ICML 2025 Papers

Generating long-term trajectories of dissipative chaotic systems autoregressively is a highly challenging task. The inherent positive Lyapunov exponents amplify prediction errors over time. Many chaotic systems possess a crucial property — ergodicity on their attractors, which makes long-term prediction possible. State-of-the-art methods address ergodicity by preserving statistical properties using optimal transport techniques. However, these methods face scalability challenges due to the curse of dimensionality when matching distributions. To overcome this bottleneck, we propose a scalable transformer-based framework capable of stably generating long-term high-dimensional and high-resolution chaotic dynamics while preserving ergodicity. Our method is grounded in a physical perspective, revisiting the Von Neumann mean ergodic theorem to ensure the preservation of long-term statistics in the $\mathcal{L}^2$ space. We introduce novel modifications to the attention mechanism, making the transformer architecture well-suited for learning large-scale chaotic systems. Compared to operator-based and transformer-based methods, our model achieves better performances across five metrics, from short-term prediction accuracy to long-term statistics. In addition to our methodological contributions, we introduce new chaotic system benchmarks: a machine learning dataset of 140$k$ snapshots of turbulent channel flow and a processed high-dimensional Kolmogorov Flow dataset, along with various evaluation metrics for both short- and long-term performances. Both are well-suited for machine learning research on chaotic systems.

Numerous studies have assessed the proficiency of AI systems, particularly large language models (LLMs), in facilitating everyday tasks such as email writing, question answering, and creative content generation. However, researchers face unique challenges and opportunities in leveraging LLMs for their own work, such as brainstorming research ideas, designing experiments, and writing or reviewing papers. In this study, we introduce AAAR-1.0, a benchmark dataset designed to evaluate LLM performance in three fundamental, expertise-intensive research tasks: (i) EquationInference, assessing the correctness of equations based on the contextual information in paper submissions; (ii) ExperimentDesign, designing experiments to validate research ideas and solutions; and (iii) PaperWeakness, identifying weaknesses in paper submissions. AAAR-1.0 differs from prior benchmarks in two key ways: first, it is explicitly research-oriented, with tasks requiring deep domain expertise; second, it is researcher-oriented, mirroring the primary activities that researchers engage in on a daily basis. An evaluation of both open-source and proprietary LLMs reveals their potential as well as limitations in conducting sophisticated research tasks. We will release the AAAR-1.0 and keep iterating it to new versions.

Multimodal large language models (MLLMs) excel at generating highly detailed captions but often produce hallucinations. Our analysis reveals that existing hallucination detection methods struggle with detailed captions. We attribute this to the increasing reliance of MLLMs on their generated text, rather than the input image, as the sequence length grows. To address this issue, we propose a multiagent approach that leverages LLM-MLLM collaboration to correct given captions. Additionally, we introduce an evaluation framework and a benchmark dataset to facilitate the systematic analysis of detailed captions. Our experiments demonstrate that the proposed evaluation method aligns better with human judgments of factuality than existing metrics. Moreover, we show that current approaches for enhancing MLLM factuality often fail in hyper-detailed image captioning tasks. In contrast, our approach significantly enhances the factual accuracy of captions, even improving those generated by GPT-4V. Finally, we highlight a limitation of VQA-centric benchmarking by demonstrating that an MLLM's performance on VQA benchmarks may not correlate with its ability to generate detailed image captions.

Large language models (LLMs) have shown great potential as general-purpose AI assistants across various domains. To fully leverage this potential in specific applications, many companies provide fine-tuning API services, enabling users to upload their own data for LLM customization. However, fine-tuning services introduce a new safety threat: user-uploaded data, whether harmful or benign, can break the model’s alignment, leading to unsafe outputs. Moreover, existing defense methods struggle to address the diversity of fine-tuning datasets (e.g., varying sizes, tasks), often sacrificing utility for safety or vice versa. To address this issue, we propose Safe Delta, a safety-aware post-training defense method that adjusts the delta parameters (i.e., the parameter change before and after fine-tuning). Specifically, Safe Delta estimates the safety degradation, selects delta parameters to maximize utility while limiting overall safety loss, and applies a safety compensation vector to mitigate residual safety loss. Through extensive experiments on four diverse datasets with varying settings, our approach consistently preserves safety while ensuring that the utility gain from benign datasets remains unaffected.

Both PAC-Bayesian and Sample Compress learning frameworks have been shown instrumental for deriving tight (non-vacuous) generalization bounds for neural networks. We leverage these results in a meta-learning scheme, relying on a hypernetwork that outputs the parameters of a downstream predictor from a dataset input. The originality of our approach lies in the investigated hypernetwork architectures that encode the dataset before decoding the parameters: (1) a PAC-Bayesian encoder that expresses a posterior distribution over a latent space, (2) a Sample Compress encoder that selects a small sample of the dataset input along with a message from a discrete set, and (3) a hybrid between both approaches motivated by a new Sample Compress theorem handling continuous messages. The latter theorem exploits the pivotal information transiting at the encoder-decoder junction in order to compute generalization guarantees for each downstream predictor obtained by our meta-learning scheme.

Message-Passing Neural Networks (MPNNs) have become a cornerstone for processing and analyzing graph-structured data. However, their effectiveness is often hindered by phenomena such as over-squashing, where long-range dependencies or interactions are inadequately captured and expressed in the MPNN output. This limitation mirrors the challenges of the Effective Receptive Field (ERF) in Convolutional Neural Networks (CNNs), where the theoretical receptive field is underutilized in practice. In this work, we show and theoretically explain the limited ERF problem in MPNNs. Furthermore, inspired by recent advances in ERF augmentation for CNNs, we propose an Interleaved Multiscale Message-Passing Neural Networks (IM-MPNN) architecture to address these problems in MPNNs. Our method incorporates a hierarchical coarsening of the graph, enabling message-passing across multiscale representations and facilitating long-range interactions without excessive depth or parameterization. Through extensive evaluations on benchmarks such as the Long-Range Graph Benchmark (LRGB), we demonstrate substantial improvements over baseline MPNNs in capturing long-range dependencies while maintaining computational efficiency.

Heterogeneity in sensors and actuators across environments poses a significant challenge to building large-scale pre-trained world models on top of this low-dimensional sensor information. In this work, we explore pre-training world models for heterogeneous environments by addressing key transfer barriers in both data diversity and model flexibility. We introduce UniTraj, a unified dataset comprising over one million trajectories from 80 environments, designed to scale data while preserving critical diversity. Additionally, we propose TrajWorld, a novel architecture capable of flexibly handling varying sensor and actuator information and capturing environment dynamics in-context. Pre-training TrajWorld on UniTraj yields substantial gains in transition prediction, achieves a new state-of-the-art for off-policy evaluation, and also delivers superior online performance of model predictive control. To the best of our knowledge, this work, for the first time, demonstrates the transfer benefits of world models across heterogeneous and complex control environments. Code and data are available at https://github.com/thuml/TrajWorld.

Sparse variational Gaussian processes (GPs) construct tractable posterior approximations to GP models. At the core of these methods is the assumption that the true posterior distribution over training function values ${\bf f}$ and inducing variables ${\bf u}$ is approximated by a variational distribution that incorporates the conditional GP prior $p({\bf f} | {\bf u})$ in its factorization. While this assumption is considered as fundamental, we show that for model training we can relax it through the use of a more general variational distribution $q({\bf f} | {\bf u} )$ that depends on $N$ extra parameters, where $N$ is the number of training examples. In GP regression, we can analytically optimize the evidence lower bound over the extra parameters and express a tractable collapsed bound that is tighter than the previous bound. The new bound is also amenable to stochastic optimization and its implementation requires minor modifications to existing sparse GP code. Further, we also describe extensions to non-Gaussian likelihoods. On several datasets we demonstrate that our method can reduce bias when learning the hyperparameters and can lead to better predictive performance.

Quantum Error Mitigation (QEM) has emerged as a pivotal technique for enhancing the reliability of noisy quantum devices in the *Noisy Intermediate-Scale Quantum* (NISQ) era. Recently, machine learning (ML)-based QEM approaches have demonstrated strong generalization capabilities without sampling overheads compared to conventional methods. However, evaluating these techniques is often hindered by a lack of standardized datasets and inconsistent experimental settings across different studies. In this work, we present **QEM-Bench**, a comprehensive benchmark suite of *twenty-two* datasets covering diverse circuit types and noise profiles, which provides a unified platform for comparing and advancing ML-based QEM methods. We further propose a refined ML-based QEM pipeline **QEMFormer**, which leverages a feature encoder that preserves local, global, and topological information, along with a two-branch model that captures short-range and long-range dependencies within the circuit. Empirical evaluations on QEM-Bench illustrate the superior performance of QEMFormer over existing baselines, underscoring the potential of integrated ML-QEM strategies.

Recent deep learning models demand larger datasets, driving the need for dataset distillation to create compact, cost-efficient datasets while maintaining performance. Due to the powerful image generation capability of diffusion, it has been introduced to this field for generating distilled images. In this paper, we systematically investigate issues present in current diffusion-based dataset distillation methods, including inaccurate distribution matching, distribution deviation with random noise, and separate sampling. Building on this, we propose D$^3$HR, a novel diffusion-based framework to generate distilled datasets with high representativeness. Specifically, we adopt DDIM inversion to map the latents of the full dataset from a low-normality latent domain to a high-normality Gaussian domain, preserving information and ensuring structural consistency to generate representative latents for the distilled dataset. Furthermore, we propose an efficient sampling scheme to better align the representative latents with the high-normality Gaussian distribution. Our comprehensive experiments demonstrate that D$^3$HR can achieve higher accuracy across different model architectures compared with state-of-the-art baselines in dataset distillation. Source code: https://github.com/lin-zhao-resoLve/D3HR.

Spectral Graph Neural Networks process graph signals using the spectral properties of the normalized graph Laplacian matrix. However, the frequent occurrence of repeated eigenvalues limits the expressiveness of spectral GNNs. To address this, we propose a higher-dimensional sheaf Laplacian matrix, which not only encodes the graph's topological information but also increases the upper bound on the number of distinct eigenvalues. The sheaf Laplacian matrix is derived from carefully designed perturbations of the block form of the normalized graph Laplacian, yielding a perturbed sheaf Laplacian (PSL) matrix with more distinct eigenvalues. We provide a theoretical analysis of the expressiveness of spectral GNNs equipped with the PSL and establish perturbation bounds for the eigenvalues. Extensive experiments on benchmark datasets for node classification demonstrate that incorporating the perturbed sheaf Laplacian enhances the performance of spectral GNNs.

This paper studies the interplay between learning algorithms and graph structure for graph neural networks (GNNs). Existing theoretical studies on the learning dynamics of GNNs primarily focus on the convergence rates of learning algorithms under the interpolation regime (noise-free) and offer only a crude connection between these dynamics and the actual graph structure (e.g., maximum degree). This paper aims to bridge this gap by investigating the excessive risk (generalization performance) of learning algorithms in GNNs within the generalization regime (with noise). Specifically, we extend the conventional settings from the learning theory literature to the context of GNNs and examine how graph structure influences the performance of learning algorithms such as stochastic gradient descent (SGD) and Ridge regression. Our study makes several key contributions toward understanding the interplay between graph structure and learning in GNNs. First, we derive the excess risk profiles of SGD and Ridge regression in GNNs and connect these profiles to the graph structure through spectral graph theory. With this established framework, we further explore how different graph structures (regular vs. power-law) impact the performance of these algorithms through comparative analysis. Additionally, we extend our analysis to multi-layer linear GNNs, revealing an increasing non-isotropic effect on the excess risk profile, thereby offering new insights into the over-smoothing issue in GNNs from the perspective of learning algorithms. Our empirical results align with our theoretical predictions, \emph{collectively showcasing a coupling relation among graph structure, GNNs and learning algorithms, and providing insights on GNN algorithm design and selection in practice.}

Dataset distillation has emerged as an effective strategy, significantly reducing training costs and facilitating more efficient model deployment. Recent advances have leveraged generative models to distill datasets by capturing the underlying data distribution. Unfortunately, existing methods require model fine-tuning with distillation losses to encourage diversity and representativeness. However, these methods do not guarantee sample diversity, limiting their performance. We propose a mode-guided diffusion model leveraging a pre-trained diffusion model without the need to fine-tune with distillation losses. Our approach addresses dataset diversity in three stages: Mode Discovery to identify distinct data modes, Mode Guidance to enhance intra-class diversity, and Stop Guidance to mitigate artifacts in synthetic samples that affect performance. We evaluate our approach on ImageNette, ImageIDC, ImageNet-100, and ImageNet-1K, achieving accuracy improvements of 4.4%, 2.9%, 1.6%, and 1.6%, respectively, over state-of-the-art methods. Our method eliminates the need for fine-tuning diffusion models with distillation losses, significantly reducing computational costs.

As AI systems continue to evolve, their rigorous evaluation becomes crucial for their development and deployment. Researchers have constructed various large-scale benchmarks to determine their capabilities, typically against a gold-standard test set and report metrics averaged across all items. However, this static evaluation paradigm increasingly shows its limitations, including high evaluation costs, data contamination, and the impact of low-quality or erroneous items on evaluation reliability and efficiency. In this Position, drawing from human psychometrics, we discuss a paradigm shift from static evaluation methods to adaptive testing. This involves estimating the characteristics or value of each test item in the benchmark, and tailoring each model's evaluation instead of relying on a fixed test set. This paradigm provides robust ability estimation, uncovering the latent traits underlying a model’s observed scores. This position paper analyze the current possibilities, prospects, and reasons for adopting psychometrics in AI evaluation. We argue that *psychometrics, a theory originating in the 20th century for human assessment, could be a powerful solution to the challenges in today's AI evaluations*.

Synthetic videos nowadays is widely used to complement data scarcity and diversity of real-world videos. Current synthetic datasets primarily replicate real-world scenarios, leaving impossible, counterfactual and anti-reality video concepts underexplored. This work aims to answer two questions: 1) Can today's video generation models effectively follow prompts to create impossible video content? 2) Are today's video understanding models good enough for understanding impossible videos? To this end, we introduce *IPV-Bench*, a novel benchmark designed to evaluate and foster progress in video understanding and generation. *IPV-Bench* is underpinned by a comprehensive taxonomy, encompassing 4 domains, 14 categories. It features diverse scenes that defy physical, biological, geographical, or social laws. Based on the taxonomy, a prompt suite is constructed to evaluate video generation models, challenging their prompt following and creativity capabilities. In addition, a video benchmark is curated to assess Video-LLMs on their ability of understanding impossible videos, which particularly requires reasoning on temporal dynamics and world knowledge. Comprehensive evaluations reveal limitations and insights for future directions of video models, paving the way for next-generation video models.

Parameter-efficient fine-tuning (PEFT) methods, such as LoRA, are commonly used to adapt LLMs. However, the effectiveness of standard PEFT methods is limited in low-resource scenarios with only a few hundred examples. Recent advances in interpretability research have inspired the emergence of activation editing (or steering) techniques, which modify the activations of specific model components. Due to their extremely small parameter counts, these methods show promise for small datasets. However, their performance is highly dependent on identifying the correct modules to edit and often lacks stability across different datasets. In this paper, we propose Joint Localization and Activation Editing (JoLA), a method that jointly learns (1) which heads in the Transformer to edit (2) whether the intervention should be additive, multiplicative, or both and (3) the intervention parameters themselves - the vectors applied as additive offsets or multiplicative scalings to the head output. Through evaluations on three benchmarks spanning commonsense reasoning, natural language understanding, and natural language generation, we demonstrate that JoLA consistently outperforms existing methods.

Benchmarks and evaluations are central to machine learning methodology and direct research in the field. Current evaluations commonly test systems in the absence of humans. This position paper argues that the machine learning community should increasingly use _centaur evaluations_, in which humans and AI jointly solve tasks. Centaur Evaluations refocus machine learning development toward human augmentation instead of human replacement, they allow for direct evaluation of human-centered desiderata, such as interpretability and helpfulness, and they can be more challenging and realistic than existing evaluations. By shifting the focus from _automation_ toward _collaboration_ between humans and AI, centaur evaluations can drive progress toward more effective and human-augmenting machine learning systems.

We study universal properties in real-world complex and synthetically generated datasets. Our approach is to analogize data to a physical system and employ tools from statistical physics and Random Matrix Theory (RMT) to reveal their underlying structure. Examining the local and global eigenvalue statistics of feature-feature covariance matrices, we find: (i) bulk eigenvalue power-law scaling vastly differs between uncorrelated Gaussian and real-world data, (ii) this power law behavior is reproducible using Gaussian data with long-range correlations, (iii) all dataset types exhibit chaotic RMT universality, (iv) RMT statistics emerge at smaller dataset sizes than typical training sets, correlating with power-law convergence, (v) Shannon entropy correlates with RMT structure and requires fewer samples in strongly correlated datasets. These results suggest natural image Gram matrices can be approximated by Wishart random matrices with simple covariance structure, enabling rigorous analysis of neural network behavior.

In continual learning scenarios, catastrophic forgetting of previously learned tasks is a critical issue, making it essential to effectively measure such forgetting. Recently, there has been growing interest in focusing on representation forgetting, the forgetting measured at the hidden layer. In this paper, we provide the first theoretical analysis of representation forgetting and use this analysis to better understand the behavior of continual learning. First, we introduce a new metric called representation discrepancy, which measures the difference between representation spaces constructed by two snapshots of a model trained through continual learning. We demonstrate that our proposed metric serves as an effective surrogate for the representation forgetting while remaining analytically tractable. Second, through mathematical analysis of our metric, we derive several key findings about the dynamics of representation forgetting: the forgetting occurs more rapidly to a higher degree as the layer index increases, while increasing the width of the network slows down the forgetting process. Third, we support our theoretical findings through experiments on real image datasets, including Split-CIFAR100 and ImageNet1K.