ICML 2025 论文列表

Large Language Models (LLMs) have become indispensable in numerous real-world applications. However, fine-tuning these models at scale, especially in federated settings where data privacy and communication efficiency are critical, presents significant challenges. Existing approaches often resort to parameter-efficient fine-tuning (PEFT) to mitigate communication overhead, but this typically comes at the cost of model accuracy. To this end, we propose *federated full-parameter tuning at scale for LLMs* (Ferret), **the first first-order method with shared randomness** to enable scalable full-parameter tuning of LLMs across decentralized data sources while maintaining competitive model accuracy. Ferret accomplishes this through three aspects: **(i)** it employs widely used first-order methods for efficient local updates; **(ii)** it projects these updates into a low-dimensional space to considerably reduce communication overhead; and **(iii)** it reconstructs local updates from this low-dimensional space with shared randomness to facilitate effective full-parameter global aggregation, ensuring fast convergence and competitive final performance. Our rigorous theoretical analyses and insights along with extensive experiments, show that Ferret significantly enhances the scalability of existing federated full-parameter tuning approaches by achieving high computational efficiency, reduced communication overhead, and fast convergence, all while maintaining competitive model accuracy. Our implementation is available at [https://github.com/allen4747/Ferret](https://github.com/allen4747/Ferret).

Vovk (2015) introduced cross-conformal prediction, a modification of split conformal designed to improve the width of prediction sets. The method, when trained with a miscoverage rate equal to $\alpha$ and $n \gg K$, ensures a marginal coverage of at least $1 - 2\alpha - 2(1-\alpha)(K-1)/(n+K)$, where $n$ is the number of observations and $K$ denotes the number of folds. A simple modification of the method achieves coverage of at least $1-2\alpha$. In this work, we propose new variants of both methods that yield smaller prediction sets without compromising the latter theoretical guarantees. The proposed methods are based on recent results deriving more statistically efficient combination of p-values that leverage exchangeability and randomization. Simulations confirm the theoretical findings and bring out some important tradeoffs.

Reward-guided text generation (RGTG) has emerged as a viable alternative to offline reinforcement learning from human feedback (RLHF). RGTG methods can align baseline language models to human preferences without further training as in standard RLHF methods. However, they rely on a reward model to score each candidate token generated by the language model at inference, incurring significant test-time overhead. Additionally, the reward model is usually only trained to score full sequences, which can lead to sub-optimal choices for partial sequences. In this work, we present a novel reward model architecture that is trained, using a Bradley-Terry loss, to prefer the optimal expansion of a sequence with just a single call to the reward model at each step of the generation process. That is, a score for all possible candidate tokens is generated simultaneously, leading to efficient inference. We theoretically analyze various RGTG reward models and demonstrate that prior techniques prefer sub-optimal sequences compared to our method during inference. Empirically, our reward model leads to significantly faster inference than other RGTG methods. It requires fewer calls to the reward model and performs competitively compared to previous RGTG and offline RLHF methods.

Symbolic Regression (SR) algorithms select expressions based on prediction performance while also keeping the expression lengths short to produce explainable white box models. In this context, SR algorithms can be evaluated by measuring the extent to which the expressions discovered are Pareto-optimal, in the sense of having the best R-squared score for a given expression length. This evaluation is most commonly done based on relative performance, in the sense that an SR algorithm is judged on whether it Pareto-dominates other SR algorithms selected in the analysis, without any indication on efficiency or attainable limits. In this paper, we explore absolute Pareto-optimal (APO) solutions instead, which have the optimal tradeoff between the multiple SR objectives, for 34 datasets in the widely-used SR benchmark, SRBench, by performing exhaustive search. Additionally, we include comparisons between eight numerical optimization methods. We extract, for every dataset, an APO front of expressions that can serve as a universal baseline for SR algorithms that informs researchers of the best attainable performance for selected sizes. The APO fronts provided serves as an important benchmark and performance limit for SR algorithms and is made publicly available at: https://github.com/kentridgeai/SRParetoFronts

In multi-view multi-label classification (MVML), each object has multiple heterogeneous views and is annotated with multiple labels. The key to deal with such problem lies in how to capture cross-view consistent correlations while excavate multi-label semantic relationships. Existing MVML methods usually employ two independent components to address them separately, and ignores their potential interaction relationships. To address this issue, we propose a novel Tensorized MVML method named TMvML, which formulates an MVML tensor classifier to excavate comprehensive cross-view feature correlations while characterize complete multi-label semantic relationships. Specifically, we first reconstruct the MVML mapping matrices as an MVML tensor classifier. Then, we rotate the tensor classifier and introduce a low-rank tensor constraint to ensure view-level feature consistency and label-level semantic co-occurrence simultaneously. To better characterize the low-rank tensor structure, we design a new Laplace Tensor Rank (LTR), which serves as a tighter surrogate of tensor rank to capture high-order fiber correlations within the tensor space. By conducting the above operations, our method can easily address the two key challenges in MVML via a concise LTR tensor classifier and achieve the extraction of both cross-view consistent correlations and multi-label semantic relationships simultaneously. Extensive experiments demonstrate that TMvML significantly outperforms state-of-the-art methods.

Partial observability presents a significant challenge for safe reinforcement learning, as it impedes the identification of potential risks and rewards. Leveraging specific types of privileged information during training to mitigate the effects of partial observability has yielded notable empirical successes. In this paper, we propose Asymmetric Constrained Partially Observable Markov Decision Processes (ACPOMDPs) to theoretically examine the advantages of incorporating privileged information. Building upon ACPOMDPs, we propose the Privileged Information Guided Dreamer, a model-based safe reinforcement learning approach that leverages privileged information to enhance the agent's safety and performance through privileged representation alignment and an asymmetric actor-critic structure. Our empirical results demonstrate that our approach significantly outperforms existing methods in terms of safety and task-centric performance. Meanwhile, compared to alternative privileged model-based reinforcement learning methods, our approach exhibits superior performance and ease of training.

Knowledge distillation typically minimizes the Kullback–Leibler (KL) divergence between teacher and student logits. However, optimizing the KL divergence can be challenging for the student and often leads to sub-optimal solutions. We further show that gradients induced by KL divergence scale with the magnitude of the teacher logits, thereby diminishing updates on low-probability channels. This imbalance weakens the transfer of inter-class information and in turn limits the performance improvements achievable by the student. To mitigate this issue, we propose a plug-and-play auxiliary ranking loss based on Kendall’s $\tau$ coefficient that can be seamlessly integrated into any logit-based distillation framework. It supplies inter-class relational information while rebalancing gradients toward low-probability channels. We demonstrate that the proposed ranking loss is largely invariant to channel scaling and optimizes an objective aligned with that of KL divergence, making it a natural complement rather than a replacement. Extensive experiments on CIFAR-100, ImageNet, and COCO datasets, as well as various CNN and ViT teacher-student architecture combinations, demonstrate that our plug-and-play ranking loss consistently boosts the performance of multiple distillation baselines.

This paper investigates the fusion of absolute (reward) and relative (dueling) feedback in stochastic bandits, where both feedback types are gathered in each decision round. We derive a regret lower bound, demonstrating that an efficient algorithm may incur only the smaller among the reward and dueling-based regret for each individual arm. We propose two fusion approaches: (1) a simple elimination fusion algorithm that leverages both feedback types to explore all arms and unifies collected information by sharing a common candidate arm set, and (2) a decomposition fusion algorithm that selects the more effective feedback to explore the corresponding arms and randomly assigns one feedback type for exploration and the other for exploitation in each round. The elimination fusion experiences a suboptimal multiplicative term of the number of arms in regret due to the intrinsic suboptimality of dueling elimination. In contrast, the decomposition fusion achieves regret matching the lower bound up to a constant under a common assumption. Extensive experiments confirm the efficacy of our algorithms and theoretical results.

Although large language models (LLMs) show promise in solving complex mathematical tasks, existing evaluation paradigms rely solely on a coarse measure of overall answer accuracy, which are insufficient for assessing their authentic capabilities. In this paper, we propose \textbf{CogMath}, which comprehensively assesses LLMs' mathematical abilities through the lens of human cognition. Specifically, inspired by psychological theories, CogMath formalizes human reasoning process into 3 stages: \emph{problem comprehension}, \emph{problem solving}, and \emph{solution summarization}. Within these stages, we investigate perspectives such as numerical calculation, knowledge, and counterfactuals, and design a total of 9 fine-grained evaluation dimensions. In each dimension, we develop an ``\emph{Inquiry}-\emph{Judge}-\emph{Reference}'' multi-agent system to generate inquiries that assess LLMs' mastery from this dimension. An LLM is considered to truly master a problem only when excelling in all inquiries from the 9 dimensions. By applying CogMath on three benchmarks, we reveal that the mathematical capabilities of 7 mainstream LLMs are overestimated by 30\%-40\%. Moreover, we locate their strengths and weaknesses across specific stages/dimensions, offering in-depth insights to further enhance their reasoning abilities.

Selecting a sample generation scheme from multiple prompt-based generative models, including large language models (LLMs) and prompt-guided image and video generation models, is typically addressed by choosing the model that maximizes an averaged evaluation score. However, this score-based selection overlooks the possibility that different models achieve the best generation performance for different types of text prompts. An online identification of the best generation model for various input prompts can reduce the costs associated with querying sub-optimal models. In this work, we explore the possibility of varying rankings of text-based generative models for different text prompts and propose an online learning framework to predict the best data generation model for a given input prompt. The proposed PAK-UCB algorithm addresses a contextual bandit (CB) setting with shared context variables across the arms, utilizing the generated data to update kernel-based functions that predict the score of each model available for unseen text prompts. Additionally, we leverage random Fourier features (RFF) to accelerate the online learning process of PAK-UCB. Our numerical experiments on real and simulated text-to-image and image-to-text generative models show that RFF-UCB performs successfully in identifying the best generation model across different sample types. The code is available at: [github.com/yannxiaoyanhu/dgm-online-select](github.com/yannxiaoyanhu/dgm-online-select).

Self-Explainable Graph Neural Networks (SE-GNNs) are popular explainable-by-design GNNs, but their explanations' properties and limitations are not well understood. Our first contribution fills this gap by formalizing the explanations extracted by some popular SE-GNNs, referred to as Minimal Explanations (MEs), and comparing them to established notions of explanations, namely Prime Implicant (PI) and faithful explanations. Our analysis reveals that MEs match PI explanations for a restricted but significant family of tasks. In general, however, they can be less informative than PI explanations and are surprisingly misaligned with widely accepted notions of faithfulness. Although faithful and PI explanations are informative, they are intractable to find and we show that they can be prohibitively large. Given these observations, a natural choice is to augment SE-GNNs with alternative modalities of explanations taking care of SE-GNNs’ limitations. To this end, we propose Dual-Channel GNNs that integrate a white-box rule extractor and a standard SE-GNN, adaptively combining both channels. Our experiments show that even a simple instantiation of Dual-Channel GNNs can recover succinct rules and perform on par or better than widely used SE-GNNs.

Mathematical diagrams have a distinctive structure. Standard feature transforms designed for natural images (e.g., CLIP) fail to process them effectively, limiting their utility in multimodal large language models (MLLMs). Current efforts to improve MLLMs have primarily focused on scaling mathematical visual instruction datasets and strengthening LLM backbones, yet fine‐grained visual recognition errors remain unaddressed. Our systematic evaluation on the visual grounding capabilities of state‐of‐the‐art MLLMs highlights that fine‐grained visual understanding remains a crucial bottleneck in visual mathematical reasoning (GPT-4o exhibits a 70% grounding error rate, and correcting these errors improves reasoning accuracy by 12%). We thus propose a novel approach featuring a geometrically‐grounded vision encoder and a feature router that dynamically selects between hierarchical visual feature maps. Our model accurately recognizes visual primitives and generates precise visual prompts aligned with the language model’s reasoning needs. In experiments, PRIMITIVE-Qwen2.5-7B outperforms other 7B models by 12% on MathVerse and is on par with GPT-4V on MathVista. Our findings highlight the need for better fine‐grained visual integration in MLLMs. Code is available at github.com/AI4Math-ShanZhang/SVE-Math.

Cooperative multi-agent reinforcement learning faces significant challenges in effectively organizing agent relationships and facilitating information exchange, particularly when agents need to adapt their coordination patterns dynamically. This paper presents a novel framework that integrates dynamic spectral clustering with hypergraph neural networks to enable adaptive group formation and efficient information processing in multi-agent systems. The proposed framework dynamically constructs and updates hypergraph structures through spectral clustering on agents' state histories, enabling higher-order relationships to emerge naturally from agent interactions. The hypergraph structure is enhanced with attention mechanisms for selective information processing, providing an expressive and efficient way to model complex agent relationships. This architecture can be implemented in both value-based and policy-based paradigms through a unified objective combining task performance with structural regularization. Extensive experiments on challenging cooperative tasks demonstrate that our method significantly outperforms state-of-the-art approaches in both sample efficiency and final performance. The code is available at: https://github.com/mysteryelder/HYGMA.

The integration of long-context capabilities with visual understanding unlocks unprecedented potential for Vision Language Models (VLMs). However, the quadratic attention complexity during the pre-filling phase remains a significant obstacle to real-world deployment. To overcome this limitation, we introduce MMInference (Multimodality Million tokens Inference), a dynamic sparse attention method that accelerates the prefilling stage for long-context multi-modal inputs. First, our analysis reveals that the temporal and spatial locality of video input leads to a unique sparse pattern, the Grid pattern. Simultaneously, VLMs exhibit markedly different sparse distributions across different modalities. We introduce a permutation-based method to leverage the unique Grid pattern and handle modality boundary issues. By offline search the optimal sparse patterns for each head, MMInference constructs the sparse distribution dynamically based on the input. We also provide optimized GPU kernels for efficient sparse computations. Notably, MMInference integrates seamlessly into existing VLM pipelines without any model modifications or fine-tuning. Experiments on multi-modal benchmarks-including Video QA, Captioning, VisionNIAH, and Mixed-Modality NIAH-with state-of-the-art long-context VLMs (LongVila, LlavaVideo, VideoChat-Flash, Qwen2.5-VL) show that MMInference accelerates the pre-filling stage by up to 8.3x at 1M tokens while maintaining accuracy. Our code is available at https://ama.ms/MMInference.

Incorporating permutation equivariance into neural networks has proven to be useful in ensuring that models respect symmetries that exist in data. Symmetric tensors, which naturally appear in statistics, machine learning, and graph theory, are essential for many applications in physics, chemistry, and materials science, amongst others. However, existing research on permutation equivariant models has not explored symmetric tensors as inputs, and most prior work on learning from these tensors has focused on equivariance to Euclidean groups. In this paper, we present two different characterisations of all linear permutation equivariant functions between symmetric power spaces of $\mathbb{R}^{n}$. We show on two tasks that these functions are highly data efficient compared to standard MLPs and have potential to generalise well to symmetric tensors of different sizes.

We investigate the fundamental limits of transformer-based foundation models, extending our analysis to include Visual Autoregressive (VAR) transformers. VAR represents a big step toward generating images using a novel, scalable, coarse-to-fine ``next-scale prediction'' framework. These models set a new quality bar, outperforming all previous methods, including Diffusion Transformers, while having state-of-the-art performance for image synthesis tasks. Our primary contributions establish that, for single-head VAR transformers with a single self-attention layer and single interpolation layer, the VAR Transformer is universal. From the statistical perspective, we prove that such simple VAR transformers are universal approximators for any word-to-image Lipschitz functions. Furthermore, we demonstrate that flow-based autoregressive transformers inherit similar approximation capabilities. Our results provide important design principles for effective and computationally efficient VAR Transformer strategies that can be used to extend their utility to more sophisticated VAR models in image generation and other related areas.

Knowledge Graph Foundation Models (KGFMs) are at the frontier for deep learning on knowledge graphs (KGs), as they can generalize to completely novel knowledge graphs with different relational vocabularies. Despite their empirical success, our theoretical understanding of KGFMs remains very limited. In this paper, we conduct a rigorous study of the expressive power of KGFMs. Specifically, we show that the expressive power of KGFMs directly depends on the *motifs* that are used to learn the relation representations. We then observe that the most typical motifs used in the existing literature are *binary*, as the representations are learned based on how pairs of relations interact, which limits the model's expressiveness. As part of our study, we design more expressive KGFMs using richer motifs, which necessitate learning relation representations based on, e.g., how triples of relations interact with each other. Finally, we empirically validate our theoretical findings, showing that the use of richer motifs results in better performance on a wide range of datasets drawn from different domains.

This paper examines the systemic risks posed by incremental advancements in artificial intelligence, developing the concept of `gradual disempowerment', in contrast to the abrupt takeover scenarios commonly discussed in AI safety. We analyze how even incremental improvements in AI capabilities can undermine human influence over large-scale systems that society depends on, including the economy, culture, and nation-states. As AI increasingly replaces human labor and cognition in these domains, it can weaken both explicit human control mechanisms (like voting and consumer choice) and the implicit alignments with human preferences that often arise from societal systems' reliance on human participation to function. Furthermore, AI systems may amplify existing misalignments with human preferences by optimizing these systems more powerfully. These distortions across domains may be mutually reinforcing: economic power shapes cultural narratives and political decisions, while cultural shifts alter economic and political behavior. We argue that this dynamic could lead to an effectively irreversible loss of human influence over crucial societal systems, precipitating an existential catastrophe through the permanent disempowerment of humanity. This analysis suggests the need for both technical research and governance approaches that specifically address the risk of incremental erosion of human influence across interconnected societal systems.

In image processing, solving inverse problems is the task of finding plausible reconstructions of an image that was corrupted by some (usually known) degradation operator. Commonly, this process is done using a generative image model that can guide the reconstruction towards solutions that appear natural. The success of diffusion models over the last few years has made them a leading candidate for this task. However, the sequential nature of diffusion models makes this conditional sampling process challenging. Furthermore, since diffusion models are often defined in the latent space of an autoencoder, the encoder-decoder transformations introduce additional difficulties. To address these challenges, we suggest a novel sampling method based on sequential Monte Carlo (SMC) in the latent space of diffusion models. We name our method LD-SMC. We define a generative model for the data using additional auxiliary observations and perform posterior inference with SMC sampling based on a backward diffusion process. Empirical evaluations on ImageNet and FFHQ show the benefits of LD-SMC over competing methods in various inverse problem tasks and especially in challenging inpainting tasks.