ICML 2025 论文列表

Motivated by liability and intellectual property concerns over open-weight models we consider the following problem: given the weights of two models, can we test whether they were trained independently---i.e., from independent random initializations? We consider two settings: *constrained* and *unconstrained*. In the constrained setting, we make assumptions about model architecture and training and propose statistical tests that yield exact p-values with respect to the null hypothesis that the models are trained from independent random initializations. We compute the p-values by simulating exchangeable copies of each model under our assumptions and comparing various similarity measures between the original two models versus these copies. We report p-values on pairs of 21 open-weight models (210 total pairs) and find we correctly identify all pairs of non-independent models. In the unconstrained setting we make none of the prior assumptions and allow for adversarial evasion attacks that do not change model output. We thus propose a new test which matches hidden activations between two models, which is robust to these transformations and to changes in model architecture and can also identify specific non-independent components of models. Though we no longer obtain exact p-values from this test, empirically we find it reliably distinguishes non-independent models like a p-value. Notably, we can use the test to identify specific parts of one model that are derived from another (e.g., how Llama 3.1-8B was pruned to initialize Llama 3.2-3B, or shared layers between Mistral-7B and StripedHyena-7B), and it is even robust to retraining individual layers of either model from scratch.

Similar to surprising performance in the standard deep learning, deep nets trained by adversarial training also generalize well for unseen clean data (natural data). However, despite adversarial training can achieve low robust training error, there exists a significant robust generalization gap. We call this phenomenon the Clean Generalization and Robust Overfitting (CGRO). In this work, we study the CGRO phenomenon in adversarial training from two views: representation complexity and training dynamics. Specifically, we consider a binary classification setting with $N$ separated training data points. First, we prove that, based on the assumption that we assume there is $\operatorname{poly}(D)$-size clean classifier (where $D$ is the data dimension), ReLU net with only $O(N D)$ extra parameters is able to leverages robust memorization to achieve the CGRO, while robust classifier still requires exponential representation complexity in worst case. Next, we focus on a structured-data case to analyze training dynamics, where we train a two-layer convolutional network with $O(N D)$ width against adversarial perturbation. We then show that a three-stage phase transition occurs during learning process and the network provably converges to robust memorization regime, which thereby results in the CGRO. Besides, we also empirically verify our theoretical analysis by experiments in real-image recognition datasets.

Relational reasoning is a central component of generally intelligent systems, enabling robust and data-efficient inductive generalization. Recent empirical evidence shows that many existing neural architectures, including Transformers, struggle with tasks requiring relational reasoning. In this work, we distinguish between two types of information: *sensory* information about the properties of individual objects, and *relational* information about the relationships between objects. While neural attention provides a powerful mechanism for controlling the flow of sensory information between objects, the Transformer lacks an explicit computational mechanism for routing and processing relational information. To address this limitation, we propose an architectural extension of the Transformer framework that we call the *Dual Attention Transformer (DAT)*, featuring two distinct attention mechanisms: sensory attention for directing the flow of sensory information, and a novel relational attention mechanism for directing the flow of relational information. We empirically evaluate *DAT* on a diverse set of tasks ranging from synthetic relational benchmarks to complex real-world tasks such as language modeling and visual processing. Our results demonstrate that integrating explicit relational computational mechanisms into the Transformer architecture leads to significant performance gains in terms of data efficiency and parameter efficiency.

Traditional network inference methods, such as Gaussian Graphical Models, which are built on continuity and homogeneity, face challenges when modeling discrete data and heterogeneous frameworks. Furthermore, under high-dimensionality, the parameter estimation of such models can be hindered by the notorious intractability of high-dimensional integrals. In this paper, we introduce a new and flexible device for graphical models, which accommodates diverse data types, including Gaussian, Poisson log-normal, and latent Gaussian copula models. The new device is driven by a new marginally recoverable parametric family, which can be effectively estimated without evaluating the high-dimensional integration in high-dimensional settings thanks to the marginal recoverability. We further introduce a mixture of marginally recoverable models to capture ubiquitous heterogeneous structures. We show the validity of the desirable properties of the models and the effective estimation methods, and demonstrate their advantages over the state-of-the-art network inference methods via extensive simulation studies and a gene regulatory network analysis of real single-cell RNA sequencing data.

The availability of machine learning systems that can effectively perform arbitrary tasks has led to synthetic labels from these systems being used in applications of statistical inference, such as data analysis or model evaluation. The Prediction Powered Inference (PPI) framework provides a way of leveraging both a large pool of pseudo-labelled data and a small sample with real, high-quality labels to produce a low-variance, unbiased estimate of the quantity being evaluated for. Most work on PPI considers a relatively sizable set of labelled samples, which can be resource intensive to obtain. However, we find that when labelled data is scarce, the PPI++ method can perform even worse than classical inference. We analyze this phenomenon by relating PPI++ to ordinary least squares regression, which also experiences high variance with small sample sizes, and use this regression framework to better understand the efficacy of PPI. Motivated by this, we present two new PPI-based techniques that leverage robust regressors to produce even lower variance estimators in the few-label regime

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 $2.83$\%.

Time-series forecasting is crucial across various domains, including finance, healthcare, and energy. Transformer models, originally developed for natural language processing, have demonstrated significant potential in addressing challenges associated with time-series data. These models utilize different tokenization strategies, point-wise, patch-wise, and variate-wise, to represent time-series data, each resulting in different scope of attention maps. Despite the emergence of sophisticated architectures, simpler transformers consistently outperform their more complex counterparts in widely used benchmarks. This study examines why point-wise transformers are generally less effective, why intra- and inter-variate attention mechanisms yield similar outcomes, and which architectural components drive the success of simpler models. By analyzing mutual information and evaluating models on synthetic datasets, we demonstrate that intra-variate dependencies are the primary contributors to prediction performance on benchmarks, while inter-variate dependencies have a minor impact. Additionally, techniques such as Z-score normalization and skip connections are also crucial. However, these results are largely influenced by the self-dependent and stationary nature of benchmark datasets. By validating our findings on real-world healthcare data, we provide insights for designing more effective transformers for practical applications.

Quantized large language models (LLMs) have gained increasing attention and significance for enabling deployment in resource-constrained environments. However, emerging studies on a few calibration dataset-free quantization methods suggest that quantization may compromise the safety capabilities of LLMs, underscoring the urgent need for systematic safety evaluations and effective mitigation strategies. In this paper, we present comprehensive safety evaluations across various mainstream quantization techniques and diverse calibration datasets, utilizing widely accepted safety benchmarks. To address the identified safety vulnerabilities, we propose a quantization-aware safety patching framework, Q-resafe, to efficiently restore the safety capabilities of quantized LLMs while minimizing any adverse impact on utility. Extensive experiment results demonstrate that Q-resafe successfully re-aligns the safety of quantized LLMs with their pre-quantization counterparts, even under challenging evaluation scenarios. Project page: https://github.com/Thecommonirin/Qresafe.

With the growing interest in embodied and spatial intelligence, accurately predicting trajectories in 3D environments has become increasingly critical. However, no datasets have been explicitly designed to study 3D trajectory prediction. To this end, we contribute a 3D motion trajectory (3DMoTraj) dataset collected from unmanned underwater vehicles (UUVs) operating in oceanic environments. Mathematically, trajectory prediction becomes significantly more complex when transitioning from 2D to 3D. To tackle this challenge, we analyze the prediction complexity of 3D trajectories and propose a new method consisting of two key components: decoupled trajectory prediction and correlated trajectory refinement. The former decouples inter-axis correlations, thereby reducing prediction complexity and generating coarse predictions. The latter refines the coarse predictions by modeling their inter-axis correlations. Extensive experiments show that our method significantly improves 3D trajectory prediction accuracy and outperforms state-of-the-art methods. Both the 3DMoTraj dataset and the method are available at https://github.com/zhouhao94/3DMoTraj.

Leveraging offline data is an attractive way to accelerate online sequential decision-making. However, it is crucial to account for latent states in users or environments in the offline data, and latent bandits form a compelling model for doing so. In this light, we design end-to-end latent bandit algorithms capable of handing uncountably many latent states. We focus on a linear latent contextual bandit — a linear bandit where each user has its own high-dimensional reward parameter in $\mathbb{R}^{d_A}$, but reward parameters across users lie in a low-rank latent subspace of dimension $d_K \ll d_A$. First, we provide an offline algorithm to learn this subspace with provable guarantees. We then present two online algorithms that utilize the output of this offline algorithm to accelerate online learning. The first enjoys $\tilde O(\min(d_A\sqrt{T}, d_K\sqrt{T}(1+\sqrt{d_AT/d_KN})))$ regret guarantees, so that the effective dimension is lower when the size $N$ of the offline dataset is larger. We prove a matching lower bound on regret, showing that our algorithm is minimax optimal. The second is a practical algorithm that enjoys only a slightly weaker guarantee, but is computationally efficient. We also establish the efficacy of our methods using experiments on both synthetic data and real-life movie recommendation data from MovieLens. Finally, we theoretically establish the generality of the latent bandit model by proving a de Finetti theorem for stateless decision processes.

Recent advancements in predictive models have demonstrated exceptional capabilities in predicting the future state of objects and scenes. However, the lack of categorization based on inherent characteristics continues to hinder the progress of predictive model development. Additionally, existing benchmarks are unable to effectively evaluate higher-capability, highly embodied predictive models from an embodied perspective. In this work, we classify the functionalities of predictive models into a hierarchy and take the first step in evaluating World Simulators by proposing a dual evaluation framework called WorldSimBench. WorldSimBench includes Explicit Perceptual Evaluation and Implicit Manipulative Evaluation, encompassing human preference assessments from the visual perspective and action-level evaluations in embodied tasks, covering three representative embodied scenarios: Open-Ended Embodied Environment, Autonomous, Driving, and Robot Manipulation. In the Explicit Perceptual Evaluation, we introduce the HF-Embodied Dataset, a video assessment dataset based on fine-grained human feedback, which we use to train a Human Preference Evaluator that aligns with human perception and explicitly assesses the visual fidelity of World Simulater. In the Implicit Manipulative Evaluation, we assess the video-action consistency of World Simulators by evaluating whether the generated situation-aware video can be accurately translated into the correct control signals in dynamic environments. Our comprehensive evaluation offers key insights that can drive further innovation in video generation models, positioning World Simulators as a pivotal advancement toward embodied artificial intelligence.

Self-supervised learning (SSL) methods via joint embedding architectures have proven remarkably effective at capturing semantically rich representations with strong clustering properties, magically in the absence of label supervision. Despite this, few of them have explored leveraging these untapped properties to improve themselves. In this paper, we provide an evidence through various metrics that the encoder's output *encoding* exhibits superior and more stable clustering properties compared to other components. Building on this insight, we propose a novel positive-feedback SSL method, termed **Re**presentation **S**elf-**A**ssignment (ReSA), which leverages the model's clustering properties to promote learning in a self-guided manner. Extensive experiments on standard SSL benchmarks reveal that models pretrained with ReSA outperform other state-of-the-art SSL methods by a significant margin. Finally, we analyze how ReSA facilitates better clustering properties, demonstrating that it effectively enhances clustering performance at both fine-grained and coarse-grained levels, shaping representations that are inherently more structured and semantically meaningful.

Unpaired point cloud completion is crucial for real-world applications, where ground-truth data for complete point clouds are often unavailable. By learning a completion map from unpaired incomplete and complete point cloud data, this task avoids the reliance on paired datasets. In this paper, we propose the \textit{Unbalanced Optimal Transport Map for Unpaired Point Cloud Completion (\textbf{UOT-UPC})} model, which formulates the unpaired completion task as the (Unbalanced) Optimal Transport (OT) problem. Our method employs a Neural OT model learning the UOT map using neural networks. Our model is the first attempt to leverage UOT for unpaired point cloud completion, achieving competitive or superior performance on both single-category and multi-category benchmarks. In particular, our approach is especially robust under the class imbalance problem, which is frequently encountered in real-world unpaired point cloud completion scenarios.

Graph contrastive learning has attracted great interest as a dominant and promising self-supervised representation learning approach in recent years. While existing works follow the basic principle of pulling positive pairs closer and pushing negative pairs far away, they still suffer from several critical problems, such as the underlying semantic disturbance brought by augmentation strategies, the failure of GCN in capturing long-range dependence, rigidness and inefficiency of node sampling techniques. To address these issues, we propose Manifold Learning Inspired Lightweight Graph Contrastive Learning (ML$^2$-GCL), which inherits the merits of both manifold learning and GCN. ML$^2$-GCL avoids the potential risks of semantic disturbance with only one single view. It achieves global nonlinear structure recovery from locally linear fits, which can make up for the defects of GCN. The most amazing advantage is about the lightweight due to its closed-form solution of positive pairs weights and removal of pairwise distances calculation. Theoretical analysis proves the existence of the optimal closed-form solution. Extensive empirical results on various benchmarks and evaluation protocols demonstrate effectiveness and lightweight of ML$^2$-GCL. We release the code at https://github.com/a-hou/ML2-GCL.

Backpropagation provides a generalized configuration for overcoming catastrophic forgetting. Optimizers such as SGD and Adam are commonly used for weight updates in continual learning and continual pre-training. However, access to gradient information is not always feasible in practice due to black-box APIs, hardware constraints, or non-differentiable systems, a challenge we refer to as the gradient bans. To bridge this gap, we introduce ZeroFlow, the first benchmark designed to evaluate gradient-free optimization algorithms for overcoming forgetting. ZeroFlow examines a suite of forward pass-based methods across various algorithms, forgetting scenarios, and datasets. Our results show that forward passes alone can be sufficient to mitigate forgetting. We uncover novel optimization principles that highlight the potential of forward pass-based methods in mitigating forgetting, managing task conflicts, and reducing memory demands. Additionally, we propose new enhancements that further improve forgetting resistance using only forward passes. This work provides essential tools and insights to advance the development of forward-pass-based methods for continual learning.

Leveraging the powerful representation learning capabilities, deep multi-view clustering methods have demonstrated reliable performance by effectively integrating multi-source information from diverse views in recent years. Most existing methods rely on the assumption of clean views. However, noise is pervasive in real-world scenarios, leading to a significant degradation in performance. To tackle this problem, we propose a novel multi-view clustering framework for the automatic identification and rectification of noisy data, termed AIRMVC. Specifically, we reformulate noisy identification as an anomaly identification problem using GMM. We then design a hybrid rectification strategy to mitigate the adverse effects of noisy data based on the identification results. Furthermore, we introduce a noise-robust contrastive mechanism to generate reliable representations. Additionally, we provide a theoretical proof demonstrating that these representations can discard noisy information, thereby improving the performance of downstream tasks. Extensive experiments on six benchmark datasets demonstrate that AIRMVC outperforms state-of-the-art algorithms in terms of robustness in noisy scenarios. The code of AIRMVC are available at https://github.com/xihongyang1999/AIRMVC on Github.

Efficient access to high-quality information is vital for online platforms. To promote more useful information, users not only create new content but also evaluate existing content, often through helpfulness voting. Although aggregated votes help service providers rank their user content, these votes are often biased by disparate accessibility per position and the cascaded influence of prior votes. For a fairer assessment of information quality, we propose the Counterfactual Voting Adjustment (CVA), a causal framework that accounts for the context in which individual votes are cast. Through preliminary and semi-synthetic experiments, we show that CVA effectively models the position and herding biases, accurately recovering the predefined content quality. In a real experiment, we demonstrate that reranking content based on the learned quality by CVA exhibits stronger alignment with both user sentiment and quality evaluation assessed by GPT-4o, outperforming system rankings based on aggregated votes and model-based rerankings without causal inference. Beyond the individual quality inference, our embeddings offer comparative insights into the behavioral dynamics of expert user groups across 120 major StackExchange communities.

Large language models (LLMs) provide excellent text-generation capabilities, but standard prompting and generation methods generally do not lead to intentional or goal-directed agents and might necessitate considerable prompt tuning. Even the best current LLMs rarely ask clarifying questions, engage in explicit information gathering, or take actions that lead to better decisions after multiple turns. Reinforcement learning has the potential to leverage the powerful modeling capabilities of LLMs, as well as their internal representation of textual interactions, to create capable goal-directed language agents. This can enable intentional and temporally extended interactions, such as with humans, the emergence of complex skills such as persuasion, and long-horizon strategic behavior, such as in the context of games. Enabling this requires the community to develop reliable reinforcement learning algorithms for training LLMs. Developing such algorithms requires tasks that can gauge progress on algorithm design, provide accessible and reproducible evaluations for multi-turn interactions, and cover a range of task properties and challenges in improving reinforcement learning algorithms. Our paper introduces the LMRL-Gym benchmark for evaluating multi-turn RL for LLMs, together with an open-source research framework for getting started on multi-turn RL with offline value-based and online policy-based RL methods. Our benchmark consists of 3 Interactive Dialogue tasks and 5 RL Capability tests for a total of 8 tasks, which require multiple rounds of language interaction and cover a range of tasks in open-ended dialogue and text games.

Depth pruning aims to reduce the inference cost of a large language model without any hardware-specific complications, by simply removing several less important transformer blocks. However, our empirical findings suggest that the importance of a transformer block may be highly task-dependent---a block that is crucial for a task can be removed without degrading the accuracy on another task. Based on this observation, we develop a dynamic depth pruning algorithm, coined PuDDing (**P**rompt-ro**u**ted **D**ynamic **D**epth Prun**ing**), which determines which blocks to omit from the model based on the input prompt. PuDDing operates by training a lightweight router to predict the best omission set among a set of options, where this option set has also been constructed in a data-driven manner. Empirical results on commonsense reasoning benchmarks demonstrate that PuDDing effectively accelerates the inference language models, and achieves better on-task performance than static depth pruning baselines.