ICML 2025 Papers

On-device learning has emerged as a promising direction for AI development, particularly because of its potential to reduce latency issues and mitigate privacy risks associated with device-server communication, while improving energy efficiency. Despite these advantages, significant memory and computational constraints still represent major challenges for its deployment. Drawing on previous studies on low-rank decomposition methods that address activation memory bottlenecks in backpropagation, we propose a novel shortcut approach as an alternative. Our analysis and experiments demonstrate that our method can reduce activation memory usage, even up to $120.09\times$ compared to vanilla training, while also reducing overall training FLOPs up to $1.86\times$ when evaluated on traditional benchmarks.

In many scenarios of practical interest, labeled data from a target distribution are scarce while labeled data from a related source distribution are abundant. One particular setting of interest arises when the target label space is a subset of the source label space, leading to the framework of partial domain adaptation (PDA). Typical approaches to PDA involve minimizing a domain alignment term and a weighted empirical loss on the source data, with the aim of transferring knowledge between domains. However, a theoretical basis for this procedure is lacking, and in particular, most existing weighting schemes are heuristic. In this work, we derive generalization bounds for the PDA problem based on partial optimal transport. These bounds corroborate the use of the partial Wasserstein distance as a domain alignment term, and lead to theoretically motivated explicit expressions for the empirical source loss weights. Inspired by these bounds, we devise a practical algorithm for PDA, termed WARMPOT. Through extensive numerical experiments, we show that WARMPOT is competitive with recent approaches, and that our proposed weights improve on existing schemes.

K-Means clustering is a classical and effective unsupervised learning method attributed to its simplicity and efficiency. However, it faces notable challenges, including sensitivity to random initial centroid selection, a limited ability to discover the intrinsic manifold structures within nonlinear datasets, and difficulty in achieving balanced clustering in practical scenarios. To overcome these weaknesses, we introduce a novel framework for K-Means that leverages manifold learning. This approach eliminates the need for centroid calculation and utilizes a cluster indicator matrix to align the manifold structures, thereby enhancing clustering accuracy. Beyond the traditional Euclidean distance, our model incorporates Gaussian kernel distance, K-nearest neighbor distance, and low-pass filtering distance to effectively manage data that is not linearly separable. Furthermore, we introduce a balanced regularizer to achieve balanced clustering results. The detailed experimental results demonstrate the efficacy of our proposed methodology.

Recent studies have shown that supervised fine-tuning of LLMs on a small number of high-quality datasets can yield strong reasoning capabilities. However, full fine-tuning (Full FT), while powerful, is computationally expensive and susceptible to overfitting and catastrophic forgetting, particularly when data is limited. Sparse fine-tuning, which previously achieved notable success by updating only a small subset of model parameters, offers a promising trade-off between efficiency and effectiveness. Yet, it has lagged behind in the LLM era due to the difficulty of identifying parameters truly critical for reasoning. In this work, we state that weights with the largest magnitude after low-rank approximation are critical weights for fine-tuning, which we call *Principal Weights*. Surprisingly, while magnitude-based sparse fine-tuning performs poorly as a baseline on LLM fine-tuning, it becomes highly effective after rank reduction. These insights motivate our method: **L**ow-rank **I**nformed Sparse **F**ine-**T**uning ($\texttt{LIFT}$). $\texttt{LIFT}$ only updates the top 5% *Principal Weights* throughout training and consistently achieves better performance on reasoning tasks than Full FT, while maintaining memory efficiency on par with popular parameter-efficient fine-tuning methods. In addition to strong performance on target domains such as arithmetic reasoning, $\texttt{LIFT}$ also retains up to 20% more source-domain knowledge, compared to Full FT and LoRA. Our code is available at: https://github.com/zihanghliu/LIFT.

We present **HealthGPT**, a powerful Medical Large Vision-Language Model (Med-LVLM) that integrates medical visual comprehension and generation capabilities within a unified autoregressive paradigm. Our bootstrapping philosophy is to progressively adapt heterogeneous comprehension and generation knowledge to pre-trained Large Language Models (LLMs). This is achieved through a novel heterogeneous low-rank adaptation **(H-LoRA)** technique, which is complemented by a tailored hierarchical visual perception **(HVP)** approach and a three-stage learning strategy **(TLS)**. To effectively learn the HealthGPT, we devise a comprehensive medical domain-specific comprehension and generation dataset called **VL-Health**. Experimental results demonstrate exceptional performance and scalability of HealthGPT in medical visual unified tasks. Our project can be accessed at https://github.com/DCDmllm/HealthGPT.

Large Language Models (LLMs) are increasingly deployed in large-scale online services, enabling sophisticated applications. However, the computational overhead of generating key-value (KV) caches in the prefill stage presents a major bottleneck, particularly for long-context inputs. Prefix caching mitigates this issue by storing KV caches for reuse, reducing redundant computation. Despite its advantages, prefix caching suffers from high latency due to the limited I/O bandwidth of storage devices, constraining inference efficiency. To address this challenge, we introduce Cake, a novel KV cache loading system that optimally utilizes both computational and I/O resources in parallel. Cake employs a bidirectional scheduling strategy that dynamically balances KV cache computation and loading, ensuring efficient resource utilization. Additionally, Cake incorporates an adaptive scheduling mechanism that seamlessly integrates with non-prefix caching requests, improving system throughput and adapting to fluctuating resource availabilty. Through extensive evaluations across various hardware configurations, datasets, and storage conditions, Cake achieves on average 2.6× reduction in Time to First Token (TTFT) compared to compute-only and I/O-only methods. Our findings highlight Cake as an effective and practical solution for optimizing long-context LLM inference, bridging the gap between computation and I/O efficiency in large-scale AI deployments.

The entropic-regularized optimal transport (OT) has gained massive attention in machine learning due to its ability to provide scalable solutions for OT-based tasks. However, most of the existing algorithms, including the Sinkhorn algorithm and its extensions, suffer from relatively slow convergence in many cases. More recently, some second-order methods have been proposed based on the idea of Hessian sparsification. Despite their promising results, they have two major issues: first, there is limited theoretical understanding on the effect of sparsification; second, in cases where the transport plan is dense, Hessian sparsification does not perform well. In this paper, we propose a new quasi-Newton method to address these problems. First, we develop new theoretical analyses to understand the benefits of Hessian sparsification, which lays the foundation for highly flexible sparsification schemes. Then we introduce an additional low-rank term in the approximate Hessian to better handle the dense case. Finally, the convergence properties of the proposed algorithm are rigorously analyzed, and various numerical experiments are conducted to demonstrate its improved performance in solving large-scale OT problems.

The Segment Anything Model (SAM), a vision foundation model, exhibits impressive zero-shot capabilities in general tasks but struggles in specialized domains. Parameter-efficient fine-tuning (PEFT) is a promising approach to unleash the potential of SAM in novel scenarios. However, existing PEFT methods for SAM neglect the domain-invariant relations encoded in the pre-trained model. To bridge this gap, we propose InfoSAM, an information-theoretic approach that enhances SAM fine-tuning by distilling and preserving its pre-trained segmentation knowledge. Specifically, we formulate the knowledge transfer process as two novel mutual information-based objectives: (i) to compress the domain-invariant relation extracted from pre-trained SAM, excluding pseudo-invariant information as possible, and (ii) to maximize mutual information between the relational knowledge learned by the teacher (pre-trained SAM) and the student (fine-tuned model). The proposed InfoSAM establishes a robust distillation framework for PEFT of SAM. Extensive experiments across diverse benchmarks validate InfoSAM's effectiveness in improving SAM family's performance on real-world tasks, demonstrating its adaptability and superiority in handling specialized scenarios. The code and models are available at https://muyaoyuan.github.io/InfoSAM_Page.

We present `GL-LowPopArt`, a novel Catoni-style estimator for generalized low-rank trace regression. Building on `LowPopArt` (Jang et al., 2024), it employs a two-stage approach: nuclear norm regularization followed by matrix Catoni estimation. We establish state-of-the-art estimation error bounds, surpassing existing guarantees (Fan et al., 2019; Kang et al., 2022), and reveal a novel experimental design objective, $\mathrm{GL}(\pi)$. The key technical challenge is controlling bias from the nonlinear inverse link function, which we address by our two-stage approach. We prove a *local* minimax lower bound, showing that our `GL-LowPopArt` enjoys instance-wise optimality up to the condition number of the ground-truth Hessian. Applications include generalized linear matrix completion, where `GL-LowPopArt` achieves a state-of-the-art Frobenius error guarantee, and **bilinear dueling bandits**, a novel setting inspired by general preference learning (Zhang et al., 2024). Our analysis of a `GL-LowPopArt`-based explore-then-commit algorithm reveals a new, potentially interesting problem-dependent quantity, along with improved Borda regret bound than vectorization (Wu et al., 2024).

Graph Domain Adaptation (GDA) transfers knowledge from labeled source graphs to unlabeled target graphs, addressing the challenge of label scarcity. In this paper, we highlight the significance of graph homophily, a pivotal factor for graph domain alignment, which, however, has long been overlooked in existing approaches. Specifically, our analysis first reveals that homophily discrepancies exist in benchmarks. Moreover, we also show that homophily discrepancies degrade GDA performance from both empirical and theoretical aspects, which further underscores the importance of homophily alignment in GDA. Inspired by this finding, we propose a novel homophily alignment algorithm that employs mixed filters to smooth graph signals, thereby effectively capturing and mitigating homophily discrepancies between graphs. Experimental results on a variety of benchmarks verify the effectiveness of our method.

Test-Time Adaptation (TTA) enables deep neural networks to adapt to arbitrary distributions during inference. Existing TTA algorithms generally tend to select benign samples that help achieve robust online prediction and stable self-training. Although malicious samples that would undermine the model's optimization should be filtered out, it also leads to a waste of test data. To alleviate this issue, we focus on how to make full use of the malicious test samples for TTA by transforming them into benign ones, and propose a plug-and-play method, PTTA. The core of our solution lies in the purification strategy, which retrieves benign samples having opposite effects on the objective function to perform Mixup with malicious samples, based on a saliency indicator for encoding benign and malicious data. This strategy results in effective utilization of the information in malicious samples and an improvement of the models' online test accuracy. In this way, we can directly apply the purification loss to existing TTA algorithms without the need to carefully adjust the sample selection threshold. Extensive experiments on four types of TTA tasks as well as classification, segmentation, and adversarial defense demonstrate the effectiveness of our method. Code is available at https://github.com/HAIV-Lab/PTTA.

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.

While Low-Rank Adaptation (LoRA) enables parameter-efficient fine-tuning for Large Language Models (LLMs), its performance often falls short of Full Fine-Tuning (Full FT). Current methods optimize LoRA by initializing with static singular value decomposition (SVD) subsets, leading to suboptimal leveraging of pre-trained knowledge. {Another path for improving LoRA is incorporating a Mixture-of-Experts (MoE) architecture.} {However, weight misalignment and complex gradient dynamics make it challenging to adopt SVD prior to the LoRA MoE architecture.} To mitigate these issues, we propose \underline{G}reat L\underline{o}R\underline{A} Mixture-of-Exper\underline{t} (GOAT), a framework that (1) adaptively integrates relevant priors using an SVD-structured MoE, and (2) aligns optimization with full fine-tuned MoE by deriving a theoretical scaling factor. We demonstrate that proper scaling, without modifying the architecture or training algorithms, boosts LoRA MoE’s efficiency and performance. Experiments across 25 datasets, including natural language understanding, commonsense reasoning, image classification, and natural language generation, demonstrate GOAT’s state-of-the-art performance, closing the gap with Full FT. Our code is available at: https://github.com/Facico/GOAT-PEFT

Pruning is a widely used technique for compressing large neural networks that eliminates weights that have minimal impact on the model's performance. Current pruning methods, exemplified by magnitude pruning, assign an importance score to each weight based on its magnitude and remove weights with scores below a certain threshold. Nonetheless, these methods often create a gap between the original dense and the pruned sparse model, potentially impairing performance. Especially when the sparsity ratio is high, the gap becomes more pronounced. To mitigate this issue, we introduce a method to bridge the gap left by pruning by utilizing a low-rank approximation of the difference between the dense and sparse matrices. Our method entails the iterative refinement of the sparse weight matrix augmented by a low-rank adjustment. This technique captures and retains the essential information often lost during pruning, thereby improving the performance of the pruned model. Furthermore, we offer a comprehensive theoretical analysis of our approach, emphasizing its convergence properties and establishing a solid basis for its efficacy. Experimental results on LLaMa models validate its effectiveness on large language models across various pruning techniques and sparsity levels. Our method shows significant improvements: at 50\% sparsity, it reduces perplexity by 53.9\% compared to conventional magnitude pruning on LLaMa-7B. Furthermore, to achieve a specific performance target, our approach enables an 8.6\% reduction in model parameters while maintaining a sparsity ratio of about 50\%.

The core component of attention is the scoring function, which transforms the inputs into low-dimensional queries and keys and takes the dot product of each pair. While the low-dimensional projection improves efficiency, it causes information loss for certain tasks that have intrinsically high-dimensional inputs. Additionally, attention uses the same scoring function for all input pairs, without imposing a distance-dependent compute bias for neighboring tokens in the sequence. In this work, we address these shortcomings by proposing new scoring functions based on computationally efficient structured matrices with high ranks, including Block Tensor-Train (BTT) and Multi-Level Low Rank (MLR) matrices. On in-context regression tasks with high-dimensional inputs, our proposed scoring functions outperform standard attention for any fixed compute budget. On language modeling, a task that exhibits locality patterns, our MLR-based attention method achieves improved scaling laws compared to both standard attention and variants of sliding window attention. Additionally, we show that both BTT and MLR fall under a broader family of efficient structured matrices capable of encoding either full-rank or distance-dependent compute biases, thereby addressing significant shortcomings of standard attention.

We propose Network Automatic Relevance Determination (NARD), an extension of ARD for linearly probabilistic models, to simultaneously model sparse relationships between inputs $X \in \mathbb R^{d \times N}$ and outputs $Y \in \mathbb R^{m \times N}$, while capturing the correlation structure among the $Y$. NARD employs a matrix normal prior which contains a sparsity-inducing parameter to identify and discard irrelevant features, thereby promoting sparsity in the model. Algorithmically, it iteratively updates both the precision matrix and the relationship between $Y$ and the refined inputs. To mitigate the computational inefficiencies of the $\mathcal O(m^3 + d^3)$ cost per iteration, we introduce Sequential NARD, which evaluates features sequentially, and a Surrogate Function Method, leveraging an efficient approximation of the marginal likelihood and simplifying the calculation of determinant and inverse of an intermediate matrix. Combining the Sequential update with the Surrogate Function method further reduces computational costs. The computational complexity per iteration for these three methods is reduced to $\mathcal O(m^3+p^3)$, $\mathcal O(m^3 + d^2)$, $\mathcal O(m^3+p^2)$ respectively, where $p \ll d$ is the final number of features in the model. Our methods demonstrate significant improvements in computational efficiency with comparable performance on both synthetic and real-world datasets.

Fine-tuning large-scale pre-trained models provides an effective solution to alleviate the label scarcity problem in cardiovascular diseases (CVDs) detection using electrocardiogram (ECG). However, as the pre-trained models scale up, the computational costs for fine-tuning and inference become unaffordable on low-level devices deployed for clinical applications. Additionally, maintaining the model performance under low budgets in computational resources remains a significant challenge. However, a comprehensive study that can address them in a joint framework is still lacking. Here, we propose a holistic method (H-Tuning) for low-cost and efficient fine-tuning of pre-trained models on downstream datasets. Then, the inference costs of the models fine-tuned by H-Tuning are further reduced significantly using a knowledge distillation technique. Experiments on four ECG datasets demonstrate that H-Tuning reduces the GPU memory consumption during fine-tuning by 6.34 times while achieving comparable CVDs detection performance to standard fine-tuning. With the knowledge distillation technique, the model inference latency and the memory consumption are reduced by 4.52 times and 19.83 times. As such, the proposed joint framework allows for the utilization of pre-trained models with high computation efficiency and robust performance, exploring a path toward low-cost and efficient CVDs detection. Code is available at https://github.com/KAZABANA/H-Tuning

We introduce GPTAQ, a novel finetuning-free quantization method for compressing large-scale transformer architectures. Unlike the previous GPTQ method, which independently calibrates each layer, we always match the quantized layer's output to the exact output in the full-precision model, resulting in a scheme that we call *asymmetric calibration*. Such a scheme can effectively reduce the quantization error accumulated in previous layers. We analyze this problem using optimal brain compression to derive a close-formed solution. The new solution explicitly minimizes the quantization error as well as the accumulated asymmetry error. Furthermore, we utilize various techniques to parallelize the solution calculation, including channel parallelization, neuron decomposition, and Cholesky reformulation for matrix fusion. As a result, GPTAQ is easy to implement, simply using 20 more lines of code than GPTQ but improving its performance under low-bit quantization. Remarkably, on a single GPU, we quantize a 405B language transformer as well as EVA-02—the rank first vision transformer that achieves 90% pretraining Imagenet accuracy. Code is available at [Github](https://github.com/Intelligent-Computing-Lab-Yale/GPTAQ).

Discovering regularities from spatiotemporal systems can benefit various scientific and social planning. Current spatiotemporal learners usually train an independent model from a specific source data that leads to limited transferability among sources, where even correlated tasks requires new design and training. The key towards increasing cross-domain knowledge is to enable collective intelligence and model evolution. In this paper, inspired by neuroscience theories, we theoretically derive the increased information boundary via learning cross-domain collective intelligence and propose a Synaptic EVOlutional spatiotemporal network, SynEVO, where SynEVO breaks the model independence and enables cross-domain knowledge to be shared and aggregated. Specifically, we first re-order the sample groups to imitate the human curriculum learning, and devise two complementary learners, elastic common container and task-independent extractor to allow model growth and task-wise commonality and personality disentanglement. Then an adaptive dynamic coupler with a new difference metric determines whether the new sample group should be incorporated into common container to achieve model evolution under various domains. Experiments show that SynEVO improves the generalization capacity by at most 42\% under cross-domain scenarios and SynEVO provides a paradigm of NeuroAI for knowledge transfer and adaptation. Code available at [https://github.com/Rodger-Lau/SynEVO](https://github.com/Rodger-Lau/SynEVO).