ICML 2025 Papers

Training large language models (LLMs) increasingly relies on geographically distributed accelerators, causing prohibitive communication costs across regions and uneven utilization of heterogeneous hardware. We propose HALoS, a hierarchical asynchronous optimization framework that tackles these issues by introducing local parameter servers (LPSs) within each region and a global parameter server (GPS) that merges updates across regions. This hierarchical design minimizes expensive inter-region communication, reduces straggler effects, and leverages fast intra-region links. We provide a rigorous convergence analysis for HALoS under non-convex objectives, including theoretical guarantees on the role of hierarchical momentum in asynchronous training. Empirically, HALoS attains up to 7.5× faster convergence than synchronous baselines in geo-distributed LLM training and improves upon existing asynchronous methods by up to 2.1×. Crucially, HALoS preserves the model quality of fully synchronous SGD—matching or exceeding accuracy on standard language modeling and downstream benchmarks—while substantially lowering total training time. These results demonstrate that hierarchical, server-side update accumulation and global model merging are powerful tools for scalable, efficient training of new-era LLMs in heterogeneous, geo-distributed environments.

We present a novel self-supervised feature learning method using Vision Transformers (ViT) as the backbone, specifically designed for object detection and instance segmentation. Our approach addresses the challenge of extracting features that capture both class and positional information, which are crucial for these tasks. The method introduces two key components: (1) a positional encoding tied to the cropping process in contrastive learning, which utilizes a novel vector field representation for positional embeddings; and (2) masking and prediction, similar to conventional Masked Image Modeling (MIM), applied in parallel to both content and positional embeddings of image patches. These components enable the effective learning of intertwined content and positional features. We evaluate our method against state-of-the-art approaches, pre-training on ImageNet-1K and fine-tuning on downstream tasks. Our method outperforms the state-of-the-art SSL methods on the COCO object detection benchmark, achieving significant improvements with fewer pre-training epochs. These results suggest that better integration of positional information into self-supervised learning can improve performance on the dense prediction tasks.

Navigating in an unfamiliar environment during deployment poses a critical challenge for a vision-language navigation (VLN) agent. Yet, test-time adaptation (TTA) remains relatively underexplored in robotic navigation, leading us to the fundamental question: what are the key properties of TTA for online VLN? In our view, effective adaptation requires three qualities: 1) flexibility in handling different navigation outcomes, 2) interactivity with external environment, and 3) maintaining a harmony between plasticity and stability. To address this, we introduce FeedTTA, a novel TTA framework for online VLN utilizing feedback-based reinforcement learning. Specifically, FeedTTA learns by maximizing binary episodic feedback, a practical setup in which the agent receives a binary scalar after each episode that indicates the success or failure of the navigation. Additionally, we propose a gradient regularization technique that leverages the binary structure of FeedTTA to achieve a balance between plasticity and stability during adaptation. Our extensive experiments on challenging VLN benchmarks demonstrate the superior adaptability of FeedTTA, even outperforming the state-of-the-art offline training methods in REVERIE benchmark with a single stream of learning.

Chest X-rays (CXRs) play an integral role in driving critical decisions in disease management and patient care. While recent innovations have led to specialized models for various CXR interpretation tasks, these solutions often operate in isolation, limiting their practical utility in clinical practice. We present MedRAX, the first versatile AI agent that seamlessly integrates state-of-the-art CXR analysis tools and multimodal large language models into a unified framework. MedRAX dynamically leverages these models to address complex medical queries without requiring additional training. To rigorously evaluate its capabilities, we introduce ChestAgentBench, a comprehensive benchmark containing 2,500 complex medical queries across 7 diverse categories. Our experiments demonstrate that MedRAX achieves state-of-the-art performance compared to both open-source and proprietary models, representing a significant step toward the practical deployment of automated CXR interpretation systems. Data and code have been publicly available at https://github.com/bowang-lab/MedRAX

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.

Effectively modeling and utilizing spatiotemporal features from RGB and other modalities (e.g., depth, thermal, and event data, denoted as X) is the core of RGB-X tracker design. Existing methods often employ two parallel branches to separately process the RGB and X input streams, requiring the model to simultaneously handle two dispersed feature spaces, which complicates both the model structure and computation process. More critically, intra-modality spatial modeling within each dispersed space incurs substantial computational overhead, limiting resources for inter-modality spatial modeling and temporal modeling. To address this, we propose a novel tracker, CSTrack, which focuses on modeling Compact Spatiotemporal features to achieve simple yet effective tracking. Specifically, we first introduce an innovative Spatial Compact Module that integrates the RGB-X dual input streams into a compact spatial feature, enabling thorough intra- and inter-modality spatial modeling. Additionally, we design an efficient Temporal Compact Module that compactly represents temporal features by constructing the refined target distribution heatmap. Extensive experiments validate the effectiveness of our compact spatiotemporal modeling method, with CSTrack achieving new SOTA results on mainstream RGB-X benchmarks. The code and models will be released at: https://github.com/XiaokunFeng/CSTrack.

Image Super-Resolution (ISR) has seen significant progress with the introduction of remarkable generative models. However, challenges such as the trade-off issues between fidelity and realism, as well as computational complexity, have also posed limitations on their application. Building upon the tremendous success of autoregressive models in the language domain, we propose \textbf{VARSR}, a novel visual autoregressive modeling for ISR framework with the form of next-scale prediction. To effectively integrate and preserve semantic information in low-resolution images, we propose using prefix tokens to incorporate the condition. Scale-aligned Rotary Positional Encodings are introduced to capture spatial structures and the diffusion refiner is utilized for modeling quantization residual loss to achieve pixel-level fidelity. Image-based Classifier-free Guidance is proposed to guide the generation of more realistic images. Furthermore, we collect large-scale data and design a training process to obtain robust generative priors. Quantitative and qualitative results show that VARSR is capable of generating high-fidelity and high-realism images with more efficiency than diffusion-based methods. Our codes are released at \url{https://github.com/quyp2000/VARSR}.

Real-world visual perception requires invariance to diverse transformations, yet current methods rely heavily on specialized architectures or training on predefined augmentations, limiting generalization. We propose FoCal, a test-time, data-driven framework that achieves robust perception by leveraging internet-scale visual priors from foundation models. By generating and optimizing candidate transformations toward visually typical, "canonical" views, FoCal enhances robustness without retraining or architectural changes. Experiments demonstrate improved robustness of CLIP and SAM across challenging transformations, including 2D/3D rotations, illumination shifts (contrast and color), and day-night variations. We also highlight potential applications in active vision. Our approach challenges the assumption that transform-specific training is necessary, instead offering a scalable path to invariance. Our code is available at: https://github.com/sutkarsh/focal.

Recent developments in prompt learning of large Vision-Language Models (VLMs) have significantly improved performance in target-specific tasks. However, these prompting methods often struggle to tackle the target-unspecific or generalizable tasks effectively. It may be attributed to the fact that overfitting training causes the model to forget its general knowledge. The general knowledge has a strong promotion on target-unspecific tasks. To alleviate this issue, we propose a novel Features Matrix (FM) approach designed to enhance these models on target-unspecific tasks. Our method extracts and leverages general knowledge, shaping a Features Matrix (FM). Specifically, the FM captures the semantics of diverse inputs from a deep and fine perspective, preserving essential general knowledge, which mitigates the risk of overfitting. Representative evaluations demonstrate that: 1) the FM is compatible with existing frameworks as a generic and flexible module, and 2) the FM significantly showcases its effectiveness in enhancing target-unspecific tasks (base-to-novel generalization, domain generalization, and cross-dataset generalization), achieving state-of-the-art performance.

Chain-of-thought (CoT) prompting has become a widely used strategy for improving large language and multimodal model performance. However, it is still an open question under which settings CoT systematically reduces performance. In this paper, we seek to identify the characteristics of tasks where CoT reduces performance by drawing inspiration from cognitive psychology, focusing on six representative tasks from the psychological literature where deliberation hurts performance in humans. In three of these tasks, state-of-the-art models exhibit significant performance drop-offs with CoT (up to 36.3\% absolute accuracy for OpenAI o1-preview compared to GPT-4o), while in others, CoT effects are mixed, with positive, neutral, and negative changes. While models and humans do not exhibit perfectly parallel cognitive processes, considering cases where thinking has negative consequences for humans helps identify settings where it negatively impacts models. By connecting the literature on human verbal thinking and deliberation with evaluations of CoT, we offer a perspective for understanding the impact of inference-time reasoning.

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.

Large vision-language models (LVLMs) have significantly advanced numerous fields. In this work, we explore how to harness their potential to address 3D scene understanding tasks, using 3D question answering (3D-QA) as a representative example. Due to the limited training data in 3D, we do not train LVLMs but infer in a zero-shot manner. Specifically, we sample 2D views from a 3D point cloud and feed them into 2D models to answer a given question. When the 2D model is chosen, e.g., LLAVA-OV, the quality of sampled views matters the most. We propose cdViews, a novel approach to automatically selecting critical and diverse Views for 3D-QA. cdViews consists of two key components: viewSelector prioritizing critical views based on their potential to provide answer-specific information, and viewNMS enhancing diversity by removing redundant views based on spatial overlap. We evaluate cdViews on the widely-used ScanQA and SQA benchmarks, demonstrating that it achieves state-of-the-art performance in 3D-QA while relying solely on 2D models without fine-tuning. These findings support our belief that 2D LVLMs are currently the most effective alternative (of the resource-intensive 3D LVLMs) for addressing 3D tasks.

Multimodal large language models (MLLMs) have achieved impressive progress in tasks such as visual question answering and visual understanding, but they still face significant challenges in emotional reasoning. Current methods to enhance emotional understanding typically rely on fine-tuning or manual annotations, which are resource-intensive and limit scalability. In this work, we focus on improving the ability of MLLMs to capture emotions during the inference phase. Specifically, MLLMs encounter two main issues: they struggle to distinguish between semantically similar emotions, leading to misclassification, and they are overwhelmed by redundant or irrelevant visual information, which distracts from key emotional cues. To address these, we propose Sharpening Emotion Perception in MLLMs (SEPM), which incorporates a Confidence-Guided Coarse-to-Fine Inference framework to refine emotion classification by guiding the model through simpler tasks. Additionally, SEPM employs Focus-on-Emotion Visual Augmentation to reduce visual redundancy by directing the attention of models to relevant emotional cues in images. Experimental results demonstrate that SEPM significantly improves MLLM performance on emotion-related tasks, providing a resource-efficient and scalable solution for emotion recognition.

Scaling laws guide the development of large language models (LLMs) by offering estimates for the optimal balance of model size, tokens, and compute. More recently, loss-to-loss scaling laws that relate losses across pretraining datasets and downstream tasks have emerged as a powerful tool for understanding and improving LLM performance and generalization. In this work, we investigate which factors most strongly influence loss-to-loss scaling. Our experiments reveal that the pretraining data determines the scaling trend. In contrast, model size, optimization hyperparameters, tokenizer and even significant architectural differences, such as between transformer-based models like Llama and state-space models like Mamba, generally have limited impact. Consequently, practitioners should carefully curate pretraining datasets for optimal downstream performance, while architectures and other settings can be freely optimized for training efficiency.

This paper proposes a simple yet effective jailbreak attack named FlipAttack against black-box LLMs. First, from the autoregressive nature, we reveal that LLMs tend to understand the text from left to right and find that they struggle to comprehend the text when the perturbation is added to the left side. Motivated by these insights, we propose to disguise the harmful prompt by constructing a left-side perturbation merely based on the prompt itself, then generalize this idea to 4 flipping modes. Second, we verify the strong ability of LLMs to perform the text-flipping task and then develop 4 variants to guide LLMs to understand and execute harmful behaviors accurately. These designs keep FlipAttack universal, stealthy, and simple, allowing it to jailbreak black-box LLMs within only 1 query. Experiments on 8 LLMs demonstrate the superiority of FlipAttack. Remarkably, it achieves $\sim$78.97\% attack success rate across 8 LLMs on average and $\sim$98\% bypass rate against 5 guard models on average.

Machine learning (ML) systems are utilized in critical sectors such as healthcare, law enforcement, and finance, but often rely on historical data that contains demographic biases, leading to decisions that perpetuate or intensify existing inequalities. Causal and counterfactual fairness provide a transparent, human-in-the-loop framework to mitigate algorithmic discrimination, aligning closely with legal doctrines of direct and indirect discrimination. However, current causal fairness frameworks hold a key limitation in that they assume prior knowledge of the correct causal model, restricting their applicability in complex fairness scenarios where causal models are unknown or difficult to identify. To bridge this gap, we propose FairPFN, a tabular foundation model pre-trained on synthetic causal fairness data to identify and mitigate the causal effects of protected attributes in its predictions. FairPFN's key contribution is that it requires no knowledge of the causal model and demonstrates strong performance across a diverse set of hand-crafted and real-world causal scenarios relative to robust baseline methods. FairPFN paves the way for a promising direction for future research, making causal fairness more accessible to a wider variety of complex fairness problems.

One approach to reducing the massive costs of large language models (LLMs) is the use of quantized or sparse representations for training or deployment. While post-training compression methods are very popular, the question of obtaining even more accurate compressed models by *directly training* over such representations, i.e., *Quantization-Aware Training (QAT)*, is still open: for example, a recent study put the "optimal" bit-width at which models can be trained using QAT, while staying accuracy-competitive with standard FP16/BF16 precision, at 8-bits weights and activations. We advance this state-of-the-art via a new method called QuEST, for which we demonstrate optimality at 4-bits and stable convergence as low as 1-bit weights and activations. QuEST achieves this by improving two key aspects of QAT methods: (1) accurate and fast quantization of the (continuous) distributions of weights and activations via Hadamard normalization and MSE-optimal fitting; (2) a new *trust gradient estimator* based on the idea of explicitly minimizing the error between the noisy gradient computed over quantized states and the "true" (but unknown) full-precision gradient. Experiments on Llama-type architectures show that QuEST induces stable scaling laws across the entire range of hardware-supported precisions, and can be extended to sparse representations. We provide GPU kernel support showing that models produced by QuEST can be executed efficiently. Our code is available at [https://github.com/IST-DASLab/QuEST](https://github.com/IST-DASLab/QuEST).

Recent advancements in Large Vision-Language Models built upon Large Language Models have established aligning visual features with LLM representations as the dominant paradigm. However, inherited LLM architectural designs introduce suboptimal characteristics for multimodal processing. First, LVLMs exhibit a bimodal distribution in attention allocation, leading to the progressive neglect of middle visual content as context expands. Second, conventional positional encoding schemes fail to preserve vital 2D structural relationships when processing dynamic high-resolution images. To address these limitations, we propose **CoMemo** - a dual-path architecture that combines a **Co**ntext image path with an image **Memo**ry path for visual processing, effectively alleviating visual information neglect. Additionally, we introduce RoPE-DHR, a novel positional encoding mechanism that employs thumbnail-based positional aggregation to maintain 2D spatial awareness while mitigating remote decay in extended sequences. Evaluations across seven benchmarks,including long-context comprehension, multi-image reasoning, and visual question answering, demonstrate CoMemo's superior performance compared to conventional LVLM architectures. Project page is available at [https://lalbj.github.io/projects/CoMemo/](https://lalbj.github.io/projects/CoMemo/).

Web-crawled image-text pairs are inherently noisy. Prior studies demonstrate that semantically aligning and enriching textual descriptions of these pairs can significantly enhance model training across various vision-language tasks, particularly text-to-image generation. However, large-scale investigations in this area remain predominantly closed-source. Our paper aims to bridge this community effort, leveraging the powerful and $\textit{open-sourced}$ LLaMA-3, a GPT-4 level LLM. Our recaptioning pipeline is simple: first, we fine-tune a LLaMA-3-8B powered LLaVA-1.5 and then employ it to recaption ~1.3 billion images from the DataComp-1B dataset. Our empirical results confirm that this enhanced dataset, Recap-DataComp-1B, offers substantial benefits in training advanced vision-language models. For discriminative models like CLIP, we observe an average of 3.1% enhanced zero-shot performance cross four cross-modal retrieval tasks using a mixed set of the original and our captions. For generative models like text-to-image Diffusion Transformers, the generated images exhibit a significant improvement in alignment with users' text instructions, especially in following complex queries. Our project page is https://www.haqtu.me/Recap-Datacomp-1B/.