ICML 2025 Papers

Large Language Models (LLMs) excel in generation tasks, yet their causal attention mechanisms limit performance in embedding tasks. While bidirectional modeling may enhance embeddings, naively fine-tuning unidirectional models bidirectionally severely degrades generative performance. To investigate this trade-off, we analyze attention weights as dependence indicators and find that bidirectional fine-tuning increases subsequent dependence, impairing unidirectional generation. Through systematic Transformer module evaluations, we discover the FFN layer is least affected by such dependence. Leveraging this discovery, we propose UBMoE-LLM, a novel Uni-Bi-directional Mixture-of-Experts LLM, which integrates the original unidirectional FFN with a bidirectionally fine-tuned FFN via unsupervised contrastive learning. This MoE-based approach enhances embedding performance while preserving robust generation. Extensive experiments across diverse datasets and model scales validate our attention dependence metric and demonstrate UBMoE-LLM’s superior generative quality and reduced hallucination. Code is available at: https://github.com/heiyonghua/ubmoe_llm.

We introduce *Domain2Vec*, a novel approach that decomposes any dataset into a linear combination of several *meta-domains*, a new concept designed to capture the key underlying features of datasets. *Domain2Vec* maintains a vocabulary of meta-domains and uses a classifier to decompose any given dataset into a domain vector that corresponds to a distribution over this vocabulary. These domain vectors enable the identification of optimal data mixture for language model (LM) pretraining in a training-free manner under the ***D**istribution **A**lignment **A**ssumption* (DA$^{2}$), which suggests that when the data distribution of the training set and the validation set is more aligned, a lower validation loss is achieved. Moreover, *Domain2Vec* can be seamlessly integrated into previous works to model the relationship between domain vectors and LM performance, greatly enhancing the efficiency and scalability of previous methods. Extensive experiments demonstrate that *Domain2Vec* helps find the data mixture that enhances downstream task performance with minimal computational overhead. Specifically, *Domain2Vec* achieves the same validation loss on Pile-CC using only $51.5$\% of the compute required when training on the original mixture of The Pile Dataset. Under equivalent compute budget, *Domain2Vec* improves downstream performance by an average of $2.83$\%.

We study the complexity of learning $k$-mixtures of Gaussians ($k$-GMMs) on $\mathbb R^d$. This task is known to have complexity $d^{\Omega(k)}$ in full generality. To circumvent this exponential lower bound on the number of components, research has focused on learning families of GMMs satisfying additional structural properties. A natural assumption posits that the component weights are not exponentially small and that the components have the same unknown covariance. Recent work gave a $d^{O(\log(1/w_{\min}))}$-time algorithm for this class of GMMs, where $w_{\min}$ is the minimum weight. Our first main result is a Statistical Query (SQ) lower bound showing that this quasi-polynomial upper bound is essentially best possible, even for the special case of uniform weights. Specifically, we show that it is SQ-hard to distinguish between such a mixture and the standard Gaussian. We further explore how the distribution of weights affects the complexity of this task. Our second main result is a quasi-polynomial upper bound for the aforementioned testing task when most of the weights are uniform while a small fraction of the weights are potentially arbitrary.

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 9.3$\times$ compression of parameters per expert in Mixtral-8$\times$7B; enables deployment on a GPU with only 11GB VRAM, reducing the memory footprint by up to 8.5$\times$; and delivers a 48.7$\times$ inference speedup compared to DeepSpeed-MII on a single GeForce RTX 3090—all with only a 4.4\% $\sim$ 7.6\% average performance degradation.

Sparse Mixture-of-Experts (SMoE) models represent a significant advancement in large language model (LLM) development through their efficient parameter utilization. These models achieve substantial performance improvements at reduced inference costs. However, the deployment of SMoE models faces constraints from extensive memory requirements of expert components in resource-limited environments. To address these limitations, this paper introduces Hierarchical Clustering for Sparsely activated Mixture of Experts (HC-SMoE), a task-agnostic expert merging framework for parameter reduction without retraining. HC-SMoE introduces a novel hierarchical clustering approach based on expert outputs to ensure merging robustness independent of routing decisions. The proposed output-based clustering method enables effective capture of functional relationships between experts for large- scale architectures. We provide theoretical analysis and comprehensive evaluations across multiple zero-shot language tasks to demonstrate HC-SMoE’s effectiveness in state-of-the-art models including Qwen and Mixtral. The experimental results validate HC-SMoE’s superior performance and practical applicability for real-world deployments. Our implementation is available at https://github.com/wazenmai/HC-SMoE.

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 competitive output quality while maintaining throughput comparable to serving a single model, and incurs only 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.

In this work, we upgrade the multi-head attention mechanism, the core of the Transformer model, to reduce computational costs while maintaining or surpassing the previous accuracy level. We show that multi-head attention can be expressed in the summation form. Drawing on the insight that not all attention heads hold equal significance, we propose Mixture-of-Head attention (MoH), a new architecture that treats attention heads as experts in the Mixture-of-Experts (MoE) mechanism. MoH has two significant advantages: First, MoH enables each token to select the appropriate attention heads, enhancing inference efficiency without compromising accuracy or increasing the number of parameters. Second, MoH replaces the standard summation in multi-head attention with a weighted summation, introducing flexibility to the attention mechanism and unlocking extra performance potential. Extensive experiments on ViT, DiT, and LLMs demonstrate that MoH outperforms multi-head attention by using only 50%$\sim$90% of the attention heads. Moreover, we demonstrate that pre-trained multi-head attention models, such as LLaMA3-8B, can be further continue-tuned into our MoH models. Notably, MoH-LLaMA3-8B achieves an average accuracy of 64.0% across 14 benchmarks, outperforming LLaMA3-8B by 2.4% by utilizing only 75% of the attention heads. We believe the proposed MoH is a promising alternative to multi-head attention and provides a strong foundation for developing advanced and efficient attention-based models.

Domain-Incremental Learning (DIL) focuses on continual learning in non-stationary environments, requiring models to adjust to evolving domains while preserving historical knowledge. DIL faces two critical challenges in the context of imbalanced data: intra-domain class imbalance and cross-domain class distribution shifts. These challenges significantly hinder model performance, as intra-domain imbalance leads to underfitting of few-shot classes, while cross-domain shifts require maintaining well-learned many-shot classes and transferring knowledge to improve few-shot class performance in old domains. To overcome these challenges, we introduce the Dual-Balance Collaborative Experts (DCE) framework. DCE employs a frequency-aware expert group, where each expert is guided by specialized loss functions to learn features for specific frequency groups, effectively addressing intra-domain class imbalance. Subsequently, a dynamic expert selector is learned by synthesizing pseudo-features through balanced Gaussian sampling from historical class statistics. This mechanism navigates the trade-off between preserving many-shot knowledge of previous domains and leveraging new data to improve few-shot class performance in earlier tasks. Extensive experimental results on four benchmark datasets demonstrate DCE’s state-of-the-art performance.

Multimodal person re-identification (Re-ID) aims to match pedestrian images across different modalities. However, most existing methods focus on limited cross-modal settings and fail to support arbitrary query-retrieval combinations, hindering practical deployment. We propose FlexiReID, a flexible framework that supports seven retrieval modes across four modalities: RGB, infrared, sketches, and text. FlexiReID introduces an adaptive mixture-of-experts (MoE) mechanism to dynamically integrate diverse modality features and a cross-modal query fusion module to enhance multimodal feature extraction. To facilitate comprehensive evaluation, we construct CIRS-PEDES, a unified dataset extending four popular Re-ID datasets to include all four modalities. Extensive experiments demonstrate that FlexiReID achieves state-of-the-art performance and offers strong generalization in complex scenarios.

Mixture of Experts (MoE) have shown remarkable success in leveraging specialized expert networks for complex machine learning tasks. However, their susceptibility to adversarial attacks presents a critical challenge for deployment in robust applications. This paper addresses the critical question of how to incorporate robustness into MoEs while maintaining high natural accuracy. We begin by analyzing the vulnerability of MoE components, finding that expert networks are notably more susceptible to adversarial attacks than the router. Based on this insight, we propose a targeted robust training technique that integrates a novel loss function to enhance the adversarial robustness of MoE, requiring only the robustification of one additional expert without compromising training or inference efficiency. Building on this, we introduce a dual-model strategy that linearly combines a standard MoE model with our robustified MoE model using a smoothing parameter. This approach allows for flexible control over the robustness-accuracy trade-off. We further provide theoretical foundations by deriving certified robustness bounds for both the single MoE and the dual-model. To push the boundaries of robustness and accuracy, we propose a novel joint training strategy JTDMoE for the dual-model. This joint training enhances both robustness and accuracy beyond what is achievable with separate models. Experimental results on CIFAR-10 and TinyImageNet datasets using ResNet18 and Vision Transformer (ViT) architectures demonstrate the effectiveness of our proposed methods.

Modern language models are trained on large, unstructured datasets consisting of trillions of tokens and obtained by crawling the web. The unstructured nature makes it difficult to reason about their contents and develop systematic approaches to data curation. In this paper, we unpack monolithic web corpora by developing taxonomies of their contents and organizing them into domains. We introduce WebOrganizer, a framework for organizing web pages in terms of both their topic and format. Using these two complementary notions of domains, we automatically annotate pre-training data by distilling annotations from a large language model into efficient classifiers. This allows us to study how data from different domains should be mixed to improve models on downstream tasks, and we show that we can combine insights about effective topics and formats to further boost performance. We demonstrate that our domain mixing also improves existing methods that select data based on quality. Furthermore, we study and compare how quality-based methods will implicitly change the domain mixture. Overall, our work demonstrates that constructing and mixing domains provides a valuable complement to quality-based data curation methods, opening new avenues for effective and insightful pre-training data curation.

LLaMA-Adapter has recently emerged as an efficient fine-tuning technique for LLaMA models, leveraging zero-initialized attention to stabilize training and enhance performance. However, despite its empirical success, the theoretical foundations of zero-initialized attention remain largely unexplored. In this paper, we provide a rigorous theoretical analysis, establishing a connection between zero-initialized attention and mixture-of-expert models. We prove that both linear and non-linear prompts, along with gating functions, can be optimally estimated, with non-linear prompts offering greater flexibility for future applications. Empirically, we validate our findings on the open LLM benchmarks, demonstrating that non-linear prompts outperform linear ones. Notably, even with limited training data, both prompt types consistently surpass vanilla attention, highlighting the robustness and adaptability of zero-initialized attention.

Large language model pre-training has traditionally relied on human experts to craft heuristics for improving the corpora quality, resulting in numerous rules developed to date. However, these fixed rules lack the flexibility to address the unique characteristics of individual examples, yet crafting sample-wise rules is impractical for human experts. In this paper, we show that even small language models, with only 0.3B parameters, can exhibit substantial data refining capabilities. We propose Programming Every Example (ProX), a novel framework that treats data refinement as a programming task, and enables the model to refine corpora by generating and executing fine-grained operations, such as string normalization, for each individual example at scale. Experiments show that models trained on ProX-refined data consistently outperform other baselines across 10 benchmarks, demonstrating effectiveness across model sizes (up to 1.7B) and pre-training corpora (C4, RedPajama-V2, FineWeb, FineWeb-Edu, and DCLM). ProX also shows great potential in continual pre-training: on math domain, ProX boosts 7B models by up to 20% within 10B tokens—results typically achieved with much larger scale training (e.g., 200B tokens). We believe ProX offers a way to curate high-quality pre-training data, and finally contributes to efficient LLM development.

Mixture of Experts (MoE) architecture improves Large Language Models (LLMs) with better scaling, but its higher parameter counts and memory demands create challenges for deployment. In this paper, we present MoE-SVD, a new decomposition-based compression framework tailored for MoE LLMs without any extra training. By harnessing the power of Singular Value Decomposition (SVD), MoE-SVD addresses the critical issues of decomposition collapse and matrix redundancy in MoE architectures. Specifically, we first decompose experts into compact low-rank matrices, resulting in accelerated inference and memory optimization. In particular, we propose selective decomposition strategy by measuring sensitivity metrics based on weight singular values and activation statistics to automatically identify decomposable expert layers. Then, we share a single V-matrix across all experts and employ a top-k selection for U-matrices. This low-rank matrix sharing and trimming scheme allows for significant parameter reduction while preserving diversity among experts. Comprehensive experiments on Mixtral, Phi-3.5, DeepSeek, and Qwen2 MoE LLMs show MoE-SVD outperforms other compression methods, achieving a 60\% compression ratio and 1.5× faster inference with minimal performance loss. Codes are available at: https://github.com/lliai/MoE-SVD.

Audio-visual speech recognition (AVSR) has become critical for enhancing speech recognition in noisy environments by integrating both auditory and visual modalities. However, existing AVSR systems struggle to scale up without compromising computational efficiency. In this study, we introduce MoHAVE (Mixture of Hierarchical Audio-Visual Experts), a novel robust AVSR framework designed to address these scalability constraints. By leveraging a Mixture-of-Experts (MoE) architecture, MoHAVE activates modality-specific expert groups, ensuring dynamic adaptation to various audio-visual inputs with minimal computational overhead. Key contributions of MoHAVE include: (1) a sparse MoE framework that efficiently scales AVSR model capacity, (2) a hierarchical gating mechanism that dynamically utilizes the expert groups based on input context, enhancing adaptability and robustness, and (3) remarkable performance across robust AVSR benchmarks, including LRS3 and MuAViC transcription and translation tasks, setting a new standard for scalable speech recognition systems.

Pretraining large language models (LLMs) is resource-intensive, often requiring months of training time even with high-end GPU clusters. There are two approaches of mitigating such computational demands: reusing smaller models to train larger ones (upcycling), and training computationally efficient models like mixture-of-experts (MoE). In this paper, we study the upcycling of LLMs to MoE models, of which the scaling behavior remains underexplored. Through extensive experiments, we identify empirical scaling laws that describe how performance depends on dataset size and model configuration. Particularly, we show that, while scaling these factors improves performance, there is a novel interaction term between the dense and upcycled training dataset that limits the efficiency of upcycling at large computational budgets. Based on these findings, we provide guidance to scale upcycling, and establish conditions under which upcycling outperforms from-scratch trainings within budget constraints.

The rapid development of Multimodal Large Language Models (MLLMs), such as GPT-4, marks a significant step toward artificial general intelligence. Existing methods typically align vision encoders with LLMs via supervised fine-tuning (SFT), but this often deteriorates their ability to handle multiple languages as training progresses. We empirically observe that imbalanced SFT datasets, largely English-centric, degrade performance on non-English languages due to the failure in multilingual token alignment. To address this, we propose Parrot, a novel approach that leverages textual guidance for visual token alignment at the language level. Parrot conditions visual tokens on diverse language inputs and uses Mixture-of-Experts (MoE) to align multilingual tokens. By computing cross-attention between initial visual features and textual embeddings, we select the most relevant experts, converting visual tokens into language-specific representations. Additionally, we introduce the Massive Multilingual Multimodal Benchmark (MMMB), a new benchmark comprising 6 languages, 15 categories, and 12,000 questions, to assess multilingual capabilities. Parrot achieves state-of-the-art performance on both the multilingual benchmarks and a wide range of multimodal tasks. Code and dataset are available at: \url{https://github.com/AIDC-AI/Parrot}.

Achieving effective unified pretraining on large time series corpora remains an open challenge in developing time series foundation models. Existing methods, such as Moirai, introduce multiple projection layers for time series of different frequencies to account for high data heterogeneity. We identify major drawbacks to this human-imposed frequency-level model specialization. First, frequency is not a reliable indicator for grouping pretraining data. Second, time series can display varied distributions even within a short window. Frequency-level specialization overlooks the diversity at this granularity. To address these issues, this paper introduces Moirai-MoE, excluding human-defined data groupings while delegating the modeling of diverse time series patterns to the sparse mixture of experts (MoE) within Transformers. With this design, Moirai-MoE eliminates reliance on heuristics and enables automatic token-level specialization. Extensive evaluations on 39 datasets demonstrate the superiority of Moirai-MoE over state-of-the-art foundation models. This study also conducts comprehensive model analyses to explore the inner workings of time series MoE foundation models.

Low-rank Adaptation (LoRA) has emerged as a powerful and efficient method for fine-tuning large-scale foundation models. Despite its popularity, the theoretical understanding of LoRA has remained underexplored. In this paper, we present a theoretical analysis of LoRA by examining its connection to the Mixture of Experts models. Under this framework, we show that a simple technique, reparameterizing LoRA matrices, can notably accelerate the low-rank matrix estimation process. In particular, we prove that reparameterization can reduce the data needed to achieve a desired estimation error from an exponential to a polynomial scale. Motivated by this insight, we propose **Rep**arameterized **Lo**w-**R**ank **A**daptation (RepLoRA), incorporating a lightweight MLP to reparameterize the LoRA matrices. Extensive experiments across multiple domains demonstrate that RepLoRA consistently outperforms vanilla LoRA. With limited data, RepLoRA surpasses LoRA by a substantial margin of up to **40.0%** and achieves LoRA's performance using only **30.0%** of the training data, highlighting the theoretical and empirical robustness of our PEFT method.