ICML 2025 Papers

The multi-domain fine-tuning of large language models (LLMs) confronts a notorious trade-off among abilities across domains. Existing studies attribute this trade-off to the conflicts between samples rooted in inherent semantics. Recent approaches attempt to mitigate these conflicts through the empirical investigation or heuristic strategies. However, without a fundamental understanding of interactions between samples, they yield only marginal improvements, while incurring substantial trial-and-error costs. To address this challenge, we move beyond empirical studies by modeling interactions between samples as their influence on each other's loss, estimated using gradients. Intriguingly, we find that these interactions **evolve throughout training** rather than being purely determined by inherent semantics. Building on this insight, we propose **EV**olving **I**nteraction-guided **C**urriculum (**EVIC**), which iteratively selects samples that positively influence the overall dataset for training. By dynamically adapting the training curriculum to prioritize samples that contribute the most to the model training, EVIC effectively mitigates conflicts and improves the sample efficiency. Extensive experiments on a mixed dataset covering coding, math, and general tasks with several model architectures show that EVIC significantly outperforms all baselines across diverse capabilities.

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.

Robust training methods typically defend against specific attack types, such as $\ell_p$ attacks with fixed budgets, and rarely account for the fact that defenders may encounter new attacks over time. A natural solution is to adapt the defended model to new adversaries as they arise via fine-tuning, a method which we call continual robust training (CRT). However, when implemented naively, fine-tuning on new attacks degrades robustness on previous attacks. This raises the question: \textit{how can we improve the initial training and fine-tuning of the model to simultaneously achieve robustness against previous and new attacks?} We present theoretical results which show that the gap in a model's robustness against different attacks is bounded by how far each attack perturbs a sample in the model's logit space, suggesting that regularizing with respect to this logit space distance can help maintain robustness against previous attacks. Extensive experiments on 3 datasets (CIFAR-10, CIFAR-100, and ImageNette) and over 100 attack combinations demonstrate that the proposed regularization improves robust accuracy with little overhead in training time. Our findings and open-source code lay the groundwork for the deployment of models robust to evolving attacks.

Contrastive Language-Image Pre-Training (CLIP) is highly instrumental in machine learning applications within a large variety of domains. We investigate the geometry of this embedding, which is still not well understood, and show that text and image reside on linearly separable ellipsoid shells, not centered at the origin. We explain the benefits of having this structure, allowing to better embed instances according to their uncertainty during contrastive training. Frequent concepts in the dataset yield more false negatives, inducing greater uncertainty. A new notion of conformity is introduced, which measures the average cosine similarity of an instance to any other instance within a representative data set. We show this measure can be accurately estimated by simply computing the cosine similarity to the modality mean vector. Furthermore, we find that CLIP's modality gap optimizes the matching of the conformity distributions of image and text.

Spatial Transcriptomics (ST) technologies provide biologists with rich insights into single-cell biology by preserving spatial context of cells. Building foundational models for ST can significantly enhance the analysis of vast and complex data sources, unlocking new perspectives on the intricacies of biological tissues. However, modeling ST data is inherently challenging due to the need to extract multi-scale information from tissue slices containing vast numbers of cells. This process requires integrating macro-scale tissue morphology, micro-scale cellular microenvironment, and gene-scale gene expression profile. To address this challenge, we propose **SToFM**, a multi-scale **S**patial **T**ranscript**o**mics **F**oundation **M**odel. SToFM first performs multi-scale information extraction on each ST slice, to construct a set of ST sub-slices that aggregate macro-, micro- and gene-scale information. Then an SE(2) Transformer is used to obtain high-quality cell representations from the sub-slices. Additionally, we construct **SToCorpus-88M**, the largest high-resolution spatial transcriptomics corpus for pretraining. SToFM achieves outstanding performance on a variety of downstream tasks, such as tissue region semantic segmentation and cell type annotation, demonstrating its comprehensive understanding of ST data through capturing and integrating multi-scale information.

Parameter-efficient fine-tuning (PEFT) techniques such as low-rank adaptation (LoRA) can effectively adapt large pre-trained foundation models to downstream tasks using only a small fraction (0.1%-10%) of the original trainable weights. An under-explored question of PEFT is in extending the pre-training phase without supervised labels; that is, can we adapt a pre-trained foundation model to a new domain via efficient self-supervised pre-training on this domain? In this work, we introduce ExPLoRA, a highly effective technique to improve transfer learning of pre-trained vision transformers (ViTs) under domain shifts. Initializing a ViT with pre-trained weights on large, natural-image datasets such as from DinoV2 or MAE, ExPLoRA continues the unsupervised pre-training objective on a new domain, unfreezing 1-2 pre-trained ViT blocks and tuning all other layers with LoRA. We then fine-tune the resulting model only with LoRA on this new domain for supervised learning. Our experiments demonstrate state-of-the-art results on satellite imagery, even outperforming fully pre-training and fine-tuning ViTs. Using the DinoV2 training objective, we demonstrate up to 8% improvement in linear probing top-1 accuracy on downstream tasks while using <10% of the number of parameters that are used in prior fully-tuned state-of-the art approaches. Our ablation studies confirm the efficacy of our approach over other baselines such as PEFT. Code is available at: https://samar-khanna.github.io/ExPLoRA/

This paper investigates the convergence properties of the hypergradient descent method ($\texttt{HDM}$), a 25-year-old heuristic originally proposed for adaptive stepsize selection in stochastic first-order methods. We provide the first rigorous convergence analysis of $\texttt{HDM}$ using the online learning framework and apply this analysis to develop a new state-of-the-art adaptive gradient methods with empirical and theoretical support. Notably, $\texttt{HDM}$ automatically identifies the optimal stepsize for the local optimization landscape and achieves local superlinear convergence. Our analysis explains the instability of $\texttt{HDM}$ reported in the literature and proposes efficient strategies to address it. We also develop two $\texttt{HDM}$ variants with heavy-ball and Nesterov momentum. Experiments on deterministic convex problems show $\texttt{HDM}$ with heavy-ball momentum ($\texttt{HDM-HB}$) exhibits robust performance and significantly outperforms other adaptive first-order methods. Moreover, $\texttt{HDM-HB}$ often matches the performance of $\texttt{L-BFGS}$, an efficient and practical quasi-Newton method, using less memory and cheaper iterations.

*Density ratio estimation* (DRE) is a paramount task in machine learning, for its broad applications across multiple domains, such as covariate shift adaptation, causal inference, independence tests and beyond. Parametric methods for estimating the density ratio possibly lead to biased results if models are misspecified, while conventional non-parametric methods suffer from the curse of dimensionality when the dimension of data is large. To address these challenges, in this paper, we propose a novel approach for DRE based on the projection pursuit (PP) approximation. The proposed method leverages PP to mitigate the impact of high dimensionality while retaining the model flexibility needed for the accuracy of DRE. We establish the consistency and the convergence rate for the proposed estimator. Experimental results demonstrate that our proposed method outperforms existing alternatives in various applications.

Fine-tuning large pretrained Transformer models can focus on either introducing a small number of new learnable parameters (parameter efficiency) or editing representations of a small number of tokens using lightweight modules (representation efficiency). While the pioneering method LoRA (Low-Rank Adaptation) inherently balances parameter, compute, and memory efficiency, many subsequent variants trade off compute and memory efficiency and/or performance to further reduce fine-tuning parameters. To address this limitation and unify parameter-efficient and representation-efficient fine-tuning, we propose Weight-Generative Fine-Tuning (WeGeFT, pronounced *wee-gift*), a novel approach that **learns to generate fine-tuning weights directly from the pretrained weights**. WeGeFT employs a simple low-rank formulation consisting of two linear layers, either shared across multiple layers of the pretrained model or individually learned for different layers. This design achieves multi-faceted efficiency in parameters, representations, compute, and memory, while maintaining or exceeding the performance of LoRA and its variants. Extensive experiments on commonsense reasoning, arithmetic reasoning, instruction following, code generation, and visual recognition verify the effectiveness of our proposed WeGeFT.

Universal Domain Adaptation (UniDA) addresses unsupervised domain adaptation where target classes may differ arbitrarily from source ones, except for a shared subset. A widely used approach, partial domain matching (PDM), aligns only shared classes but struggles in extreme cases where many source classes are absent in the target domain, underperforming the most naive baseline that trains on only source data. In this work, we identify that the failure of PDM for extreme UniDA stems from dimensional collapse (DC) in target representations. To address target DC, we propose to jointly leverage the alignment and uniformity techniques in self-supervised learning on the unlabeled target data to preserve the intrinsic structure of the learned representations. Our experimental results confirm that SSL consistently advances PDM and delivers new state-of-the-art results across a broader benchmark of UniDA scenarios with different portions of shared classes, representing a crucial step toward truly comprehensive UniDA. Project page: https://dc-unida.github.io/

Foundation Models have demonstrated significant success across various domains in Artificial Intelligence (AI), yet their capabilities for brainwave modeling remain unclear. In this paper, we comprehensively evaluate current Large Brainwave Foundation Models (LBMs) through systematic fine-tuning experiments across multiple Brain-Computer Interface (BCI) benchmark tasks, including memory tasks and sleep stage classification. Our extensive analysis shows that state-of-the-art LBMs achieve only marginal improvements (0.5\%) over traditional deep architectures while requiring significantly more parameters (millions vs thousands), raising important questions about their efficiency and applicability in BCI contexts. Moreover, through detailed ablation studies and Low-Rank Adaptation (LoRA), we significantly reduce trainable parameters without performance degradation, while demonstrating that architectural and training inefficiencies limit LBMs' current capabilities. Our experiments span both full model fine-tuning and parameter-efficient adaptation techniques, providing insights into optimal training strategies for BCI applications. We pioneer the application of LoRA to LBMs, revealing that performance benefits generally emerge when adapting multiple neural network components simultaneously. These findings highlight the critical need for domain-specific development strategies to advance LBMs, suggesting that current architectures may require redesign to fully leverage the potential of foundation models in brainwave analysis.

We study the problem of distributed distinct element estimation, where $\alpha$ servers each receive a subset of a universe $[n]$ and aim to compute a $(1+\varepsilon)$-approximation to the number of distinct elements using minimal communication. While prior work establishes a worst-case bound of $\Theta\left(\alpha\log n+\frac{\alpha}{\varepsilon^2}\right)$ bits, these results rely on assumptions that may not hold in practice. We introduce a new parameterization based on the number $C = \frac{\beta}{\varepsilon^2}$ of pairwise collisions, i.e., instances where the same element appears on multiple servers, and design a protocol that uses only $O\left(\alpha\log n\log\log n+\frac{\sqrt{\beta}}{\varepsilon^2} \log n\right)$ bits, breaking previous lower bounds when $C$ is small. We further improve our algorithm under assumptions on the number of distinct elements or collisions and provide matching lower bounds in all regimes, establishing $C$ as a tight complexity measure for the problem. Finally, we consider streaming algorithms for distinct element estimation parameterized by the number of items with frequency larger than $1$. Overall, our results offer insight into why statistical problems with known hardness results can be efficiently solved in practice.

Recent studies have uncovered a troubling vulnerability in the fine-tuning stage of large language models (LLMs): even fine-tuning on entirely benign datasets can lead to a significant increase in the harmfulness of LLM outputs. Building on this finding, our red teaming study takes this threat one step further by developing a more effective attack. Specifically, we analyze and identify samples within benign datasets that contribute most to safety degradation, then fine-tune LLMs exclusively on these samples. We approach this problem from an outlier detection perspective and propose Self-Inf-N, to detect and extract outliers for fine-tuning. Our findings reveal that fine-tuning LLMs on 100 outlier samples selected by Self-Inf-N in the benign datasets severely compromises LLM safety alignment. Extensive experiments across seven mainstream LLMs demonstrate that our attack exhibits high transferability across different architectures and remains effective in practical scenarios. Alarmingly, our results indicate that most existing mitigation strategies fail to defend against this attack, underscoring the urgent need for more robust alignment safeguards. Codes are available at https://github.com/GuanZihan/Benign-Samples-Matter.

Multimodal Large Language Model (MLLM) has demonstrated strong generalization capabilities across diverse distributions and tasks, largely due to extensive pre-training datasets. Fine-tuning MLLM has become a common practice to improve performance on specific downstream tasks. However, during fine-tuning, MLLM often faces the risk of forgetting knowledge acquired during pre-training, which can result in a decline in generalization abilities. To balance the trade-off between generalization and specialization, we propose measuring the parameter importance for both pre-trained and fine-tuning distributions, based on frozen pre-trained weight magnitude and accumulated fine-tuning gradient values. We further apply an importance-aware weight allocation strategy, selectively updating relatively important parameters for downstream tasks. We conduct empirical evaluations on both image captioning and visual question-answering tasks using various MLLM architectures. The comprehensive experimental analysis demonstrates the effectiveness of the proposed solution, highlighting the efficiency of the crucial modules in enhancing downstream specialization performance while mitigating generalization degradation in MLLM Fine-Tuning.

Wearable devices record physiological and behavioral signals that can improve health predictions. While foundation models are increasingly used for such predictions, they have been primarily applied to low-level sensor data, despite behavioral data often being more informative due to their alignment with physiologically relevant timescales and quantities. We develop foundation models of such behavioral signals using over 2.5B hours of wearable data from 162K individuals, systematically optimizing architectures and tokenization strategies for this unique dataset. Evaluated on 57 health-related tasks, our model shows strong performance across diverse real-world applications including individual-level classification and time-varying health state prediction. The model excels in behavior-driven tasks like sleep prediction, and improves further when combined with representations of raw sensor data. These results underscore the importance of tailoring foundation model design to wearables and demonstrate the potential to enable new health applications.

Advances in Parameter-efficient Fine-tuning (PEFT) bridged the performance gap with Full Fine-Tuning (FFT) through sophisticated analysis of pre-trained parameter spaces. Starting from drawing insights from Neural Engrams (NE) in Biological Neural Networks (BNNs), we establish a connection between the low-rank property observed during PEFT's parameter space shifting and neurobiological mechanisms. This observation leads to our proposed method, **S**ynapse and **N**euron (**SAN**), which decomposes and propagates the scaling component from anterior feature adjustment vectors towards posterior weight matrices. Our approach is theoretically grounded in Long-Term Potentiation/Depression (LTP/D) phenomena, which govern synapse development through neurotransmitter release modulation. Extensive experiments demonstrate its effectiveness: on **vision tasks** across VTAB, FGVC, and GIC (25 datasets) using ViT, Swin-T and ConvNeXt architectures, SAN outperforms FFT up to *8.7%* and LoRA by *3.2%*; on **language tasks** using Commonsense Reasoning (8 datasets) with LLaMA models (all generations), surpassing ChatGPT up to *8.5%* and LoRA by *4.7%*; on **vision-language tasks** using Visual Instruction Tuning (7 datasets) with LLaVA models, it exceeds FFT up to *2.4%* and LoRA by *1.9%*. Our code and W&B log will be released

Foundation models are pretrained on large-scale corpora to learn generalizable patterns across domains and tasks---such as contours, textures, and edges in images, or tokens and sentences in text. In contrast, discovering such generalities in graph-structured data, especially across heterogeneous graph tasks, remains an open challenge. To address this, we propose a novel approach to cross-task generalization in graphs via task-trees, which serve as unified learning instances aligning node-, edge-, and graph-level tasks. We theoretically analyze the stability, transferability, and generalization properties of task-trees, showing that pretraining a graph neural network (GNN) on diverse task-trees with a reconstruction objective induces transferable knowledge. This enables efficient adaptation to downstream tasks with minimal fine-tuning. To validate our framework, we introduce Graph Generality Identifier on Task-Trees (GIT), a graph foundation model that demonstrates strong performance on over 30 graphs across five domains via fine-tuning, in-context learning, and zero-shot generalization. Code and data are available at https://github.com/Zehong-Wang/GIT.

We address the problem of reward hacking, where maximising a proxy reward does not necessarily increase the true reward. This is a key concern for Large Language Models (LLMs), as they are often fine-tuned on human preferences that may not accurately reflect a true objective. Existing work uses various tricks such as regularisation, tweaks to the reward model, and reward hacking detectors, to limit the influence that such proxy preferences have on a model. Luckily, in many contexts such as medicine, education, and law, a sparse amount of expert data is often available. In these cases, it is often unclear whether the addition of proxy data can improve policy learning. We outline a set of sufficient conditions on proxy feedback that, if satisfied, indicate that proxy data can provably improve the sample complexity of learning the ground truth policy. These conditions can inform the data collection process for specific tasks. The result implies a parameterisation for LLMs that achieves this improved sample complexity. We detail how one can adapt existing architectures to yield this improved sample complexity.

How to best develop foundational models for time series forecasting remains an important open question. Tokenization is a crucial consideration in this effort: what is an effective discrete vocabulary for a real-valued sequential input? To address this question, we develop WaveToken, a wavelet-based tokenizer that allows models to learn complex representations directly in the space of time-localized frequencies. Our method first scales and decomposes the input time series, then thresholds and quantizes the wavelet coefficients, and finally pre-trains an autoregressive model to forecast coefficients for the forecast horizon. By decomposing coarse and fine structures in the inputs, wavelets provide an eloquent and compact language for time series forecasting that simplifies learning. Empirical results on a comprehensive benchmark, including 42 datasets for both in-domain and zero-shot settings, show that WaveToken: i) performs on par or better than recently proposed foundation models for forecasting while using a much smaller vocabulary (1024 tokens), and is competitive with modern deep learning models trained specifically on each dataset; ii) exhibits superior generalization capabilities, achieving the best average rank across all datasets for three complementary metrics; and iii) easily captures complex temporal patterns of practical relevance that are challenging for other recent pre-trained models, including trends, sparse spikes, and non-stationary time series with varying frequencies evolving over time.