ICML 2026 Papers

Large Language Models (LLMs) often assign disproportionate attention to the first token, a phenomenon known as the attention sink. Several recent approaches aim to address this issue, including Sink Attention in GPT-OSS and Gated Attention in Qwen3-Next. However, a comprehensive analysis of the relationship among these attention mechanisms is lacking. In this work, we provide both theoretical and empirical evidence demonstrating that the sink in Vanilla Attention and Sink Attention naturally construct a Mixture-of-Experts (MoE) mechanism within attention layers. This insight explains the head collapse phenomenon observed in prior work, where only a fixed subset of attention heads contributes to generation. To mitigate head collapse, we propose a sink-aware training algorithm with an auxiliary load balancing loss designed for attention layers. Extensive experiments show that our method achieves effective head load balancing and improves model performance across Vanilla Attention, Sink Attention, and Gated Attention. We hope this study offers a new perspective on attention mechanisms and encourages further exploration of the inherent MoE structure within attention layers.

Mixture-of-Experts (MoE) scales model capacity efficiently by selectively routing inputs to a specialized subset of experts. However, input-expert specialization, the core motivation of MoE, critically depends on whether the router is actually aware of input structure. In practice, MoE routing is typically implemented as a shallow linear projection with limited awareness of input representation, which often leads to unstable routing. We propose STAR, a Structure Aware Routing that rethinks MoE routing as a subspace learning problem by augmenting standard learnable routing with an evolving principal subspace that tracks dominant input structure via Generalized Hebbian Algorithm (GHA). By aligning routing decisions directly with input structure, STAR enables stable expert specialization. We evaluate STAR on controlled synthetic setup and large-scale language and vision tasks, where it consistently improves routing quality and downstream performance over strong MoE baselines. Moreover, optional test-time subspace updates further enhance routing robustness and generalization under input distribution shifts. Code is available at \url{https://github.com/psmiz/STAR}.

Synthesizing realistic time series with generative models has wide-ranging applications in real-world scenarios. Despite recent progress, most existing methods are trained under the assumption of abundant training data, which substantially limits their effectiveness in data-scarce settings. In this paper, we propose TimeMoDE, a novel framework that integrates Diffusion Transformers with Mixture-of-Experts to exploit both domain adaptability and diffusion-stage awareness for time series generation under data scarcity. It is pre-trained on a large-scale collection of multi-domain datasets to extract domain-agnostic temporal representations and domain-specific information benefiting generalization during fine-tuning. We propose Domain Prompts to condition expert assignment for indistinguishable noised tokens, mitigating the limitations of capturing inter-dataset relationships. Moreover, we incorporate diffusion timestep signals to equip the experts with awareness of time series degradation variations, facilitating adaptive calibrate to stage-dependent denoising requirements. Extensive experiments demonstrate that TimeMoDE outperforms existing methods under diverse low-data settings. It establishes an innovative paradigm for advanced time series few-shot generation.

Mixture-of-Experts (MoEs) have become a central component of many state-of-the-art open-source and proprietary large language models. Despite their widespread adoption, it remains unclear how close existing MoE architectures are to optimal for inference cost, as measured by accuracy per floating-point operation and per parameter. In this work, we revisit MoE design from a hardware-software co-design perspective, grounded in empirical and theoretical considerations. We characterize key performance bottlenecks across diverse deployment regimes, spanning offline high-throughput execution and online, latency-critical inference. Guided by these insights, we introduce **LatentMoE**, a new model architecture resulting from systematic design exploration and optimized for maximal accuracy per unit of compute. Empirical design space exploration at scales of up to 95B parameters and over a 1T-token training horizon, together with supporting theoretical analysis, shows that LatentMoE consistently outperforms standard MoE architectures in terms of accuracy per FLOP and per parameter. Given its strong performance, the LatentMoE architecture has been adopted by the flagship Nemotron-3 Super and Ultra models and scaled to substantially larger regimes, including longer token horizons and larger model sizes, as reported in (NVIDIA et al., 2025, arXiv:2512.20856).

Graph Mixture-of-Experts (Graph-MoE) offers a way to scale GNNs via adaptive capacity allocation, with the goal of allowing different experts to capture diverse graph patterns. Its effectiveness heavily depends on the coordination between routing decisions and expert specialization. However, through extensive empirical study, we identify two critical phenomena. First, discrimination loss occurs on both the expert and routing sides, where GNN experts become highly homogenized and the router collapses to a small subset of experts, failing to reflect diverse graph semantics. Second, routing uncertainty is prevalent, as existing routers produce uncertain expert assignments for most nodes, and such uncertainty exhibits a strong negative correlation with model performance. To address these issues, we propose C$^2$GMoE, a novel Graph-MoE framework featuring Contrastive routing and Confidence-aware fusion. We introduce a group-wise contrastive routing strategy that provides explicit guidance for routing optimization by aligning node-level routing decisions with semantic clusters while satisfying load-balancing constraints. Moreover, through a theoretical analysis of generalization error, we develop a confidence-aware fusion mechanism that adaptively reweights expert predictions according to their confidence. Extensive experiments across multiple benchmarks demonstrate the effectiveness of our proposed C$^2$GMoE.

Diffusion language models (DLMs) offer substantial speed advantages through parallel decoding, but the lack of token dependencies limits generation quality compared to autoregressive (AR) models. Recent progress attempts to bridge the gap via importance sampling, with DLM being the proposal and AR being the target. However, due to the huge gap between their distributions, the sampling requires a large number of particles and is thus expensive to compute. In this paper, we introduce PoE-Bridge, a novel decoding framework that drastically improves generation speed and accuracy by introducing an intermediate distribution to bridge the gap. The distribution is constructed as a Product-of-Experts (PoE) of the DLM proposal and the AR target. With the intermediate distribution, we first use the DLM to draft multiple continuations in parallel, then apply rejection sampling to verify the drafted tokens and move the resulting candidates toward the PoE. We then use importance sampling to further correct the PoE-aligned candidates toward the AR target. We further propose several improved techniques, including mixed-temperature sampling for enhanced diversity and elastic rejection windows for reducing wasted verification. Empirically, PoE-Bridge achieves significantly improved accuracy with $5\times$ speedup over the standard DLM decoding approach, and recovers at least 95% of the target AR model's performance, efficiently advancing most of the quality gap on challenging mathematical reasoning and coding tasks. Our code is available at [https://github.com/juntongshi48/poe-bridge](https://github.com/juntongshi48/poe-bridge).

Mixture-of-Experts (MoE) models enable sparse expert activation, meaning that only a subset of the model’s parameters is used during each inference. However, to translate this sparsity into practical performance, an expert caching mechanism is required. Previous works have proposed hardware-centric caching policies, but how these various caching policies interact with each other and different hardware specification remains poorly understood. To address this gap, we develop \textbf{SpecMD}, a standardized framework for benchmarking ad-hoc cache policies on various hardware configurations. Using SpecMD, we perform an exhaustive benchmarking of several MoE caching strategies, reproducing and extending prior approaches in controlled settings with realistic constraints. Our experiments reveal that MoE expert access is not consistent with temporal locality assumptions (e.g LRU, LFU). Motivated by this observation, we propose \textbf{Least-Stale}, a novel eviction policy that exploits MoE's predictable expert access patterns to reduce collision misses by up to $85\times$ over LRU. With such gains, we achieve over $88\%$ hit rates with up to $34.7\%$ Time-to-first-token (TTFT) reduction on OLMoE at only $5\%$ or $0.6GB$ of VRAM cache capacity.

In DRL, an agent is trained from a stream of experience. In a continual learning setting, such agents can suffer from \emph{plasticity loss}: their ability to learn new skills from new experiences diminishes over training. Recently, Mixture-of-Experts (MoE) networks have been reported to enable scaling laws and facilitate the learning of diverse skills. However, in continual reinforcement learning settings, their performance can degenerate as learning proceeds, indicating a loss of plasticity. To address this, building on Neural Tangent Kernel (NTK) theory, we formalize the plasticity loss in MoE policies as a loss of \emph{spectral plasticity}. We then derive a tractable proxy for spectral plasticity, one expressible in terms of individual expert feature matrices. Leveraging this proxy, we introduce \emph{SPHERE}, a practical Parseval penalty tailored for MoE-based policies that alleviates the loss of spectral plasticity. On MetaWorld and HumanoidBench, SPHERE improves average success under continual RL by 133\% and 50\% over an unregularized MoE baseline, while maintaining higher spectral plasticity throughout training.

Sparse Mixture-of-Experts architectures are essential for scaling model capacity efficiently, yet the standard Top-$k$ routing imposes a rigid sparsity pattern that ignores the intrinsic variance in token difficulty and layer-specific computational needs. Top-$p$ routing is more adaptive because it selects experts until their cumulative routing probability reaches a threshold, allowing confident tokens to use fewer experts and ambiguous tokens to recruit more. However, we demonstrate that existing naive Top-$p$ implementations with fixed global probability thresholds provide only marginal gains over Top-$k$, suffer from hyperparameter sensitivity, and result in uncontrolled computational costs. In this paper, we propose **DTop-$p$**, a sparsity-controllable dynamic routing mechanism that learns the Top-$p$ probability threshold with a Proportional-Integral controller and uses dynamic routing normalization to support layer-wise expert selection under a global sparsity constraint. Extensive experiments on Large Language Models and Diffusion Transformers demonstrate that **DTop-$p$** consistently outperforms both Top-$k$ and fixed Top-$p$ baselines while matching the average FLOPs of Top-$k$ MoE. Our analysis confirms that **DTop-$p$** exhibits strong scaling properties across expert granularity, total expert capacity, model size, and dataset size, offering a robust and efficient MoE framework for foundation model pre-training.

Current fMRI decoders face a performance-fidelity trade-off where efficient ID encoders outperform geometrically faithful surface-based models. We argue this is partly driven by inefficient surface tokenization and the failure to use anatomy as a predictive signal. We present **NeurIPS**, a framework that improves surface-based decoding by reframing anatomical variation from a nuisance to a powerful inductive prior. NeurIPS unites two innovations: a **Selective ROI Spherical Tokenizer (SRST)** for efficient geometric encoding, and a **Structure-Guided Mixture of Experts (SG-MoE)** that explicitly models individual anatomy using cortical features. On the Natural Scenes Dataset, NeurIPS establishes a new state-of-the-art for surface decoders and achieves performance comparable to strong 1D baselines. This is achieved with unprecedented efficiency, as the model converges dramatically faster (**10 vs. 600 epochs**). This efficiency enables rapid adaptation to new subjects using only **20\%** of data and ensures robust scalability as the training cohort is expanded. Ablations provide causal evidence that these gains are driven by the model's use of cortical features, not by memorizing subject IDs. By leveraging anatomical priors, NeurIPS provides a principled and scalable path toward robust, generalizable brain decoding.

Tabular anomaly detection (TAD) aims to identify samples that deviate from the majority in tabular data and is critical in many real-world applications. However, existing methods follow a ``one model for one dataset (OFO)'' paradigm, which relies on dataset-specific training and thus incurs high computational cost and yields limited generalization to unseen domains. To address these limitations, we propose OFA-TAD, a generalist one-for-all (OFA) TAD framework that only requires one-time training on multiple source datasets and can generalize to unseen datasets from diverse domains on-the-fly. To realize one-for-all tabular anomaly detection, OFA-TAD extracts neighbor-distance patterns as transferable cues, and introduces multi-view neighbor-distance representations from multiple transformation-induced metric spaces to mitigate the transformation sensitivity of distance profiles. To adaptively combine multi-view distance evidence, a Mixture-of-Experts (MoE) scoring network is employed for view-specific anomaly scoring and entropy-regularized gated fusion, with a multi-strategy anomaly synthesis mechanism to support training under the one-class constraint. Extensive experiments on 34 datasets from 14 domains demonstrate that OFA-TAD achieves superior anomaly detection performance and strong cross-domain generalizability under the strict OFA setting. The source code is available at https://github.com/Shiy-Li/OFA-TAD.

By introducing routers to selectively activate experts in Transformer layers, the mixture-of-experts (MoE) architecture significantly reduces computational costs in large language models (LLMs) while maintaining competitive performance, especially for models with massive parameters. However, prior work has largely focused on utility and efficiency, leaving the safety risks associated with this sparse architecture underexplored. In this work, we show that the safety of MoE LLMs is as sparse as their architecture by discovering $\text{\emph{unsafe routes}}$: routing configurations that, once activated, convert safe outputs into harmful ones. Specifically, we first introduce the $\underline{\text{Ro}}$uter $\underline{\text{Sa}}$fety $\underline{\text{i}}$mportance $\underline{\text{s}}$core ($\textbf{RoSais}$) to quantify the safety criticality of each layer's router. Manipulation of only the high-RoSais router(s) can flip the default route into an unsafe one. We further propose a $\underline{\text{F}}$ine-grained token-layer-wise $\underline{\text{S}}$tochastic $\underline{\text{O}}$ptimization framework to discover more concrete $\underline{\text{U}}$nsafe $\underline{\text{R}}$outes ($\textbf{F-SOUR}$), which explicitly considers the sequentiality and dynamics of input tokens. Across four representative MoE LLM families, F-SOUR achieves an average ASR of 0.90 and 0.98 on JailbreakBench and AdvBench, respectively. Finally, we outline defensive perspectives, including safety-aware route disabling and router training, as promising directions to safeguard MoE LLMs. We hope our work can inform future red-teaming and safeguarding of MoE LLMs. Our code is available at https://github.com/TrustAIRLab/UnsafeMoE.

As the computational demands for pre-training Large Language Models (LLMs) continue to surge, the need for efficient training paradigms becomes critical. Despite the vast resources already invested in existing pre-trained checkpoints, these assets often remain under-leveraged due to architectural limitations. We introduce an "orthogonal growth" strategy designed to "recycle" these checkpoints by strategically expanding their parameters prior to continued training. Our method focuses on optimizing converged Mixture-of-Experts (MoE) models through two dimensions: interpositional layer copying for increased depth and noisy expert duplication for expanded width. Through extensive scaling laws analysis, we demonstrate a strong positive correlation between the "sunk cost" (prior investment) and the final model accuracy. Empirical results on models up to 70B parameters and 1T tokens show that our recycling approach yields a 10.6\% accuracy improvement compared to training from scratch under identical extra compute budgets. This work provides a cost-effective blueprint for sustainable large-scale LLM development.

Mixture-of-Experts (MoE) layers have emerged as an important tool in scaling up modern neural networks by decoupling total trainable parameters from activated parameters in the forward pass for each token. However, sparse MoEs add complexity to training due to (i) new trainable parameters (router weights) that, like all other parameter groups, require hyperparameter (HP) tuning; (ii) new architecture scale dimensions (number of and size of experts) that must be chosen and potentially taken large. To make HP selection cheap and reliable, we propose a new parameterization for transformer models with MoE layers when scaling model width, depth, number of experts, *and* expert (hidden) size. Our parameterization is justified by a novel dynamical mean-field theory (DMFT) analysis. When varying different model dimensions trained at a fixed token budget, we find empirically that our parameterization enables reliable HP transfer across models from 51M to 2B total parameters. We further take HPs identified from sweeping small models on a short token horizon to train larger models on longer horizons and report performant model behaviors.

Test-time scaling improves LLM performance by generating multiple candidate solutions, yet token-level sampling requires temperature tuning that trades off diversity against stability. Fine-grained MoE, featuring hundreds of well-trained experts per layer and multi-expert activation per token, offers an unexplored alternative through its rich routing space. We empirically characterize fine-grained MoE routing and uncover an informative pattern: router scores exhibit a certain head of high-confidence experts followed by an uncertain tail of low-confidence candidates. While single-run greedy accuracy remains stable when fewer experts are activated, multi-sample pass@n degrades significantly—suggesting that the certain head governs core reasoning capability while the uncertain tail correlates with reasoning diversity. Motivated by these findings, we propose Expert-Sample, a training-free method that preserves high-confidence selections while injecting controlled stochasticity into the uncertain tail, enabling diverse generation without destabilizing outputs. Evaluated on multiple fine-grained MoE models across math, knowledge reasoning, and code tasks, Expert-Sample consistently improves pass@n and verification-based accuracy. On Qwen3-30B-A3B-Instruct evaluated on GPQA-Diamond with 32 parallel samples, pass@32 rises from 85.4% to 91.9%, and accuracy improves from 59.1% to 62.6% with Best-of-N verification.

Dynamic graph learning, which focuses on modeling the merging, vanishing, and reconnection of nodes and edges, is crucial for real-world applications. In dynamic graphs, node neighborhoods often exhibit diverse and time-evolving topologies, including hierarchical, grid-like, and cyclic patterns. Existing methods typically embed graphs into a single curvature space, which limits the quality of node representations when the embedding geometry is not aligned well with the local graph topology. In this paper, we propose **DyGMoCE**, a **Dy**namic **G**raph Transformer with a **M**ixture **o**f **C**urvature-aware **E**xperts, which efficiently embeds each node at every timestamp into an adaptive curvature space. Specifically, DyGMoCE incorporates a mixture-of-experts framework to both the attention and feed-forward modules, where each expert operates on a Riemannian manifold with a distinct curvature. Then, motivated by the geometric continuity across the experts, we introduce a routing mechanism with a ranking constraint. To improve efficiency, we design a fast Riemannian attention module for DyGMoCE, achieving an average speedup of 27.5% and memory reduction of 52.6%. Notably, the fast Riemannian attention module is broadly applicable to Transformer models with sequence inputs. Extensive experimental results show that DyGMoCE significantly outperforms other state-of-the-art methods.

Real-world graph datasets often arise from mixtures of populations, where graphs are generated by multiple distinct underlying distributions. In this work, we propose a unified framework that explicitly models graph data as a mixture of probabilistic graph generative models represented by graphons. To characterize and estimate these graphons, we leverage graph moments (motif densities) to cluster graphs generated from the same underlying model. We establish a novel theoretical guarantee, deriving a tighter bound showing that graphs sampled from structurally similar graphons exhibit similar motif densities with high probability. This result enables principled estimation of graphon mixture components. We show how incorporating estimated graphon mixture components enhances two widely used downstream paradigms: graph data augmentation via mixup and graph contrastive learning. By conditioning these methods on the underlying generative models, we develop graphon-mixture-aware mixup (GMAM) and model-aware graph contrastive learning (MGCL). Extensive experiments on both simulated and real-world datasets demonstrate strong empirical performance. In supervised learning, GMAM outperforms existing augmentation strategies, achieving new state-of-the-art accuracy on 6 out of 7 datasets. In unsupervised learning, MGCL performs competitively across seven benchmark datasets and achieves the lowest average rank overall.

Quantization has significantly improved the compute and memory efficiency of Large Language Model (LLM) training. However, existing approaches still rely on accumulating their updates in high-precision: concretely, gradient updates must be applied to a high-precision weight buffer, known as $\textit{master weights}$. This buffer introduces substantial memory overhead, particularly for Sparse Mixture of Experts (SMoE) models, where model parameters and optimizer states dominate memory usage. To address this, we introduce the Error-Compensating Optimizer (ECO), which eliminates master weights by applying updates directly to quantized parameters. ECO quantizes weights after each step and carefully injects the resulting quantization error into the optimizer momentum, forming an error-feedback loop with no additional memory. We prove that, under standard assumptions and a decaying learning rate, ECO converges to a constant-radius neighborhood of the optimum, while naive master-weight removal can incur an error that is inversely proportional to the learning rate. We show empirical results for pretraining small Transformers (30--800M), a Gemma-3 1B model, and a 2.1B parameter Sparse MoE model with FP8 quantization, and fine-tuning DeepSeek-MoE-16B in INT4 precision. Throughout, ECO matches baselines with master weights up to near-lossless accuracy, significantly shifting the static memory vs validation loss Pareto frontier.

In this paper, we introduce a model of evolution and learning in robots that co-optimizes a distribution of latent design vectors (genotypes) and a mixture of control experts (neural modules), which are gated by the latent coordinates of each decoded design (phenotype). This provides a scalable alternative to co-design algorithms that either train an individual policy for every robot, which is inefficient, or a monolithic universal controller for all robots, which results in overly conservative structures and behaviors. Our approach lies somewhere between these two extremes, preserving ancestral knowledge in a unified yet modular framework in which different body plans activate and deactivate different combinations of learned sensorimotor circuits for goal-directed behavior. This allows one part of the controller to be overhauled to better suit new species of designs as they emerge without disrupting the hard-earned knowledge contained within other expert modules. It also allows pretrained expert policies to be directly plugged into the mixture, which can steer evolution into otherwise unexplored areas of latent space containing desired morphological traits. We refer to this process as "evo by demo" and explore how it may be used to guide freeform evolution toward canonical structures defined by the pretrained model. Videos and code can be found at: https://eco-moe.github.io.