ICML 2025 论文列表

Deep networks are prone to catastrophic forgetting during sequential task learning, i.e., losing the knowledge about old tasks upon learning new tasks. To this end, continual learning (CL) has emerged, whose existing methods focus mostly on regulating or protecting the parameters associated with the previous tasks. However, parameter protection is often impractical, since the size of parameters for storing the old-task knowledge increases linearly with the number of tasks, otherwise it is hard to preserve the parameters related to the old-task knowledge. In this work, we bring a dual opinion from neuroscience and physics to CL: in the whole networks, the pathways matter more than the parameters when concerning the knowledge acquired from the old tasks. Following this opinion, we propose a novel CL framework, learning without isolation (LwI), where model fusion is formulated as graph matching and the pathways occupied by the old tasks are protected without being isolated. Thanks to the sparsity of activation channels in a deep network, LwI can adaptively allocate available pathways for a new task, realizing pathway protection and addressing catastrophic forgetting in a parameter-effcient manner. Experiments on popular benchmark datasets demonstrate the superiority of the proposed LwI.

We study how inherent randomness in the training process—where each sample (or client in federated learning) contributes only to a randomly selected portion of training—can be leveraged for privacy amplification. This includes (1) data partitioning, where a sample participates in only a subset of training iterations, and (2) model partitioning, where a sample updates only a subset of the model parameters. We apply our framework to model parallelism in federated learning, where each client updates a randomly selected subnetwork to reduce memory and computational overhead, and show that existing methods, e.g. model splitting or dropout, provide a significant privacy amplification gain not captured by previous privacy analysis techniques. Additionally, we introduce balanced iteration subsampling, a new data partitioning method where each sample (or client) participates in a fixed number of training iterations. We show that in certain regimes, this method yields stronger privacy amplification than Poisson (i.i.d.) sampling of data (or clients). Our results demonstrate that randomness in the training process, which is structured rather than i.i.d. and interacts with data in complex ways, can be systematically leveraged for nontrivial privacy amplification.

This work investigates the problem of best arm identification for multi-agent multi-armed bandits. We consider $N$ agents grouped into $M$ clusters, where each cluster solves a stochastic bandit problem. The mapping between agents and bandits is \textit{a priori} unknown. Each bandit is associated with $K$ arms, and the goal is to identify the best arm for each agent under a $\delta$-probably correct ($\delta$-PC) framework, while minimizing sample complexity and communication overhead. We propose two novel algorithms: \emph{Clustering then Best Arm Identification} (\texttt{Cl-BAI}) and \emph{Best Arm Identification then Clustering} (\texttt{BAI-Cl}). \texttt{Cl-BAI} employs a two-phase approach that first clusters agents based on the bandit problems they are learning, followed by identifying the best arm for each cluster. \texttt{BAI-Cl} reverses the sequence by identifying the best arms first and then clustering agents accordingly. Both algorithms exploit the successive elimination framework to ensure computational efficiency and high accuracy. Theoretical analysis establishes $\delta$-PC guarantees for both methods, derives bounds on their sample complexity, and provides a lower bound for the problem class. Moreover, when $M$ is small (a constant), we show that the sample complexity of (a variant of) \texttt{BAI-Cl} is (order-wise) minimax optimal. Experiments on synthetic and real-world (Movie Lens, Yelp) data demonstrates the superior performance of the proposed algorithms in terms of sample and communication efficiency, particularly in settings where $M \ll N$.

Offline algorithms for Reinforcement Learning from Human Preferences (RLHF), which use only a fixed dataset of sampled responses given an input, and preference feedback among these responses, have gained increasing prominence in the literature on aligning language models. In this paper, we study how the different design choices made in methods such as DPO, IPO, SLiC and many variants influence the quality of the learned policy, from a theoretical perspective. Our treatment yields insights into the choices of loss function, the policy which is used to normalize log-likelihoods, and also the role of the data sampling policy. Notably, our results do not rely on the standard reparameterization-style arguments used to motivate some of the algorithms in this family, which allows us to give a unified treatment to a broad class of methods. We also conduct a small empirical study to verify some of the theoretical findings on a standard summarization benchmark.

Human vision is dynamic and continuous. However, in video understanding with multimodal large language models (LLMs), existing methods primarily rely on static features extracted from images sampled at a fixed low frame rate of frame-per-second (FPS) $\leqslant$2, leading to critical visual information loss. In this paper, we introduce F-16, the first multimodal LLM designed for high-frame-rate video understanding. By increasing the frame rate to 16 FPS and compressing visual tokens within each 1-second clip, F-16 efficiently captures dynamic visual features while preserving key semantic information. Experimental results demonstrate that higher frame rates considerably enhance video understanding across multiple benchmarks, providing a new approach to improving video LLMs beyond scaling model size or training data. F-16 achieves state-of-the-art performance among 7-billion-parameter video LLMs on both general and fine-grained video understanding benchmarks, such as Video-MME and TemporalBench. Furthermore, F-16 excels in complex spatiotemporal tasks, including high-speed sports analysis (*e.g.*, basketball, football, gymnastics, and diving), outperforming SOTA proprietary visual models like GPT-4o and Gemini-1.5-pro. Additionally, we introduce a novel decoding method for F-16 that enables highly efficient low-frame-rate inference without requiring model retraining. We will release the source code, model checkpoints, and data at [https://github.com/bytedance/F-16](https://github.com/bytedance/F-16).

Despite the superior capability in complementary information exploration and consistent clustering structure learning, most current weight-based multi-modal clustering methods still contain three limitations: 1) lack of trustworthiness in learned weights; 2) isolated view weight learning; 3) extra weight parameters. Motivated by the peer-review mechanism in the academia, we in this paper give a new peer-review look on the multi-modal clustering problem and propose to iteratively treat one modality as "author" and the remaining modalities as "reviewers" so as to reach a peer-review score for each modality. It essentially explores the underlying relationships among modalities. To improve the trustworthiness, we further design a new trustworthy score with a self-supervision working mechanism. Following that, we propose a novel Peer-review Trustworthy Information Bottleneck (PTIB) method for weighted multi-modal clustering, where both the above scores are simultaneously taken into account for accurate and parameter-free modality weight learning. Extensive experiments on eight multi-modal datasets suggest that PTIB can outperform the state-of-the-art multi-modal clustering methods.

This work identifies the first privacy-aware stratified sampling scheme that minimizes the variance for general private mean estimation under the Laplace, Discrete Laplace (DLap) and Truncated-Uniform-Laplace (TuLap) mechanisms within the framework of differential privacy (DP). We view stratified sampling as a subsampling operation, which amplifies the privacy guarantee; however, to have the same final privacy guarantee for each group, different nominal privacy budgets need to be used depending on the subsampling rate. Ignoring the effect of DP, traditional stratified sampling strategies risk significant variance inflation. We phrase our optimal survey design as an optimization problem, where we determine the optimal subsampling sizes for each group with the goal of minimizing the variance of the resulting estimator. We establish strong convexity of the variance objective, propose an efficient algorithm to identify the integer-optimal design, and offer insights on the structure of the optimal design.

LLM-based multi-agent systems (MAS) have shown significant potential in tackling diverse tasks. However, to design effective MAS, existing approaches heavily rely on manual configurations or multiple calls of advanced LLMs, resulting in inadaptability and high inference costs. In this paper, we simplify the process of building an MAS by reframing it as a generative language task, where the input is a user query and the output is a corresponding MAS. To address this novel task, we unify the representation of MAS as executable code and propose a consistency-oriented data construction pipeline to create a high-quality dataset comprising coherent and consistent query-MAS pairs. Using this dataset, we train MAS-GPT, an open-source medium-sized LLM that is capable of generating query-adaptive MAS within a single LLM inference. The generated MAS can be seamlessly applied to process user queries and deliver high-quality responses. Extensive experiments on 9 benchmarks and 5 LLMs show that the proposed MAS-GPT consistently outperforms 10+ baseline MAS methods on diverse settings, indicating MAS-GPT's high effectiveness, efficiency and strong generalization ability. The codes are released at \url{https://github.com/rui-ye/MAS-GPT}.

While large-scale training data is fundamental for developing capable large language models (LLMs), strategically selecting high-quality data has emerged as a critical approach to enhance training efficiency and reduce computational costs. Current data selection methodologies predominantly rely on static, training-agnostic criteria, failing to account for the dynamic model training and data interactions. In this paper, we propose a new Data Weighting Model (DWM) to adjust the weight of selected data within each batch to achieve a dynamic data utilization during LLM training. Specially, to better capture the dynamic data preference of the trained model, a bi-level optimization framework is implemented to update the weighting model. Our experiments demonstrate that DWM enhances the performance of models trained with randomly-selected data, and the learned weighting model can be transferred to enhance other data selection methods and models of different sizes. Moreover, we further analyze how a model’s data preferences evolve throughout training, providing new insights into the data preference of the model during training.

In this paper, we address the problem of identifying the Pareto Set under feasibility constraints in a multivariate bandit setting. Specifically, given a $K$-armed bandit with unknown means $\mu_1, \dots, \mu_K \in \mathbb{R}^d$, the goal is to identify the set of arms whose mean is not uniformly worse than that of another arm (i.e., not smaller for all objectives), while satisfying some known set of linear constraints, expressing, for example, some minimal performance on each objective. Our focus lies in fixed-confidence identification, for which we introduce an algorithm that significantly outperforms racing-like algorithms and the intuitive two-stage approach that first identifies feasible arms and then their Pareto Set. We further prove an information-theoretic lower bound on the sample complexity of any algorithm for constrained Pareto Set identification, showing that the sample complexity of our approach is near-optimal. Our theoretical results are supported by an extensive empirical evaluation on a series of benchmarks.

We introduce PDE-Transformer, an improved transformer-based architecture for surrogate modeling of physics simulations on regular grids. We combine recent architectural improvements of diffusion transformers with adjustments specific for large-scale simulations to yield a more scalable and versatile general-purpose transformer architecture, which can be used as the backbone for building large-scale foundation models in physical sciences. We demonstrate that our proposed architecture outperforms state-of-the-art transformer architectures for computer vision on a large dataset of 16 different types of PDEs. We propose to embed different physical channels individually as spatio-temporal tokens, which interact via channel-wise self-attention. This helps to maintain a consistent information density of tokens when learning multiple types of PDEs simultaneously. We demonstrate that our pre-trained models achieve improved performance on several challenging downstream tasks compared to training from scratch and also beat other foundation model architectures for physics simulations. Our source code is available at https://github.com/tum-pbs/pde-transformer.

The rapid advancement of code large language models (Code LLMs) underscores the critical need for effective and transparent benchmarking methods. However, current benchmarking predominantly relies on publicly available, human-created datasets. The widespread use of these static benchmark datasets makes the evaluation process particularly susceptible to data contamination—an unavoidable consequence of the extensive data collection processes employed during LLM training. Existing methods for addressing data contamination typically face significant limitations, including reliance on substantial human effort and difficulty in managing class imbalances. To overcome these challenges, we propose DyCodeEval, a novel benchmarking suite specifically designed to evaluate Code LLMs under realistic contamination scenarios. Given an initial seed programming problem, DyCodeEval utilizes multiple agents to systematically extract and modify contextual information without changing the core logic, generating semantically equivalent variations. We introduce a dynamic data generation method and conduct extensive empirical studies on two seed datasets involving 18 Code LLMs. The results demonstrate that DyCodeEval effectively assesses the reasoning capabilities of Code LLMs under contamination conditions while producing diverse problem variants, thereby ensuring robust and consistent benchmarking outcomes.

Software engineering activities frequently involve edits to existing code. However, contemporary code language models (LMs) lack the ability to handle diverse types of code-edit requirements. In this work, we attempt to overcome this shortcoming through (1) a novel synthetic data generation pipeline and (2) a robust model adaptation algorithm. Starting with seed code examples and diverse editing criteria, our pipeline generates high-quality samples comprising original and modified code, along with natural language instructions in different styles and verbosity. Today's code LMs come bundled with strong abilities, such as code generation and instruction following, which should not be lost due to fine-tuning. To ensure this, we propose a novel adaptation algorithm, SeleKT, that (a) leverages a dense gradient-based step to identify the weights that are most important for code editing, and (b) does a sparse projection onto the base model to avoid overfitting. Using our approach, we obtain a new series of models NextCoder (adapted from QwenCoder-2.5) that achieves strong results on five code-editing benchmarks, outperforming comparable size models and even several larger ones. We show the generality of our approach on two model families DeepSeekCoder and QwenCoder), compare against other fine-tuning approaches, and demonstrate robustness by showing retention of code generation and general problem-solving abilities post adaptation. We opensource the models, synthetic dataset, and implementation at http://aka.ms/nextcoder.

This paper investigates causal effect identification in latent variable Linear Non-Gaussian Acyclic Models (lvLiNGAM) using higher-order cumulants, addressing two prominent setups that are challenging in the presence of latent confounding: (1) a single proxy variable that may causally influence the treatment and (2) underspecified instrumental variable cases where fewer instruments exist than treatments. We prove that causal effects are identifiable with a single proxy or instrument and provide corresponding estimation methods. Experimental results demonstrate the accuracy and robustness of our approaches compared to existing methods, advancing the theoretical and practical understanding of causal inference in linear systems with latent confounders.

Subgraphs of a complete graph are usually distributed across multiple devices and can only be accessed locally because the raw data cannot be directly shared. However, existing node-level federated graph learning suffers from at least one of the following issues: 1) heavily relying on labeled graph samples that are difficult to obtain in real-world applications, and 2) partitioning a complete graph into several subgraphs inevitably causes missing links, leading to sub-optimal sample representations. To solve these issues, we propose a novel $\underline{\text{Fed}}$erated $\underline{\text{N}}$ode-level $\underline{\text{C}}$lustering $\underline{\text{N}}$etwork (FedNCN), which mends the destroyed cross-subgraph links using clustering prior knowledge. Specifically, within each client, we first design an MLP-based projector to implicitly preserve key clustering properties of a subgraph in a denoising learning-like manner, and then upload the resultant clustering signals that are hard to reconstruct for subsequent cross-subgraph links restoration. In the server, we maximize the potential affinity between subgraphs stemming from clustering signals by graph similarity estimation and minimize redundant links via the N-Cut criterion. Moreover, we employ a GNN-based generator to learn consensus prototypes from this mended graph, enabling the MLP-GNN joint-optimized learner to enhance data privacy during data transmission and further promote the local model for better clustering. Extensive experiments demonstrate the superiority of FedNCN.

Distributional Compositional Generalization (DCG) refers to the ability to tackle tasks from new distributions by leveraging the knowledge of concepts learned from supporting distributions. In this work, we aim to explore the statistical mechanisms of DCG, which have been largely overlooked in previous studies. By statistically formulating the problem, this paper seeks to address two key research questions: 1) Can a method to one DCG problem be applicable to another? 2) What statistical properties can indicate a learning algorithm's capacity for knowledge composition in DCG tasks? \textbf{To address the first question}, an invariant measure is proposed to provide a dimension where all different methods converge. This measure underscores the critical role of data in enabling improvements without trade-offs. \textbf{As for the second question}, we reveal that by decoupling the impacts of insufficient data and knowledge composition, the ability of the learning algorithm to compose knowledge relies on the compatibility and sensitivity between the learning algorithm and the composition rule. In summary, the statistical analysis of the generalization mechanisms provided in this paper deepens our understanding of compositional generalization, offering a complementary evidence on the importance of data in DCG task.

We revisit the recently introduced Local Glivenko-Cantelli setting, which studies distribution-dependent uniform convergence rates of the Empirical Mean Estimator (EME). In this work, we investigate generalizations of this setting where arbitrary estimators are allowed rather than just the EME. Can a strictly larger class of measures be learned? Can better risk decay rates be obtained? We provide exhaustive answers to these questions—which are both negative, provided the learner is barred from exploiting some infinite-dimensional pathologies. On the other hand, allowing such exploits does lead to a strictly larger class of learnable measures.

Mathematical optimization is fundamental to decision-making across diverse domains, from operations research to healthcare. Yet, translating real-world problems into optimization models remains a difficult task, often demanding specialized expertise. This paper approaches the problem of $\textit{autoformulation}$: the automated creation of solver-ready optimization models from natural language problem descriptions. We identify three core challenges of autoformulation: $\textit{(1)}$ the vast, problem-dependent hypothesis space, $\textit{(2)}$ efficient and diverse exploration of this space under uncertainty, and $\textit{(3)}$ evaluation of formulation correctness against problem description. To address these challenges, we present a novel method leveraging $\textit{Large Language Models}$ (LLMs) with $\textit{Monte-Carlo Tree Search}$, exploiting the hierarchical nature of optimization modeling to generate and systematically explore possible formulations. To enhance search efficiency, we introduce symbolic pruning to eliminate trivially equivalent search paths (branches), and employ LLM-based evaluation of partial formulations to guide search. Empirical analysis on linear and mixed-integer programming benchmarks demonstrates our method's effectiveness, with significant performance gains from both LLM-based value estimation and symbolic pruning techniques.

We study the policy evaluation problem in an online multi-reward multi-policy discounted setting, where multiple reward functions must be evaluated simultaneously for different policies. We adopt an $(\epsilon,\delta)$-PAC perspective to achieve $\epsilon$-accurate estimates with high confidence across finite or convex sets of rewards, a setting that has not been investigated in the literature. Building on prior work on Multi-Reward Best Policy Identification, we adapt the MR-NaS exploration scheme to jointly minimize sample complexity for evaluating different policies across different reward sets. Our approach leverages an instance-specific lower bound revealing how the sample complexity scales with a measure of value deviation, guiding the design of an efficient exploration policy. Although computing this bound entails a hard non-convex optimization, we propose an efficient convex approximation that holds for both finite and convex reward sets. Experiments in tabular domains demonstrate the effectiveness of this adaptive exploration scheme.