Linear Mode Connectivity (LMC) is a notable phenomenon in the loss landscapes of neural networks, wherein independently trained models have been observed to be connected--up to permutation symmetries--by linear paths in parameter space along which the loss remains consistently low. This observation challenges classical views of non-convex optimization and has implications for model ensembling, generalization, and our understanding of neural loss geometry. Inspired by recent studies on LMC in standard neural networks, we systematically investigate this phenomenon within Mixture-of-Experts (MoE) architectures--a class of models known for their scalability and computational efficiency, which combine traditional neural networks--referred to as experts--through a learnable gating mechanism. We begin by conducting a comprehensive analysis of both dense and sparse gating regimes, demonstrating that the symmetries inherent to MoE architectures are fully characterized by permutations acting on both the expert components and the gating function. Building on these foundational findings, we propose a matching algorithm that enables alignment between independently trained MoEs, thereby facilitating the discovery of LMC. Finally, we empirically validate the presence of LMC using our proposed algorithm across diverse MoE configurations--including dense, sparse, and shared-expert variants--under a wide range of model settings and datasets of varying scales and modalities. Our results confirm the existence of LMC in MoE architectures and offer fundamental insights into the functional landscape and optimization dynamics of deep learning models.

Inspired by recent advancements in LLM reasoning, the field of multimodal reasoning has seen remarkable progress, achieving significant performance gains on intricate tasks such as mathematical problem-solving. Despite this progress, current multimodal large reasoning models exhibit two key limitations. They tend to employ computationally expensive reasoning even for simple queries, leading to inefficiency. Furthermore, this focus on specialized reasoning often impairs their broader, more general understanding capabilities. In this paper, we propose Metis-HOME: a Hybrid Optimized Mixture-of-Experts framework designed to address this trade-off. Metis-HOME enables a ''Hybrid Thinking'' paradigm by structuring the original dense model into two distinct expert branches: a thinking branch tailored for complex, multi-step reasoning, and a non-thinking branch optimized for rapid, direct inference on tasks like general VQA and OCR. A lightweight, trainable router dynamically allocates queries to the most suitable expert. We instantiate Metis-HOME by adapting the Qwen2.5-VL-7B into an MoE architecture. Comprehensive evaluations reveal that our approach not only substantially enhances complex reasoning abilities but also improves the model's general capabilities, reversing the degradation trend observed in other reasoning-specialized models. Our work establishes a new paradigm for building powerful and versatile MLLMs, effectively resolving the prevalent reasoning-vs-generalization dilemma. Code and weights are available at https://github.com/MM-Thinking/Metis-HOME.

Deploying deep neural networks on resource-constrained devices faces two critical challenges: maintaining accuracy under aggressive quantization while ensuring predictable inference latency. We present a curiosity-driven quantized Mixture-of-Experts framework that addresses both through Bayesian epistemic uncertainty-based routing across heterogeneous experts (BitNet ternary, 1-16 bit BitLinear, post-training quantization). Evaluated on audio classification benchmarks (ESC-50, Quinn, UrbanSound8K), our 4-bit quantization maintains 99.9 percent of 16-bit accuracy (0.858 vs 0.859 F1) with 4x compression and 41 percent energy savings versus 8-bit. Crucially, curiosity-driven routing reduces MoE latency variance by 82 percent (p = 0.008, Levene's test) from 230 ms to 29 ms standard deviation, enabling stable inference for battery-constrained devices. Statistical analysis confirms 4-bit/8-bit achieve practical equivalence with full precision (p > 0.05), while MoE architectures introduce 11 percent latency overhead (p < 0.001) without accuracy gains. At scale, deployment emissions dominate training by 10000x for models serving more than 1,000 inferences, making inference efficiency critical. Our information-theoretic routing demonstrates that adaptive quantization yields accurate (0.858 F1, 1.2M params), energy-efficient (3.87 F1/mJ), and predictable edge models, with simple 4-bit quantized architectures outperforming complex MoE for most deployments.

Autonomous driving holds transformative potential but remains fundamentally constrained by the limited perception and isolated decision-making with standalone intelligence. While recent multi-agent approaches introduce cooperation, they often focus merely on perception-level tasks, overlooking the alignment with downstream planning and control, or fall short in leveraging the full capacity of the recent emerging end-to-end autonomous driving. In this paper, we present UniMM-V2X, a novel end-to-end multi-agent framework that enables hierarchical cooperation across perception, prediction, and planning. At the core of our framework is a multi-level fusion strategy that unifies perception and prediction cooperation, allowing agents to share queries and reason cooperatively for consistent and safe decision-making. To adapt to diverse downstream tasks and further enhance the quality of multi-level fusion, we incorporate a Mixture-of-Experts (MoE) architecture to dynamically enhance the BEV representations. We further extend MoE into the decoder to better capture diverse motion patterns. Extensive experiments on the DAIR-V2X dataset demonstrate our approach achieves state-of-the-art (SOTA) performance with a 39.7% improvement in perception accuracy, a 7.2% reduction in prediction error, and a 33.2% improvement in planning performance compared with UniV2X, showcasing the strength of our MoE-enhanced multi-level cooperative paradigm.

Graph Neural Networks (GNNs) have demonstrated impressive performance on task-specific benchmarks, yet their ability to generalize across diverse domains and tasks remains limited. Existing approaches often struggle with negative transfer, scalability issues, and high adaptation costs. To address these challenges, we propose GMoPE (Graph Mixture of Prompt-Experts), a novel framework that seamlessly integrates the Mixture-of-Experts (MoE) architecture with prompt-based learning for graphs. GMoPE leverages expert-specific prompt vectors and structure-aware MoE routing to enable each expert to specialize in distinct subdomains and dynamically contribute to predictions. To promote diversity and prevent expert collapse, we introduce a soft orthogonality constraint across prompt vectors, encouraging expert specialization and facilitating a more balanced expert utilization. Additionally, we adopt a prompt-only fine-tuning strategy that significantly reduces spatiotemporal complexity during transfer. We validate GMoPE through extensive experiments under various pretraining strategies and multiple downstream tasks. Results show that GMoPE consistently outperforms state-of-the-art baselines and achieves performance comparable to full parameter fine-tuning-while requiring only a fraction of the adaptation overhead. Our work provides a principled and scalable framework for advancing generalizable and efficient graph foundation models.

Mixture-of-experts (MoE) architectures have expanded from language modeling to automatic speech recognition (ASR). Traditional MoE methods, such as the Switch Transformer, route experts independently within each layer. Our analysis reveals that routers in most layers make expert choices that are not strongly correlated with the choices of the routers in other layers. To increase the cooperation between experts in different layers and encourage greater specialization, we use a shared router across different MoE layers. We call this model Omni-router Transformer. Extensive experiments on a large-scale pseudo-labeled dataset and evaluations across 10 diverse, out-of-domain ASR benchmarks demonstrate that the Omni-router Transformer is able to achieve lower training loss and consistently outperform dense and Switch Transformer models, reducing average word error rates by 11.2% and 8.2%, respectively, while providing structured expert usage and improved robustness to diverse data.

Pre-training has proven effective in addressing data scarcity and performance limitations in solving PDE problems with neural operators. However, challenges remain due to the heterogeneity of PDE datasets in equation types, which leads to high errors in mixed training. Additionally, dense pre-training models that scale parameters by increasing network width or depth incur significant inference costs. To tackle these challenges, we propose a novel Mixture-of-Experts Pre-training Operator Transformer (MoE-POT), a sparse-activated architecture that scales parameters efficiently while controlling inference costs. Specifically, our model adopts a layer-wise router-gating network to dynamically select 4 routed experts from 16 expert networks during inference, enabling the model to focus on equation-specific features. Meanwhile, we also integrate 2 shared experts, aiming to capture common properties of PDE and reduce redundancy among routed experts. The final output is computed as the weighted average of the results from all activated experts. We pre-train models with parameters from 30M to 0.5B on 6 public PDE datasets. Our model with 90M activated parameters achieves up to a 40% reduction in zero-shot error compared with existing models with 120M activated parameters. Additionally, we conduct interpretability analysis, showing that dataset types can be inferred from router-gating network decisions, which validates the rationality and effectiveness of the MoE architecture.

Mixture-of-Experts (MoE) has become a dominant architecture for scaling Large Language Models (LLMs) efficiently by decoupling total parameters from computational cost. However, this decoupling creates a critical challenge: predicting the model capacity of a given MoE configurations (e.g., expert activation ratio and granularity) remains an unresolved problem. To address this gap, we introduce Efficiency Leverage (EL), a metric quantifying the computational advantage of an MoE model over a dense equivalent. We conduct a large-scale empirical study, training over 300 models up to 28B parameters, to systematically investigate the relationship between MoE architectural configurations and EL. Our findings reveal that EL is primarily driven by the expert activation ratio and the total compute budget, both following predictable power laws, while expert granularity acts as a non-linear modulator with a clear optimal range. We integrate these discoveries into a unified scaling law that accurately predicts the EL of an MoE architecture based on its configuration. To validate our derived scaling laws, we designed and trained Ling-mini-beta, a pilot model for Ling-2.0 series with only 0.85B active parameters, alongside a 6.1B dense model for comparison. When trained on an identical 1T high-quality token dataset, Ling-mini-beta matched the performance of the 6.1B dense model while consuming over 7x fewer computational resources, thereby confirming the accuracy of our scaling laws. This work provides a principled and empirically-grounded foundation for the scaling of efficient MoE models.

Vision-Language-Action (VLA) models are experiencing rapid development and demonstrating promising capabilities in robotic manipulation tasks. However, scaling up VLA models presents several critical challenges: (1) Training new VLA models from scratch demands substantial computational resources and extensive datasets. Given the current scarcity of robot data, it becomes particularly valuable to fully leverage well-pretrained VLA model weights during the scaling process. (2) Real-time control requires carefully balancing model capacity with computational efficiency. To address these challenges, We propose AdaMoE, a Mixture-of-Experts (MoE) architecture that inherits pretrained weights from dense VLA models, and scales up the action expert by substituting the feedforward layers into sparsely activated MoE layers. AdaMoE employs a decoupling technique that decouples expert selection from expert weighting through an independent scale adapter working alongside the traditional router. This enables experts to be selected based on task relevance while contributing with independently controlled weights, allowing collaborative expert utilization rather than winner-takes-all dynamics. Our approach demonstrates that expertise need not monopolize. Instead, through collaborative expert utilization, we can achieve superior performance while maintaining computational efficiency. AdaMoE consistently outperforms the baseline model across key benchmarks, delivering performance gains of 1.8% on LIBERO and 9.3% on RoboTwin. Most importantly, a substantial 21.5% improvement in real-world experiments validates its practical effectiveness for robotic manipulation tasks.

Mixture-of-Experts (MoE) models have recently demonstrated exceptional performance across a diverse range of applications. The principle of sparse activation in MoE models facilitates an offloading strategy, wherein active experts are maintained in GPU HBM, while inactive experts are stored in CPU DRAM. The efficacy of this approach, however, is fundamentally constrained by the limited bandwidth of the CPU-GPU interconnect. To mitigate this bottleneck, existing approaches have employed prefetching to accelerate MoE inference. These methods attempt to predict and prefetch the required experts using specially trained modules. Nevertheless, such techniques are often encumbered by significant training overhead and have shown diminished effectiveness on recent MoE models with fine-grained expert segmentation. In this paper, we propose MoBiLE, a plug-and-play offloading-based MoE inference framework with \textit{mixture of big-little experts}. It reduces the number of experts for unimportant tokens to half for acceleration while maintaining full experts for important tokens to guarantee model quality. Further, a dedicated fallback and prefetching mechanism is designed for switching between little and big experts to improve memory efficiency. We evaluate MoBiLE on four typical modern MoE architectures and challenging generative tasks. Our results show that MoBiLE achieves a speedup of 1.60x to 1.72x compared to the baseline on a consumer GPU system, with negligible degradation in accuracy.