ICLR 2026 Papers

Compositional Zero-Shot Learning (CZSL) investigates compositional generalization capacity to recognize unknown state-object pairs based on learned primitive concepts. Existing CZSL methods typically derive primitives features through a simple composition-prototype mapping, which is suboptimal for a set of individuals that can be divided into distinct semantic subsets. Moreover, the one-to-all cross-modal primitives matching neglects compositional divergence within identical states or objects, limiting fine-grained image-composition alignment. In this study, we propose EVA, a Mixture-of-Experts Framework for Semantic Variant Alignment. Specifically, we introduce domain-expert adaption, leveraging multiple experts to achieve token-aware learning and model high-quality primitive representations. To enable accurate compositional generalization, we further present semantic variant alignment to select semantically relevant representation for image-primitives matching. Our method significantly outperforms other state-of-the-art CZSL methods on three popular benchmarks in both closed- and open-world settings, demonstrating the efficacy of the proposed insight.

Test-time adaptation (TTA) has been widely explored to prevent performance degradation when test data differ from the training distribution. However, fully leveraging the rich representations of large pretrained models with minimal parameter updates remains underexplored. In this paper, we propose a novel approach, Intrinsic Mixture of Spectral Experts (IMSE), that leverages the spectral experts inherently embedded in Vision Transformers. We decompose each linear layer via singular value decomposition (SVD) and adapt only the singular values, referring to each decomposed rank-1 component as a spectral expert while keeping the singular vectors fixed. We further identify a key limitation of entropy minimization in TTA: it often reduces feature variance, causing the model to rely on domain-specific cues rather than class-discriminative features. To address this, we propose a diversity maximization loss based on singular vector–input alignment, which maximizing diversity of response pattern. In the continual test-time adaptation (CTTA) scenario, beyond preserving pretrained knowledge, it is crucial to retain and reuse knowledge from previously observed domains. We introduce Domain-Aware Spectral Code Retrieval, which estimates input distributions to detect domain shifts, and retrieves adapted singular values for rapid adaptation. Extensive experiments show that our method achieves state-of-the-art performance on ImageNet-C/R/A under single-domain TTA. In CTTA, it improves accuracy by 3.4pp with 2,000$\times$ fewer trainable parameters.

Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) are two prominent post-training paradigms for refining the capabilities and aligning the behavior of Large Language Models (LLMs). Existing approaches that integrate SFT and RL often face the risk of disrupting established response patterns and inducing overfitting to expert data. To address this, we present a novel investigation into the unified view of SFT and RL through an off-policy versus on-policy lens. We propose CHORD, a framework for Controllable Harmonization of On- and Off-Policy Reinforcement Learning via Dynamic Weighting, which reframes SFT not as a separate stage but as a dynamically weighted auxiliary objective within the on-policy RL process. Based on an analysis of off-policy expert data's influence at both holistic and granular levels, we incorporate a dual-control mechanism in CHORD. Specifically, the framework first employs a global coefficient to holistically guide the transition from off-policy imitation to on-policy exploration, and then applies a token-wise weighting function that enables granular learning from the expert, which promotes on-policy exploration and mitigates disruption from off-policy data. We conduct extensive experiments on mathematical reasoning problems and practical tool-use tasks, providing empirical evidence that CHORD achieves a stable and efficient learning process. By effectively harmonizing off-policy expert data with on-policy exploration, CHORD demonstrates significant improvements over baselines. We will release the source code to inspire further research.

Mixture-of-Experts (MoE) models have demonstrated exceptional performance in large-scale language models. Existing routers typically rely on non-differentiable Top-$k$+Softmax, limiting their performance and scalability. We argue that two distinct decisions, which experts to activate and how to distribute expert contributions among them, are conflated in standard Top-$k$+Softmax. We introduce Dirichlet-Routed MoE (DirMoE), a novel end-to-end differentiable routing mechanism built on a Dirichlet variational autoencoder framework. This design fundamentally disentangles the core routing problems: expert selection, modeled by a Bernoulli component, and expert contribution among chosen experts, handled by a Dirichlet component. The entire forward pass remains fully differentiable through the use of Gumbel-Sigmoid relaxation for the expert selection and implicit reparameterization for the Dirichlet distribution. Our training objective, a variational ELBO, includes a direct sparsity penalty that precisely controls the number of active experts in expectation, alongside a schedule for key hyperparameters that guides the model from an exploratory to a definitive routing state. Moreover, our DirMoE router matches or exceeds other methods while improving expert specialization.

Mixture-of-Experts (MoE) architectures have become the key to scaling modern LLMs, yet little is understood about how their sparse routing dynamics respond to multilingual data. In this work, we analyze expert routing patterns using parallel multilingual datasets and present highly interpretable layer-wise phenomena. We find that MoE models route tokens in language-specific ways in the early and late decoder layers but exhibit significant cross-lingual routing alignment in middle layers, mirroring parameter-sharing trends observed in dense LLMs. In particular, we reveal a clear, strong correlation between a model's performance in a given language and how similarly its tokens are routed to English in these layers. Extending beyond correlation, we explore inference-time interventions that induce higher cross-lingual routing alignment. We introduce a method that steers the router by promoting middle-layer task experts frequently activated in English, and it successfully increases multilingual performance. These 1-2% gains are remarkably consistent across two evaluation tasks, three models, and 15+ languages, especially given that these simple interventions override routers of extensively trained, state-of-the-art LLMs. In comparison, interventions outside of the middle layers or targeting multilingual-specialized experts only yield performance degradation. Altogether, we present numerous findings that explain how MoEs process non-English text and demonstrate that generalization is limited by the model’s ability to leverage language-universal experts in all languages.

Empirical scaling laws have driven the evolution of large language models (LLMs), yet their coefficients shift whenever the model architecture or data pipeline changes. Mixture‑of‑Experts (MoE) models, now standard in state‑of‑the‑art systems, introduce a new sparsity dimension that current dense‑model frontiers overlook. We investigate how MoE sparsity influences two distinct capability regimes: memorization skills and reasoning skills. By training MoE families that vary total parameters, active parameters, and top-$k$ routing under fixed compute budgets, we disentangle pre-training loss from downstream accuracy. Our results reveal two principles. First, Active FLOPs: models with identical training loss but greater active compute achieve higher reasoning accuracy. Second, Total tokens per parameter (TPP): memorization tasks improve with more parameters, while reasoning tasks benefit from optimal TPP, indicating that reasoning is data-hungry. Neither reinforcement learning post-training (GRPO) nor increased test-time compute alters these trends. We therefore argue that optimal MoE sparsity must be determined jointly by active FLOPs and TPP, revising the classical picture of compute-optimal scaling. All code, data sources, and logs are released to facilitate reproducibility and future work.

Recent advances in dynamic scene reconstruction have significantly benefited from 3D Gaussian Splatting, yet existing methods show inconsistent performance across diverse scenes, indicating no single approach effectively handles all dynamic challenges. To overcome these limitations, we propose Mixture of Experts for Dynamic Gaussian Splatting (MoE-GS), a unified framework integrating multiple specialized experts via a novel Volume-aware Pixel Router. Unlike sparsity-oriented MoE architectures in large language models, MoE-GS is designed to improve dynamic novel view synthesis quality by combining heterogeneous deformation priors, rather than to reduce training or inference-time FLOPs. Our router adaptively blends expert outputs by projecting volumetric Gaussian-level weights into pixel space through differentiable weight splatting, ensuring spatially and temporally coherent results. Although MoE-GS improves rendering quality, the increased model capacity and reduced FPS are inherent to the MoE architecture. To mitigate this, we explore two complementary directions: (1) single-pass multi-expert rendering and gate-aware Gaussian pruning, which improve efficiency within the MoE framework, and (2) a distillation strategy that transfers MoE performance to individual experts, enabling lightweight deployment without architectural changes. To the best of our knowledge, MoE-GS is the first approach incorporating Mixture-of-Experts techniques into dynamic Gaussian splatting. Extensive experiments on the N3V and Technicolor datasets demonstrate that MoE-GS consistently outperforms state-of-the-art methods with improved efficiency. Video demonstrations are available at https://huggingface.co/spaces/moegs/MoE-GS.

Large language models (LLMs) with Mixture-of-Experts (MoE) architectures achieve strong performance under sparse activation. However, their expertise is often fragmented across experts and redundant across layers. Prior studies primarily diagnosed redundancy or parameter importance, revealing overlaps but lacking mechanisms to transform them into reusable knowledge. In contrast, human learning succeeds not by memorizing isolated facts but by reusing shared strategies across domains, which motivates the question: do MoE models similarly encode universal knowledge that can be systematically extracted and reused? We propose Universal kNowledge Integration from Task-specific Experts (UNITE), a framework that consolidates experts through Fisher-weighted fusion and then applies Tucker decomposition to disentangle shared low-rank input/output subspaces as universal knowledge from layer-specific variations. This universal component provides a compact basis for reconstructing target models with flexible depth, enabling lightweight yet competitive adaptation across tasks. To assess effectiveness, we evaluate data efficiency, convergence speed, and generalization across multiple MoE-based LLMs and diverse datasets. The results show that UNITE not only extracts universal knowledge, but also flexibly recomposes it into once-for-all target models that generalize across domains.

Distributed inference of large-scale Mixture-of-Experts (MoE) models faces a critical challenge: expert load imbalance. Numerous system-level approaches have been proposed for load balancing, but they either fail to achieve a satisfactory level of balance or introduce new bottlenecks due to the overhead of the load balancing mechanism itself. To this end, we propose Libra, a system that achieves near-optimal load balancing with minimal overhead. Libra adopts sophisticated mechanisms that accurately predict future expert activations and, based on these predictions, systematically perform load balancing. At the same time, it effectively hides the associated overhead by reconstructing the execution flow so that these costs are overlapped with MoE computation. Evaluations with two large-scale state-of-the-art MoE models on 8 H200 GPUs demonstrate that Libra improves throughput by up to 19.2\%.

Recent large language models (LLMs) have increasingly adopted the Mixture-of-Experts (MoE) architecture for efficiency. MoE-based LLMs heavily depend on a superficial safety mechanism in which harmful inputs are routed safety-critical experts. However, our analysis reveals that routing decisions for harmful inputs drift significantly after fine-tuning, exposing a critical vulnerability to harmful fine-tuning (HFT) attacks. Existing defenses, primarily designed for monolithic LLMs, are less effective for MoE LLMs as they fail to prevent drift in harmful input routing. To address this limitation, we propose SafeMoE, a safe fine-tuning method tailored to MoE LLMs. SafeMoE directly mitigates routing drift by penalizing the gap between the routing weights of a fine-tuned model and those of the initial safety-aligned model, thereby preserving the safety-aligned routing of harmful inputs to safety-critical experts. Experiments on open-source MoE LLMs ranging from 7B to 141B parameters demonstrate that SafeMoE effectively mitigates HFT attacks, reducing the harmfulness score of OLMoE from 62.0 to 5.0, for example, while maintaining task utility within 1% degradation and incurring only 2\% overhead. It significantly outperforms state-of-the-art defense methods for safeguarding LLM fine-tuning and remains effective in recent large-scale MoE LLMs such as gpt-oss and Llama 4. Our implementation is available at https://anonymous.4open.science/r/SafeMoE.

Diffusion-based policies have recently shown strong results in robot manipulation, but their extension to multi-task scenarios is hindered by the high cost of scaling model size and demonstrations. We introduce Skill Mixture-of-Experts Policy (SMP), a diffusion-based mixture-of-experts policy that learns a compact orthogonal skill basis and uses sticky routing to compose actions from a small, task-relevant subset of experts at each step. A variational training objective supports this design, and adaptive expert activation at inference yields fast sampling without oversized backbones. We validate SMP in simulation and on a real dual-arm platform with multi-task learning and transfer learning tasks, where SMP achieves higher success rates and markedly lower inference cost than large diffusion baselines. These results indicate a practical path toward scalable, transferable multi-task manipulation: learn reusable skills once, activate only what is needed, and adapt quickly when tasks change.

Video Class-Incremental Learning (VCIL) seeks to develop models that continuously learn new action categories over time without forgetting previously acquired knowledge. Unlike traditional Class-Incremental Learning (CIL), VCIL introduces the added complexity of spatiotemporal structures, making it particularly challenging to mitigate catastrophic forgetting while effectively capturing both frame-shared semantics and temporal dynamics. Existing approaches either rely on exemplar rehearsal, raising concerns over memory and privacy, or adapt static image-based methods that neglect temporal modeling. To address these limitations, we propose Spatiotemporal Preservation and Routing (StPR), a unified and exemplar-free VCIL framework that explicitly disentangles and preserves spatiotemporal information. We begin by introducing Frame-Shared Semantics Distillation (FSSD), which identifies semantically stable and meaningful channels by jointly considering channel-wise sensitivity and classification contribution. By selectively regularizing these important semantic channels, FSSD preserves prior knowledge while allowing for adaptation. Building on this preserved semantic space, we further design a Temporal Decomposition-based Mixture-of-Experts (TD-MoE), which dynamically routes task-specific experts according to temporal dynamics, thereby enabling inference without task IDs or stored exemplars. Through the synergy of FSSD and TD-MoE, StPR progressively leverages spatial semantics and temporal dynamics, culminating in a unified, exemplar-free VCIL framework. Extensive experiments on UCF101, HMDB51, SSv2 and Kinetics400 show that our method outperforms existing baselines while offering improved interpretability and efficiency in VCIL. Code is available in the supplementary materials.

Reliable cognitive load (CL) prediction in real-world settings is fundamentally constrained by the poor and variable quality of physiological signals. In safety-critical tasks such as driving, degraded signal quality can severely compromise prediction accuracy, limiting the deployment of existing models outside controlled lab conditions. To address this challenge, we propose CogMoE, a signal quality–guided Mixture-of-Experts (MoE) framework that dynamically adapts to heterogeneous and noisy inputs. CogMoE flexibly integrates physiological modalities, including EEG, ECG, EDA, and gaze, through quality-aware gating, enabling context-sensitive fusion. The framework operates in two stages: (1) quality-aware multi-modal synchronization and recovery to mitigate artifacts, temporal misalignment, and missing data, and (2) signal-quality-specific expert modeling via a cross-modal MoE transformer that regulates information flow based on signal reliability. To further improve stability, we introduce CORTEX Loss, which balances reconstruction fidelity and expert utilization under noise. Experiments on CL-Drive and ADABase show that CogMoE outperforms strong baselines, delivering consistent improvements across diverse signal qualities.

Aligning large language models (LLMs) to simultaneously satisfy multiple objectives remains a significant challenge, especially given the diverse and often conflicting nature of human preferences. Existing alignment methods struggle to balance trade-offs effectively, often requiring costly retraining or yielding suboptimal results across the Pareto frontier of preferences. In this paper, we introduce HoE (Hierarchical Mixture-of-Experts), a lightweight, parameter-efficient, and plug-and-play approach that eliminates the need for model retraining, while enabling LLMs to adapt across the entire Pareto frontier and accommodate diverse user preferences. In particular, HoE consists of three hierarchical components: LoRA Experts, Router Experts and Weighting Router, reaching optimal Pareto frontiers and achieving a trade-off between parameter size, training cost, and performance. We evaluate HoE across various tasks on 16 objectives and 200 different preferences among 8 benchmarks, demonstrating superior performance over 15 recent baselines.

Sequence modelling requires determining which past tokens are causally relevant from the context and their importance: a process inherent to the attention layers in transformers, yet whose underlying learned mechanisms remain poorly understood. In this work, we formalize the task of estimating token importance as an in-context learning problem by introducing a novel framework based on Mixture of Transition Distributions, whereby a latent variable, whose distribution is parameterized by a set of unobserved mixture weights, determines the influence of past tokens on the next. To correctly predict the next token, transformers need to learn the mixture weights in-context. We demonstrate that transformers can implement Mirror Descent to learn the mixture weights from the context. To this end, we give an explicit construction of a three-layer transformer that exactly implements one step of Mirror Descent and prove that the resulting estimator is a first-order approximation of the Bayes-optimal predictor. Corroborating our construction and its learnability via gradient descent, we empirically show that transformers trained from scratch converge to this solution: attention maps match our construction, and deeper models’ performance aligns with multi-step Mirror Descent.

Multiple Instance Learning (MIL) is the predominant approach for classifying gigapixel whole-slide images in computational pathology. MIL follows a sequence of 1) extracting patch features, 2) applying a linear layer to obtain task-specific patch features, and 3) aggregating the patches into a slide feature for classification. While substantial efforts have been devoted to optimizing patch feature extraction and aggregation, none have yet addressed the second point, the critical layer which transforms general-purpose features into task-specific features. We hypothesize that this layer constitutes an overlooked performance bottleneck and that stronger representations can be achieved with a low-rank transformation tailored to each patch's phenotype, yielding synergistic effects with existing MIL approaches. To this end, we introduce MAMMOTH, a parameter-efficient, multi-head mixture of experts module designed to improve the performance of any MIL model with minimal alterations to the total number of parameters. Across 8 MIL methods and 19 different tasks, we find that this improvement to the task-specific transformation has a larger effect on performance than the choice of aggregation method. For instance, when equipped with MAMMOTH, even simple methods such as max or mean pooling attain higher average performance than any method with the standard linear layer. Finally, we identify Instance-Gradient Interference (IGI)—a limitation where heterogeneous instances produce conflicting gradients when processed by a single linear layer—and show that MAMMOTH effectively mitigates IGI by decoupling gradient flows between experts, yielding consistent performance gains in 130 of the 152 examined configurations.

Mixture-of-Experts (MoE) architectures have emerged as a promising approach for scaling deep learning models while maintaining computational efficiency. However, existing MoE adaptations for Contrastive Language-Image Pre-training (CLIP) models suffer from significant computational overhead during sequential training and degradation of zero-shot capabilities. To address these limitations, we propose CLIP-FMoE, a novel approach that integrates MoE architecture into CLIP fine-tuning. Our method uses Isolated Constrained Contrastive Learning, a pipeline that trains specialized experts on cluster-based data partitions to accelerate expert specialization. Additionally, we introduce a Fusion Gate mechanism to mitigate catastrophic forgetting of pre-trained knowledge. Extensive experiments across multiple benchmarks demonstrate that our approach achieves consistent improvements on downstream tasks while preserving zero-shot capabilities. Furthermore, our method demonstrates robust performance across varying context lengths, making it particularly suitable for diverse real-world applications.

Mixture-of-Experts (MoE) has become a cornerstone in recent state-of-the-art large language models (LLMs). Traditionally, MoE relies on $\mathrm{Softmax}$ as the router score function to aggregate expert output, a designed choice that has persisted from the earliest MoE models to modern LLMs, and is now widely regarded as standard practice. However, the necessity of using $\mathrm{Softmax}$ to project router weights into a probability simplex remains an unchallenged assumption rather than a principled design choice. In this work, we first revisit the classical Nadaraya–Watson regression and observe that MoE shares the same mathematical formulation as Nadaraya–Watson regression. Furthermore, we show that both feed-forward neural network (FFN) and Mixture-of-Experts (MoE) can be interpreted as a special case of Nadaraya–Watson regression, where the kernel function corresponds to the input neurons of the output layer. Motivated by these insights, we propose the **zero-additional-cost** Kernel Inspired Router with Normalization ($\mathrm{KERN}$), an FFN-style router function, as an alternative to $\mathrm{Softmax}$. We demonstrate that this router generalizes both $\mathrm{Sigmoid}$- and $\mathrm{Softmax}$-based routers. **Based on empirical observations and established practices in FFN implementation, we recommend the use of $\mathrm{ReLU}$ activation and $\ell_2$-normalization in $\mathrm{KERN}$ router function.** Comprehensive experiments in MoE and LLM validate the effectiveness of the proposed FFN-style router function $\mathrm{KERN}$.

Traditional Mixture-of-Experts (MoE) models lack explicit constraints to ensure the router’s decisions align well with the experts’ capabilities, which ultimately limits model performance. To address this, we propose expert-router coupling loss (ERC loss), a lightweight auxiliary loss that couples expert capabilities and the router’s decisions. We treat each row of the router matrix as a cluster center for the tokens assigned to a particular expert. From these centers, we create proxy tokens by applying a perturbation with noise. Using these proxy tokens, the ERC loss forces the router and experts to satisfy two constraints: (1) each expert exhibits higher activation for its corresponding proxy token than for any other proxy token, and (2) each proxy token elicits stronger activation in its designated expert than in any other expert. This optimization leads to two key effects: each row of the router matrix is an accurate representation of its expert’s capabilities, while each expert develops expertise that closely match the tokens routed to it. Our experiments involve pre-training multiple 3B-parameter MoE-LLMs on trillions of tokens in total, providing detailed evidence of the ERC loss’s effectiveness. Additionally, the ERC loss offers flexible control and quantitative tracking of expert specialization levels during training, providing many valuable insights into MoEs.