ICML 2025 Papers

Multimodal Large Language Models (MLLMs) for tabular understanding have made significant progress in tasks such as financial report analysis and public data tests. However, our comprehensive analysis shows that these models are still limited in certain simple scenarios, particularly when handling compositional conditions in QA. Further investigation reveals that the poor performance can be attributed to two main challenges: the visual encoder's inability to accurately recognize the content of a row, and the model's tendency to overlook conditions in the question. To address these, we introduce a new Compositional Condition Tabular Understanding method, called {\sc CoCoTab}. Specifically, to capture the structural relationships within tables, we enhance the visual encoder with additional row and column patches. Moreover, we introduce the conditional tokens between the visual patches and query embeddings, ensuring the model focuses on relevant parts of the table according to the conditions specified in the query. Additionally, we also introduce the Massive Multimodal Tabular Understanding (MMTU) benchmark, which comprehensively assesses the full capabilities of MLLMs in tabular understanding. Our proposed method achieves state-of-the-art performance on both existing tabular understanding benchmarks and MMTU. Our code can be available at \url{https://github.com/LAMDA-Tabular/MMTU}.

We present TUMTraf VideoQA, a novel dataset and benchmark designed for spatio-temporal video understanding in complex roadside traffic scenarios. The dataset comprises 1,000 videos, featuring 85,000 multiple-choice QA pairs, 2,300 object captioning, and 5,700 object grounding annotations, encompassing diverse real-world conditions such as adverse weather and traffic anomalies. By incorporating tuple-based spatio-temporal object expressions, TUMTraf VideoQA unifies three essential tasks—multiple-choice video question answering, referred object captioning, and spatio-temporal object grounding—within a cohesive evaluation framework. We further introduce the TraffiX-Qwen baseline model, enhanced with visual token sampling strategies, providing valuable insights into the challenges of fine-grained spatio-temporal reasoning. Extensive experiments demonstrate the dataset’s complexity, highlight the limitations of existing models, and position TUMTraf VideoQA as a robust foundation for advancing research in intelligent transportation systems. The dataset and benchmark are publicly available to facilitate further exploration.

We examine the problem of learning sequential tasks from a single visual demonstration. A key challenge arises when demonstrations are temporally misaligned due to variations in timing, differences in embodiment, or inconsistencies in execution. Existing approaches treat imitation as a distribution-matching problem, aligning individual frames between the agent and the demonstration. However, we show that such frame-level matching fails to enforce temporal ordering or ensure consistent progress. Our key insight is that matching should instead be defined at the level of sequences. We propose that perfect matching occurs when one sequence successfully covers all the subgoals in the same order as the other sequence. We present ORCA (ORdered Coverage Alignment), a dense per-timestep reward function that measures the probability of the agent covering demonstration frames in the correct order. On temporally misaligned demonstrations, we show that agents trained with the ORCA reward achieve $4.5$x improvement ($0.11 \rightarrow 0.50$ average normalized returns) for Meta-world tasks and $6.6$x improvement ($6.55 \rightarrow 43.3$ average returns) for Humanoid-v4 tasks compared to the best frame-level matching algorithms. We also provide empirical analysis showing that ORCA is robust to varying levels of temporal misalignment. The project website is at https://portal-cornell.github.io/orca/

Multimodal Large Language Models (MLLMs) have shown promising progress in understanding and analyzing video content. However, processing long videos remains a significant challenge constrained by LLM's context size. To address this limitation, we propose \textbf{LongVU}, a spatiotemporal adaptive compression mechanism that reduces the number of video tokens while preserving visual details of long videos. Our idea is based on leveraging cross-modal query and inter-frame dependencies to adaptively reduce temporal and spatial redundancy in videos. Specifically, we leverage DINOv2 features to remove redundant frames that exhibit high similarity. Then we utilize text-guided cross-modal query for selective frame feature reduction. Further, we perform spatial token reduction across frames based on their temporal dependencies. Our adaptive compression strategy effectively processes a large number of frames with little visual information loss within given context length. Our LongVU consistently surpass existing methods across a variety of video understanding benchmarks, especially on hour-long video understanding tasks such as VideoMME and MLVU. Given a light-weight LLM, our LongVU also scales effectively into a smaller size with state-of-the-art video understanding performance.

When training deep neural networks with gradient descent, sharpness often increases---a phenomenon known as *progressive sharpening*---before saturating at the *edge of stability*. Although commonly observed in practice, the underlying mechanisms behind progressive sharpening remain poorly understood. In this work, we study this phenomenon using a minimalist model: a deep linear network with a single neuron per layer. We show that this simple model effectively captures the sharpness dynamics observed in recent empirical studies, offering a simple testbed to better understand neural network training. Moreover, we theoretically analyze how dataset properties, network depth, stochasticity of optimizers, and step size affect the degree of progressive sharpening in the minimalist model. We then empirically demonstrate how these theoretical insights extend to practical scenarios. This study offers a deeper understanding of sharpness dynamics in neural network training, highlighting the interplay between depth, training data, and optimizers.

Continual learning (CL) focuses on the ability of models to learn sequentially from a stream of tasks. A major challenge in CL is catastrophic forgetting (CF). CF is a phenomenon where the model experiences significant performance degradation on previously learned tasks after training on new tasks. Although CF is commonly observed in convolutional neural networks (CNNs), the theoretical understanding about CF within CNNs remains limited. To fill the gap, we present a theoretical analysis of CF in a two-layer CNN. By employing a multi-view data model, we analyze the learning dynamics of different features throughout CL and derive theoretical insights. The findings are supported by empirical results from both simulated and real-world datasets.

Predictive posterior densities (PPDs) are essential in approximate inference for quantifying predictive uncertainty and comparing inference methods. Typically, PPDs are estimated by simple Monte Carlo (MC) averages. In this paper, we expose a critical under-recognized issue: the signal-to-noise ratio (SNR) of the simple MC estimator can sometimes be extremely low, leading to unreliable estimates. Our main contribution is a theoretical analysis demonstrating that even with exact inference, SNR can decay rapidly with an increase in (a) the mismatch between training and test data, (b) the dimensionality of the latent space, or (c) the size of test data relative to training data. Through several examples, we empirically verify these claims and show that these factors indeed lead to poor SNR and unreliable PPD estimates (sometimes, estimates are off by hundreds of nats even with a million samples). While not the primary focus, we also explore an adaptive importance sampling approach as an illustrative way to mitigate the problem, where we learn the proposal distribution by maximizing a variational proxy to the SNR. Taken together, our findings highlight an important challenge and provide essential insights for reliable estimation.

The success of sequential decision-making approaches, such as *reinforcement learning* (RL), is closely tied to the availability of a reward feedback. However, designing a reward function that encodes the desired objective is a challenging task. In this work, we address a more realistic scenario: sequential decision making with preference feedback provided, for instance, by a human expert. We aim to build a theoretical basis linking *preferences*, (non-Markovian) *utilities*, and (Markovian) *rewards*, and we study the connections between them. First, we model preference feedback using a partial (pre)order over trajectories, enabling the presence of incomparabilities that are common when preferences are provided by humans but are surprisingly overlooked in existing works. Second, to provide a theoretical justification for a common practice, we investigate how a preference relation can be approximated by a multi-objective utility. We introduce a notion of preference-utility compatibility and analyze the computational complexity of this transformation, showing that constructing the minimum-dimensional utility is NP-hard. Third, we propose a novel concept of preference-based policy dominance that does not rely on utilities or rewards and discuss the computational complexity of assessing it. Fourth, we develop a computationally efficient algorithm to approximate a utility using (Markovian) rewards and quantify the error in terms of the suboptimality of the optimal policy induced by the approximating reward. This work aims to lay the foundation for a principled approach to sequential decision making from preference feedback, with promising potential applications in RL from human feedback.

Computer Vision (CV) has yet to fully achieve the zero-shot task generalization observed in Natural Language Processing (NLP), despite following many of the milestones established in NLP, such as large transformer models, extensive pre-training, and the auto-regression paradigm, among others. In this paper, we rethink the reality that CV adopts discrete and terminological task definitions (e.g., "image segmentation"), and conjecture it is a key barrier that hampers zero-shot task generalization. Our hypothesis is that without truly understanding previously-seen tasks—due to these terminological definitions—deep models struggle to generalize to novel tasks. To verify this, we introduce Explanatory Instructions, which provide an intuitive way to define CV task objectives through detailed linguistic transformations from input images to outputs. We create a large-scale dataset comprising 12 million "image input $\to$ explanatory instruction $\to$ output" triplets, and train an auto-regressive-based vision-language model (AR-based VLM) that takes both images and explanatory instructions as input. By learning to follow these instructions, the AR-based VLM achieves instruction-level zero-shot capabilities for previously-seen tasks and demonstrates strong zero-shot generalization for unseen CV tasks. Code and dataset will be open-sourced.

Predicting future scene representations is a crucial task for enabling robots to understand and interact with the environment. However, most existing methods rely on videos and simulations with precise action annotations, limiting their ability to leverage the large amount of avail- able unlabeled video data. To address this challenge, we propose PlaySlot, an object-centric video prediction model that infers object representations and latent actions from unlabeled video sequences. It then uses these representations to forecast future object states and video frames. PlaySlot allows the generation of multiple possible futures conditioned on latent actions, which can be inferred from video dynamics, provided by a user, or generated by a learned action policy, thus enabling versatile and interpretable world modeling. Our results show that PlaySlot outperforms both stochastic and object-centric baselines for video prediction across different environments. Furthermore, we show that our inferred latent actions can be used to learn robot behaviors sample-efficiently from unlabeled video demonstrations. Videos and code are available on our project website.

In this paper, we investigate the properties of the Sliced Wasserstein Distance (SW) when employed as an objective functional. The SW metric has gained significant interest in the optimal transport and machine learning literature, due to its ability to capture intricate geometric properties of probability distributions while remaining computationally tractable, making it a valuable tool for various applications, including generative modeling and domain adaptation. Our study aims to provide a rigorous analysis of the critical points arising from the optimization of the SW objective. By computing explicit perturbations, we establish that stable critical points of SW cannot concentrate on segments. This stability analysis is crucial for understanding the behaviour of optimization algorithms for models trained using the SW objective. Furthermore, we investigate the properties of the SW objective, shedding light on the existence and convergence behavior of critical points. We illustrate our theoretical results through numerical experiments.

Large Video Models (LVMs) built upon Large Language Models (LLMs) have shown promise in video understanding but often suffer from misalignment with human intuition and video hallucination issues. To address these challenges, we introduce **VistaDPO**, a novel framework for Video Hierarchical Spatial-Temporal Direct Preference Optimization. VistaDPO enhances text-video preference alignment across three hierarchical levels: i) **Instance Level**, aligning overall video content with responses; ii) **Temporal Level**, aligning video temporal semantics with event descriptions; and iii) **Perceptive Level**, aligning spatial objects with language tokens. Given the lack of datasets for fine-grained video-language preference alignment, we construct **VistaDPO-7k**, a dataset of 7.2K QA pairs annotated with chosen and rejected responses, along with spatial-temporal grounding information such as timestamps, keyframes, and bounding boxes. Extensive experiments on benchmarks such as Video Hallucination, Video QA, and Captioning performance tasks demonstrate that VistaDPO significantly improves the performance of existing LVMs, effectively mitigating video-language misalignment and hallucination.

Model reprogramming adapts pretrained models to downstream tasks by modifying only the input and output spaces. *Visual reprogramming* (VR) is one instance for vision tasks that adds a trainable noise pattern (i.e., a visual prompt) to input images to facilitate downstream classification. The existing VR approaches for CLIP train a single visual prompt using all descriptions of different downstream classes. However, the limited learning capacity may result in (1) a failure to capture diverse aspects of the descriptions (e.g., shape, color, and texture), and (2) a possible bias toward less informative attributes that do not help distinguish between classes. In this paper, we introduce a decoupling-and-reweighting framework. Our *decoupled visual prompts* (DVP) are optimized using descriptions grouped by explicit **c**au**se**s (DVP-cse) or unsupervised **cl**u**s**ters (DVP-cls). Then, we integrate the outputs of these visual prompts with a *probabilistic reweighting matrix* (PRM) that measures their contributions to each downstream class. Theoretically, DVP lowers the empirical risk bound. Experimentally, DVP outperforms baselines on average across 11 downstream datasets. Notably, the DVP-PRM integration enables insights into how individual visual prompts influence classification decisions, providing a probabilistic framework for understanding reprogramming.

Personalized federated learning (PFL) offers a flexible framework for aggregating information across distributed clients with heterogeneous data. This work considers a personalized federated learning setting that simultaneously learns global and local models. While purely local training has no communication cost, collaborative learning among the clients can leverage shared knowledge to improve statistical accuracy, presenting an accuracy-communication trade-off in personalized federated learning. However, the theoretical analysis of how personalization quantitatively influences sample and algorithmic efficiency and their inherent trade-off is largely unexplored. This paper makes a contribution towards filling this gap, by providing a quantitative characterization of the personalization degree on the tradeoff. The results further offer theoretical insights for choosing the personalization degree. As a side contribution, we establish the minimax optimality in terms of statistical accuracy for a widely studied PFL formulation. The theoretical result is validated on both synthetic and real-world datasets and its generalizability is verified in a non-convex setting.

Generalized Category Discovery (GCD) aims to identify both known and novel categories in unlabeled data by leveraging knowledge from old classes. However, existing methods are limited to non-graph data; lack theoretical foundations to answer *When and how known classes can help GCD*. We introduce the Graph GCD task; provide the first rigorous theoretical analysis of *parametric GCD*. By quantifying the relationship between old and new classes in the embedding space using the Wasserstein distance W, we derive the first provable GCD loss bound based on W. This analysis highlights two necessary conditions for effective GCD. However, we uncover, through a Pairwise Markov Random Field perspective, that popular graph contrastive learning (GCL) methods inherently violate these conditions. To address this limitation, we propose SWIRL, a novel GCL method for GCD. Experimental results validate our (theoretical) findings and demonstrate SWIRL's effectiveness.

Tabular data generation has attracted significant research interest in recent years, with the tabular diffusion models greatly improving the quality of synthetic data. However, while memorization—where models inadvertently replicate exact or near-identical training data—has been thoroughly investigated in image and text generation, its effects on tabular data remain largely unexplored. In this paper, we conduct the first comprehensive investigation of memorization phenomena in diffusion models for tabular data. Our empirical analysis reveals that memorization appears in tabular diffusion models and increases with larger training epochs. We further examine the influence of factors such as dataset sizes, feature dimensions, and different diffusion models on memorization. Additionally, we provide a theoretical explanation for why memorization occurs in tabular diffusion models. To address this issue, we propose TabCutMix, a simple yet effective data augmentation technique that exchanges randomly selected feature segments between random same-class training sample pairs. Building upon this, we introduce TabCutMixPlus, an enhanced method that clusters features based on feature correlations and ensures that features within the same cluster are exchanged together during augmentation. This clustering mechanism mitigates out-of-distribution (OOD) generation issues by maintaining feature coherence. Experimental results across various datasets and diffusion models demonstrate that TabCutMix effectively mitigates memorization while maintaining high-quality data generation. Our code is available at https://github.com/fangzy96/TabCutMix.

Generative models have demonstrated remarkable abilities in generating high-fidelity visual content. In this work, we explore how generative models can further be used not only to synthesize visual content but also to understand the properties of a scene given a natural image. We formulate scene understanding as an inverse generative modeling problem, where we seek to find conditional parameters of a visual generative model to best fit a given natural image. To enable this procedure to infer scene structure from images substantially different than those seen during training, we further propose to build this visual generative model compositionally from smaller models over pieces of a scene. We illustrate how this procedure enables us to infer the set of objects in a scene, enabling robust generalization to new test scenes with an increased number of objects of new shapes. We further illustrate how this enables us to infer global scene factors, likewise enabling robust generalization to new scenes. Finally, we illustrate how this approach can be directly applied to existing pretrained text-to-image generative models for zero-shot multi-object perception. Code and visualizations are at https://energy-based-model.github.io/compositional-inference.

Weakly supervised video anomaly detection (WS-VAD) is tasked with pinpointing temporal intervals containing anomalous events within untrimmed videos, utilizing only video-level annotations. However, a significant challenge arises due to the absence of dense frame-level annotations, often leading to incomplete localization in existing WS-VAD methods. To address this issue, we present a novel LEC-VAD, Learning Event Completeness for Weakly Supervised Video Anomaly Detection, which features a dual structure designed to encode both category-aware and category-agnostic semantics between vision and language. Within LEC-VAD, we devise semantic regularities that leverage an anomaly-aware Gaussian mixture to learn precise event boundaries, thereby yielding more complete event instances. Besides, we develop a novel memory bank-based prototype learning mechanism to enrich concise text descriptions associated with anomaly-event categories. This innovation bolsters the text's expressiveness, which is crucial for advancing WS-VAD. Our LEC-VAD demonstrates remarkable advancements over the current state-of-the-art methods on two benchmark datasets XD-Violence and UCF-Crime.

We introduce MedXpertQA, a highly challenging and comprehensive benchmark to evaluate expert-level medical knowledge and advanced reasoning. MedXpertQA includes 4,460 questions spanning 17 specialties and 11 body systems. It includes two subsets, Text for text evaluation and MM for multimodal evaluation. Notably, MM introduces expert-level exam questions with diverse images and rich clinical information, including patient records and examination results, setting it apart from traditional medical multimodal benchmarks with simple QA pairs generated from image captions. MedXpertQA applies rigorous filtering and augmentation to address the insufficient difficulty of existing benchmarks like MedQA, and incorporates specialty board questions to improve clinical relevance and comprehensiveness. We perform data synthesis to mitigate data leakage risk and conduct multiple rounds of expert reviews to ensure accuracy and reliability. We evaluate 18 leading models on MedXpertQA. Moreover, medicine is deeply connected to real-world decision-making, providing a rich and representative setting for assessing reasoning abilities beyond mathematics and code. To this end, we develop a reasoning-oriented subset to facilitate the assessment of o1-like models.