ICML 2025 Papers

This paper formalizes an emerging learning paradigm that uses a trained model as a reference to guide and enhance the training of a target model through strategic data selection or weighting, named **model steering**. While ad-hoc methods have been used in various contexts, including the training of large foundation models, its underlying principles remain insufficiently understood, leading to sub-optimal performance. In this work, we propose a theory-driven framework for model steering called **DRRho risk minimization**, which is rooted in Distributionally Robust Optimization (DRO). Through a generalization analysis, we provide theoretical insights into why this approach improves generalization and data efficiency compared to training without a reference model. To the best of our knowledge, this is the first time such theoretical insights are provided for the new learning paradigm, which significantly enhance our understanding and practice of model steering. Building on these insights and the connection between contrastive learning and DRO, we introduce a novel method for Contrastive Language-Image Pretraining (CLIP) with a reference model, termed DRRho-CLIP. Extensive experiments validate the theoretical insights, reveal a superior scaling law compared to CLIP without a reference model, and demonstrate its strength over existing heuristic approaches. Code is released at [github.com/Optimization-AI/DRRho-CLIP](https://github.com/Optimization-AI/DRRho-CLIP)

Semi-supervised learning (SSL) has made notable advancements in medical image segmentation (MIS), particularly in scenarios with limited labeled data and significantly enhancing data utilization efficiency. Previous methods primarily focus on complex training strategies to utilize unlabeled data but neglect the importance of graph structural information. Different from existing methods, we propose a graph-based clustering for semi-supervised medical image segmentation (GraphCL) by jointly modeling graph data structure in a unified deep model. The proposed GraphCL model enjoys several advantages. Firstly, to the best of our knowledge, this is the first work to model the data structure information for semi-supervised medical image segmentation (SSMIS). Secondly, to get the clustered features across different graphs, we integrate both pairwise affinities between local image features and raw features as inputs. Extensive experimental results on three standard benchmarks show that the proposed GraphCL algorithm outperforms state-of-the-art semi-supervised medical image segmentation methods.

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.

Transformer-based large language models (LLMs) have already achieved remarkable results on long-text tasks, but the limited GPU memory (VRAM) resources struggle to accommodate the linearly growing demand for key-value (KV) cache as the sequence length increases, which has become a bottleneck for the application of LLMs on long sequences. Existing KV cache compression methods include eviction, merging, or quantization of the KV cache to reduce its size. However, compression results in irreversible information forgetting, potentially affecting the accuracy of subsequent decoding. In this paper, we propose SpeCache, which takes full advantage of the large and easily expandable CPU memory to offload the complete KV cache, and dynamically fetches KV pairs back in each decoding step based on their importance measured by low-precision KV cache copy in VRAM. To avoid inference latency caused by CPU-GPU communication, SpeCache speculatively predicts the KV pairs that the next token might attend to, allowing us to prefetch them before the next decoding step which enables parallelization of prefetching and computation. Experiments on LongBench and Needle-in-a-Haystack benchmarks verify that SpeCache effectively reduces VRAM usage while avoiding information forgetting for long sequences without re-training, even with a 10x high KV cache compression ratio.

Vision Transformer (ViT) has achieved remarkable success due to its large-scale pretraining on general domains, but it still faces challenges when applying it to downstream distant domains that have only scarce training data, which gives rise to the Cross-Domain Few-Shot Learning (CDFSL) task. Inspired by Self-Attention's insensitivity to token orders, we find an interesting phenomenon neglected in current works: disrupting the continuity of image tokens (i.e., making pixels not smoothly transited across patches) in ViT leads to a noticeable performance decline in the general (source) domain but only a marginal decrease in downstream target domains. This questions the role of image tokens' continuity in ViT's generalization under large domain gaps. In this paper, we delve into this phenomenon for an interpretation. We find continuity aids ViT in learning larger spatial patterns, which are harder to transfer than smaller ones, enlarging domain distances. Meanwhile, it implies that only smaller patterns within each patch could be transferred under extreme domain gaps. Based on this interpretation, we further propose a simple yet effective method for CDFSL that better disrupts the continuity of image tokens, encouraging the model to rely less on large patterns and more on smaller ones. Extensive experiments show the effectiveness of our method in reducing domain gaps and outperforming state-of-the-art works. Codes and models are available at https://github.com/shuaiyi308/ReCIT.

Pretrained vision-language models (VLMs), e.g., CLIP, demonstrate impressive zero-shot capabilities on downstream tasks. Prior research highlights the crucial role of visual augmentation techniques, like random cropping, in alignment with fine-grained class descriptions generated by large language models (LLMs), significantly enhancing zero-shot performance by incorporating multi-view information. However, the inherent randomness of these augmentations can inevitably introduce background artifacts and cause models to overly focus on local details, compromising global semantic understanding. To address these issues, we propose an **A**ttention-**B**ased **S**election (**ABS**) method from local details to global context, which applies attention-guided cropping in both raw images and feature space, supplement global semantic information through strategic feature selection. Additionally, we introduce a soft matching technique to effectively filter LLM descriptions for better alignment. **ABS** achieves state-of-the-art performance on out-of-distribution generalization and zero-shot classification tasks. Notably, **ABS** is training-free and even rivals few-shot and test-time adaptation methods.

Node classification is a fundamental task in graph analysis, with broad applications across various fields. Recent breakthroughs in Large Language Models (LLMs) have enabled LLM-based approaches for this task. Although many studies demonstrate the impressive performance of LLM-based methods, the lack of clear design guidelines may hinder their practical application. In this work, we aim to establish such guidelines through a fair and systematic comparison of these algorithms. As a first step, we developed LLMNodeBed, a comprehensive codebase and testbed for node classification using LLMs. It includes 10 homophilic datasets, 4 heterophilic datasets, 8 LLM-based algorithms, 8 classic baselines, and 3 learning paradigms. Subsequently, we conducted extensive experiments, training and evaluating over 2,700 models, to determine the key settings (e.g., learning paradigms and homophily) and components (e.g., model size and prompt) that affect performance. Our findings uncover 8 insights, e.g., (1) LLM-based methods can significantly outperform traditional methods in a semi-supervised setting, while the advantage is marginal in a supervised setting; (2) Graph Foundation Models can beat open-source LLMs but still fall short of strong LLMs like GPT-4o in a zero-shot setting. We hope that the release of LLMNodeBed, along with our insights, will facilitate reproducible research and inspire future studies in this field. Codes and datasets are released at https://llmnodebed.github.io/.

Medical imaging faces significant challenges in single-domain generalization (SDG) due to the diversity of imaging devices and the variability among data collection centers. To address these challenges, we propose \textbf{TinyMIG}, a framework designed to transfer generalization capabilities from vision foundation models to medical imaging SDG. TinyMIG aims to enable lightweight specialized models to mimic the strong generalization capabilities of foundation models in terms of both global feature distribution and local fine-grained details during training. Specifically, for global feature distribution, we propose a Global Distribution Consistency Learning strategy that mimics the prior distributions of the foundation model layer by layer. For local fine-grained details, we further design a Localized Representation Alignment method, which promotes semantic alignment and generalization distillation between the specialized model and the foundation model. These mechanisms collectively enable the specialized model to achieve robust performance in diverse medical imaging scenarios. Extensive experiments on large-scale benchmarks demonstrate that TinyMIG, with extremely low computational cost, significantly outperforms state-of-the-art models, showcasing its superior SDG capabilities. All the code and model weights will be publicly available.

Despite their tremendous success and versatility, Deep Neural Networks (DNNs) such as Large Language Models (LLMs) suffer from inference inefficiency and rely on advanced computational infrastructure. To address these challenges and make these models more accessible and cost-effective, in this paper, we propose algorithms to improve the inference time and memory efficiency of DNNs with binary and ternary weight matrices. Particularly focusing on matrix multiplication as the bottleneck operation of inference, we observe that, once trained, the weight matrices of a model no longer change. This allows us to preprocess these matrices and create indices that help reduce the storage requirements by a logarithmic factor while enabling our efficient inference algorithms. Specifically, for a $n\times n$ weight matrix, our efficient algorithm guarantees a time complexity of $O(\frac{n^2}{\log n})$, a logarithmic factor improvement over the standard vector-matrix multiplication. Besides theoretical analysis, we conduct extensive experiments to evaluate the practical efficiency of our algorithms. Our results confirm the superiority of our approach both with respect to time and memory, as we observed a reduction in the multiplication time up to 29x and memory usage up to 6x. When applied to LLMs, our experiments show up to a 5.24x speedup in the inference time.

Instruction-based image editing, which aims to modify the image faithfully towards instruction while preserving irrelevant content unchanged, has made advanced progresses. However, there still lacks a comprehensive metric for assessing the editing quality. Existing metrics either require high costs concerning human evaluation, which hinders large-scale evaluation, or adapt from other tasks and lose specified concerns, failing to comprehensively evaluate the modification of instruction and the preservation of irrelevant regions, resulting in biased evaluation. To tackle it, we introduce a new metric Balancing Preservation Modification (BPM), that tailored for instruction-based image editing by explicitly disentangling the image into editing-relevant and irrelevant regions for specific consideration. We first identify and locate editing-relevant regions, followed by a two-tier process to assess editing quality: Region-Aware Judge evaluates whether the position and size of the edited region align with instruction, and Semantic-Aware Judge further assesses the instruction content compliance within editing-relevant regions as well as content preservation within irrelevant regions, yielding comprehensive and interpretable quality assessment. Moreover, the editing-relevant region localization in BPM can be integrated into image editing approaches to improve the editing quality, manifesting its wild application. We verify the effectiveness of BPM metric on comprehensive instruction-editing data, and the re- sults show that we yield the highest alignment with human evaluation compared to existing metrics, indicating efficacy. The code is available at https://joyli-x.github.io/BPM/.

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.

Diffusion Transformers (DiTs) excel in generative tasks but face practical deployment challenges due to high inference costs. Feature caching, which stores and retrieves redundant computations, offers the potential for acceleration. Existing learning-based caching, though adaptive, overlooks the impact of the prior timestep. It also suffers from misaligned objectives-*aligned predicted noise vs. high-quality images*-between training and inference. These two discrepancies compromise both performance and efficiency. To this end, we *harmonize* training and inference with a novel learning-based *caching* framework dubbed **HarmoniCa**. It first incorporates *Step-Wise Denoising Training* (SDT) to ensure the continuity of the denoising process, where prior steps can be leveraged. In addition, an *Image Error Proxy-Guided Objective* (IEPO) is applied to balance image quality against cache utilization through an efficient proxy to approximate the image error. Extensive experiments across $8$ models, $4$ samplers, and resolutions from $256\times256$ to $2K$ demonstrate superior performance and speedup of our framework. For instance, it achieves over $40\\%$ latency reduction (*i.e.*, $2.07\times$ theoretical speedup) and improved performance on PixArt-$\alpha$. Remarkably, our *image-free* approach reduces training time by $25\\%$ compared with the previous method. Our code is available at https://github.com/ModelTC/HarmoniCa.

We present Self-Organizing Visual Prototypes (SOP), a new training technique for unsupervised visual feature learning. Unlike existing prototypical self-supervised learning (SSL) methods that rely on a single prototype to encode all relevant features of a hidden cluster in the data, we propose the SOP strategy. In this strategy, a prototype is represented by many semantically similar representations, or support embeddings (SEs), each containing a complementary set of features that together better characterize their region in space and maximize training performance. We reaffirm the feasibility of non-parametric SSL by introducing novel non-parametric adaptations of two loss functions that implement the SOP strategy. Notably, we introduce the SOP Masked Image Modeling (SOP-MIM) task, where masked representations are reconstructed from the perspective of multiple non-parametric local SEs. We comprehensively evaluate the representations learned using the SOP strategy on a range of benchmarks, including retrieval, linear evaluation, fine-tuning, and object detection. Our pre-trained encoders achieve state-of-the-art performance on many retrieval benchmarks and demonstrate increasing performance gains with more complex encoders.

Vision-Language Models (VLMs) face two critical limitations in visual representation learning: degraded supervision due to information loss during gradient propagation, and the inherent semantic sparsity of textual supervision compared to visual data. We propose the Diffusion Supervision Vision-Language Model (DS-VLM), a plug-and-play framework that introduces diffusion-based direct supervision for vision-language alignment. By reconstructing input images through a diffusion model conditioned on outputs of the visual encoder and the connector, our method establishes a short-path gradient propagation channel from pixel space to visual features. This approach simultaneously preserves high-level semantic alignment through conventional text supervision while enhancing visual feature quality via pixel-level reconstruction constraints. Extensive experiments conducted across various visual encoders and LLMs of different scales demonstrate the effectiveness of our approach.

Abstract visual reasoning (AVR) involves discovering shared concepts across images through analogy, akin to solving IQ test problems. Bongard Problems (BPs) remain a key challenge in AVR, requiring both visual reasoning and verbal description. We investigate whether multimodal large language models (MLLMs) can solve BPs by formulating a set of diverse MLLM-suited solution strategies and testing $4$ proprietary and $4$ open-access models on $3$ BP datasets featuring synthetic (classic BPs) and real-world (Bongard HOI and Bongard-OpenWorld) images. Despite some successes on real-world datasets, MLLMs struggle with synthetic BPs. To explore this gap, we introduce Bongard-RWR, a dataset representing synthetic BP concepts using real-world images. Our findings suggest that weak MLLM performance on classical BPs is not due to the domain specificity, but rather comes from their general AVR limitations. Code and dataset are available at: https://github.com/pavonism/bongard-rwr

Vision-Language-Action (VLA) models have shown great potential in general robotic decision-making tasks via imitation learning. However, the variable quality of training data often constrains the performance of these models. On the other hand, offline Reinforcement Learning (RL) excels at learning robust policy models from mixed-quality data. In this paper, we introduce Reinforced robot GPT (ReinboT), a novel end-to-end VLA model that integrates the RL principle of maximizing cumulative reward. ReinboT achieves a deeper understanding of the data quality distribution by predicting dense returns that capture the nuances of manipulation tasks. The dense return prediction capability enables the robot to generate more robust decision-making actions, oriented towards maximizing future benefits. Extensive experiments show that ReinboT achieves state-of-the-art performance on the CALVIN mixed-quality dataset and exhibits superior few-shot learning and out-of-distribution generalization capabilities in real-world tasks.

In the Correlation Clustering problem, we are given an undirected graph and are tasked with computing a clustering (partition of the nodes) that minimizes the sum of the number of edges across different clusters and the number of non-edges within clusters. In the constrained version of this problem, the goal is to compute a clustering that satisfies additional hard constraints mandating certain pairs to be in the same cluster and certain pairs to be in different clusters. Constrained Correlation Clustering is APX-Hard, and the best known approximation factor is 3 (van Zuylen et al. [SODA '07]). In this work, we show that in order to obtain a better-than-2 approximation, solving the (exponentially large) Constrained Cluster LP would be sufficient. [The peer-reviewed version of this article claimed an efficient algorithm for solving the Constrained Cluster LP. An error in the proof, that the authors discovered after the review process, led them to revise the results to be conditional on the existence of a valid LP solution.]

The remarkable generalization performance of contrastive vision-language models like CLIP is often attributed to the diversity of their training distributions. However, key questions remain unanswered: Can CLIP generalize to an entirely unseen domain when trained on a diverse mixture of domains (domain generalization)? Can it generalize to unseen classes within partially seen domains (compositional generalization)? What factors affect such generalization? To answer these questions, we trained CLIP models on systematically constructed training distributions with controlled domain diversity and object class exposure. Our experiments show that domain diversity is essential for both domain and compositional generalization, yet compositional generalization can be surprisingly weaker than domain generalization when the training distribution contains a suboptimal subset of the test domain. Through data-centric *and* mechanistic analyses, we find that successful generalization requires the learning of sufficiently shared representations in intermediate layers and circuits.

Recent breakthroughs in solving reasoning, math and coding problems with Large Language Models (LLMs) have been enabled by investing substantial computation budgets at inference time. Therefore, inference speed is one of the most critical properties of LLM architectures, and there is a growing need for LLMs that are efficient and fast at inference. Recently, LLMs built on the xLSTM architecture have emerged as a powerful alternative to Transformers, offering linear compute scaling with sequence length and constant memory usage, both highly desirable properties for efficient inference. However, such xLSTM-based LLMs have yet to be scaled to larger models and assessed and compared with respect to inference speed and efficiency. In this work, we introduce xLSTM 7B, a 7-billion-parameter LLM that combines xLSTM’s architectural benefits with targeted optimizations for fast and efficient inference. Our experiments demonstrate that xLSTM 7B achieves performance on downstream tasks comparable to other similar-sized LLMs, while providing significantly faster inference speeds and greater efficiency compared to Llama- and Mamba-based LLMs. These results establish xLSTM 7B as the fastest and most efficient 7B LLM, offering a solution for tasks that require large amounts of test-time computation. Our work highlights xLSTM’s potential as a foundational architecture for methods building on heavy use of LLM inference. Our model weights, model code and training code are open-source. Model: https://huggingface.co/NX-AI/xLSTM-7b Code: https://github.com/NX-AI/xlstm and https://github.com/NX-AI/xlstm-jax.