepSpeed-MII on a single GeForce RTX 3090.

MoE Parallel Folding: Heterogeneous Parallelism Mappings for Efficient Large-Scale MoE Model Training with Megatron Core

Mixture of Experts (MoE) models enhance neural network scalability by dynamically selecting relevant experts per input token, enabling larger model sizes while maintaining manageable computation costs. However, efficient training of large-scale MoE models across thousands of GPUs presents significant challenges due to limitations in existing parallelism strategies. We introduce an end-to-end training framework for large-scale MoE models that utilizes five-dimensional hybrid parallelism: Tensor Parallelism, Expert Parallelism, Context Parallelism, Data Parallelism, and Pipeline Parallelism. Central to our approach is MoE Parallel Folding, a novel strategy that decouples the parallelization of attention and MoE layers in Transformer models, allowing each layer type to adopt optimal parallel configurations. Additionally, we develop a flexible token-level dispatcher that supports both token-dropping and token-dropless MoE training across all five dimensions of parallelism. This dispatcher accommodates dynamic tensor shapes and coordinates different parallelism schemes for Attention and MoE layers, facilitating complex parallelism implementations. Our experiments demonstrate significant improvements in training efficiency and scalability. We achieve up to 49.3% Model Flops Utilization (MFU) for the Mixtral 8x22B model and 39.0% MFU for the Qwen2-57B-A14B model on H100 GPUs, outperforming existing methods. The framework scales efficiently up to 1,024 GPUs and maintains high performance with sequence lengths up to 128K tokens, validating its effectiveness for large-scale MoE model training. The code is available in Megatron-Core.