ICML 2025 论文列表

Large transformer models are known to produce high-norm tokens. In vision transformers (ViTs), such tokens have been mathematically modeled through the singular vectors of the linear approximations of layers. However, in large language models (LLMs), the underlying causes of high-norm tokens remain largely unexplored, and their different properties from those of ViTs require a new analysis framework. In this paper, we provide both theoretical insights and empirical validation across a range of recent models, leading to the following observations: i) The layer-wise singular direction predicts the abrupt explosion of token norms in LLMs. ii) The negative eigenvalues of a layer explain its sudden decay. iii) The computational pathways leading to high-norm tokens differ between initial and noninitial tokens. iv) High-norm tokens are triggered by the right leading singular vector of the matrix approximating the corresponding modules. We showcase two practical applications of these findings: the improvement of quantization schemes and the design of LLM signatures. Our findings not only advance the understanding of singular defects in LLMs but also open new avenues for their application. We expect that this work will stimulate further research into the internal mechanisms of LLMs. Code is released at https://github.com/haoqiwang/singular_defect.

Message-passing graph neural networks (MPNNs) and structural features (SFs) are cornerstones for the link prediction task. However, as a common and intuitive mode of understanding, the potential of visual perception has been overlooked in the MPNN community. For the first time, we equip MPNNs with vision structural awareness by proposing an effective framework called Graph Vision Network (GVN), along with a more efficient variant (E-GVN). Extensive empirical results demonstrate that with the proposed frameworks, GVN consistently benefits from the vision enhancement across seven link prediction datasets, including challenging large-scale graphs. Such improvements are compatible with existing state-of-the-art (SOTA) methods and GVNs achieve new SOTA results, thereby underscoring a promising novel direction for link prediction.

Human body restoration, as a specific application of image restoration, is widely applied in practice and plays a vital role across diverse fields. However, thorough research remains difficult, particularly due to the lack of benchmark datasets. In this study, we propose a high-quality dataset automated cropping and filtering (HQ-ACF) pipeline. This pipeline leverages existing object detection datasets and other unlabeled images to automatically crop and filter high-quality human images. Using this pipeline, we constructed a person-based restoration with sophisticated objects and natural activities (*PERSONA*) dataset, which includes training, validation, and test sets. The dataset significantly surpasses other human-related datasets in both quality and content richness. Finally, we propose *OSDHuman*, a novel one-step diffusion model for human body restoration. Specifically, we propose a high-fidelity image embedder (HFIE) as the prompt generator to better guide the model with low-quality human image information, effectively avoiding misleading prompts. Experimental results show that OSDHuman outperforms existing methods in both visual quality and quantitative metrics. The dataset and code are available at: https://github.com/gobunu/OSDHuman.

Recent advancements in large language models (LLMs) have significantly enhanced the ability of LLM-based systems to perform complex tasks through natural language processing and tool interaction. However, optimizing these LLM-based systems for specific tasks remains challenging, often requiring manual interventions like prompt engineering and hyperparameter tuning. Existing automatic optimization methods, such as textual feedback-based techniques (*e.g.*, TextGrad), tend to focus on immediate feedback, analogous to using immediate derivatives in traditional numerical gradient descent. However, relying solely on such feedback can be limited when the adjustments made in response to this feedback are either too small or fluctuate irregularly, potentially slowing down or even stalling the optimization process. In this paper, we introduce $\textbf{REVOLVE}$, an optimization method that tracks how $\textbf{R}$esponses $\textbf{EVOLVE}$ across iterations in LLM systems. By focusing on the evolution of responses over time, REVOLVE enables more stable and effective optimization by making thoughtful, progressive adjustments at each step. Experiments across three tasks demonstrate the adaptability and efficiency of our proposal. Beyond its practical contributions, REVOLVE highlights a promising direction, where the rich knowledge from established optimization principles can be leveraged to enhance LLM systems, which paves the way for further advancements in this hybrid domain. Code is available at: https://llm-revolve.netlify.app.

While Graph Neural Networks (GNNs) have shown outstanding performance in node classification tasks, they are vulnerable to adversarial attacks, which are imperceptible changes to input samples. Adversarial training, as a widely used tool to enhance the adversarial robustness of GNNs, has presented remarkable effectiveness in node classification tasks. However, the generalization properties for explaining their behaviors remain not well understood from the theoretical viewpoint. To fill this gap, we develop a high probability generalization bound of general GNNs in adversarial learning through covering number analysis. We estimate the covering number of the GNN model class based on the entire perturbed feature matrix by constructing a cover for the perturbation set. Our results are generally applicable to a series of GNNs. We demonstrate their applicability by investigating the generalization performance of several popular GNN models under adversarial attacks, which reveal the architecture-related factors influencing the generalization gap. Our experimental results on benchmark datasets provide evidence that supports the established theoretical findings.

We consider molecule generation in 3D space using language models (LMs), which requires discrete tokenization of 3D molecular geometries. Although tokenization of molecular graphs exists, that for 3D geometries is largely unexplored. Here, we attempt to bridge this gap by proposing a novel method which converts molecular geometries into SE(3)-invariant 1D discrete sequences. Our method consists of canonical labeling and invariant spherical representation steps, which together maintain geometric and atomic fidelity in a format conducive to LMs. Our experiments show that, when coupled with our proposed method, various LMs excel in molecular geometry generation, especially in controlled generation tasks. Our code has been released as part of the AIRS library (https://github.com/divelab/AIRS/).

Retrieval-augmented generation (RAG) enhances language models by integrating external knowledge, but its effectiveness is highly dependent on system configuration. Improper retrieval settings can degrade performance, making RAG less reliable than closed-book generation. In this work, we introduce RAGGED, a framework for systematically evaluating RAG systems across diverse retriever-reader configurations, retrieval depths, and datasets. Our analysis reveals that reader robustness to noise is the key determinant of RAG stability and scalability. Some readers benefit from increased retrieval depth, while others degrade due to their sensitivity to distracting content. Through large-scale experiments on open-domain, multi-hop, and specialized-domain datasets, we show that retrievers, rerankers, and prompts influence performance but do not fundamentally alter these reader-driven trends. By providing a principled framework and new metrics to assess RAG stability and scalability, RAGGED enables systematic evaluation of retrieval-augmented generation systems, guiding future research on optimizing retrieval depth and model robustness.

This paper introduces a checks-and-balances framework for ethical alignment of Large Language Models (LLMs), inspired by three-branch governmental systems. It implements three independent yet interacting components: LLMs as the executive branch for knowledge generation, DIKE as the legislative branch establishing ethical guardrails, and ERIS as the judicial branch for contextual interpretation. Beyond structural separation, we address a fundamental challenge: regulating emotion to shape behaviors. Drawing from psychological theories where managing emotional responses prevents harmful behaviors, we develop a self-supervised learning pipeline that maps emotions to linguistic behaviors, enabling precise behavioral modulation through emotional conditioning. By integrating this approach with adversarial testing, our framework demonstrates how DIKE and ERIS direct linguistic behaviors toward ethical outcomes while preserving independence throughout knowledge generation, ethical oversight, and contextual interpretation.

Chain-of-Thought (CoT) prompting enhances mathematical reasoning in large language models (LLMs) by enabling detailed step-by-step solutions. However, due to the verbosity of LLMs, the resulting reasoning chains can be long, making it harder to verify the reasoning steps and trace issues resulting from dependencies between the steps that may be farther away in the sequence of steps. Importantly, mathematical reasoning allows each step to be derived from a small set of premises, which are a subset of the preceding steps in the reasoning chain. In this paper, we present a framework that identifies the premises for each step, to improve the evaluation of reasoning. We restructure conventional linear reasoning chains into Premise Augmented Reasoning Chains (PARC) by introducing premise links, resulting in a directed acyclic graph where the nodes are the steps and the edges are the premise links. Through experiments with a PARC-based dataset that we built, namely (Premises and ERrors identification in LLMs), we demonstrate that LLMs can reliably identify premises within complex reasoning chains. In particular, even open-source LLMs achieve 90% recall in premise identification. We also show that PARC helps to identify errors in reasoning chains more reliably. The accuracy of error identification improves by 6% to 16% absolute when step-by-step verification is carried out in PARC under the premises. Our findings highlight the utility of premise-centric representations in addressing complex problem-solving tasks and open new avenues for improving the reliability of LLM-based reasoning evaluations.

Text-to-image diffusion models have achieved remarkable progress in recent years. However, training models for high-resolution image generation remains challenging, particularly when training data and computational resources are limited. In this paper, we explore this practical problem from two key perspectives: data and parameter efficiency, and propose a set of key guidelines for ultra-resolution adaptation termed URAE. For data efficiency, we theoretically and empirically demonstrate that synthetic data generated by some teacher models can significantly promote training convergence. For parameter efficiency, we find that tuning minor components of the weight matrices outperforms widely-used low-rank adapters when synthetic data are unavailable, offering substantial performance gains while maintaining efficiency. Additionally, for models leveraging guidance distillation, such as FLUX, we show that disabling classifier-free guidance, i.e., setting the guidance scale to 1 during adaptation, is crucial for satisfactory performance. Extensive experiments validate that URAE achieves comparable 2K-generation performance to state-of-the-art closed-source models like FLUX1.1 [Pro] Ultra with only 3K samples and 2K iterations, while setting new benchmarks for 4K-resolution generation. Codes are available [here](https://github.com/Huage001/URAE).

Unsupervised skill discovery aims to learn diverse and distinguishable behaviors in open-ended reinforcement learning. For the existing methods, they focus on improving the diversity via pure exploration, mutual information optimization and learning temporal representation. Despite they perform well on exploration, they remain limited in terms of efficiency, especially for the high-dimensional situations. In this work, we frame the skill discovery as a min-max game of skill generation and policy learning, proposing a regret-aware method on top of temporal representation learning that expands the discovered skill space along the direction of upgradable policy strength. The key insight behind the proposed method is that the skill discovery is adversarial to the policy learning, i.e., skills with weak strength should be further explored while less exploration for the skills with converged strength. As an implementation, we score the degree of strength convergence with regret, and guide the skill discovery with a learnable skill generator. To avoid degeneration, the skill generation comes from an upgradable population of skill generators. We conduct experiments on environments with varying complexities and dimension sizes. Empirical results show that our method outperforms baselines on both efficiency and diversity. Moreover, our method achieves 15\% zero-shot improvement on high-dimensional environments, compared to existing methods.

The Conditional Sequence Modeling (CSM) paradigm, benefiting from the transformer's powerful distribution modeling capabilities, has demonstrated considerable promise in offline Reinforcement Learning (RL) tasks. Depending on the task's nature, it is crucial to carefully balance the interplay between inherent local features and long-term dependencies in Markov decision trajectories to mitigate potential performance degradation and unnecessary computational overhead. In this paper, we propose Decision Mixer (DM), which addresses the conflict between features of different scales in the modeling process from the perspective of dynamic integration. Drawing inspiration from conditional computation, we design a plug-and-play dynamic token selection mechanism to ensure the model can effectively allocate attention to different features based on task characteristics. Additionally, we employ an auxiliary predictor to alleviate the short-sightedness issue in the autoregressive sampling process. DM achieves state-of-the-art performance on various standard RL benchmarks while requiring significantly fewer computational resources, offering a viable solution for building efficient and scalable RL foundation models. Code is available at here.

Simulation modeling offers a flexible approach to constructing high-fidelity synthetic representations of complex real-world systems. However, the increased complexity of such models introduces additional complications, for example when carrying out statistical inference procedures. This has motivated a large and growing literature on *likelihood-free* or *simulation-based* inference methods, which approximate (e.g., Bayesian) inference without assuming access to the simulator's intractable likelihood function. A hitherto neglected problem in the simulation-based Bayesian inference literature is the challenge of constructing minimally informative *reference priors* for complex simulation models. Such priors maximise an expected Kullback-Leibler distance from the prior to the posterior, thereby influencing posterior inferences minimally and enabling an ``objective'' approach to Bayesian inference that does not necessitate the incorporation of strong subjective prior beliefs. In this paper, we propose and test a selection of likelihood-free methods for learning reference priors for simulation models, using variational approximations to these priors and a variety of mutual information estimators. Our experiments demonstrate that good approximations to reference priors for simulation models are in this way attainable, providing a first step towards the development of likelihood-free objective Bayesian inference procedures.

Many federated learning scenarios encounter label noises in the client-side datasets. The resulting degradation in global model performance raises the urgent need to address label noise. This paper proposes FedClean -- a novel general robust label noise correction for federated learning. FedClean first uses the local centralized noisy label learning to select clean samples to train a global model. Then, it employs a two-stage correction scheme to correct the noisy labels from two distinct perspectives of local noisy label learning and the global model. FedClean also proposes a novel model aggregation method, further reducing the impact of label noises. FedClean neither assumes the existence of clean clients nor the specific noise distributions, showing the maximum versatility. Extensive experimental results show that FedClean effectively identifies and rectifies label noises even if all clients exhibit label noises, which outperforms the state-of-the-art noise-label learning methods for federated learning.

List learning is an important topic in both theoretical and empirical machine learning research, playing a key role in the recent breakthrough result of (Brukhim et al., 2022) on the characterization of multiclass PAC learnability, as well as addressing label ambiguity in computer vision classification tasks, among others. In this paper, we study the problem of list transductive online learning. In this framework, the learner outputs a list of multiple labels for each instance rather than just one, as in traditional multiclass classification. In the realizable setting, we demonstrate a trichotomy of possible rates of the minimax number of mistakes. In particular, if the learner plays for $\text{T} \in \mathbb{N}$ rounds, its minimax number of mistakes can only be of the orders $\Theta(\text{T})$, $\Theta(\log \text{T})$, or $\Theta(1)$. This resolves an open question raised by (Hanneke et al., 2024). On the other hand, in the agnostic setting, we characterize the learnability by constructively proving the $\widetilde{\mathcal{O}}(\sqrt{\text{T}})$ upper bound on the minimax expected regret. Along the way, we also answer another open question asked by (Moran et al., 2023). To establish these results, we introduce two new combinatorial complexity dimensions, called the Level-constrained $(\mathrm{L+1})$-Littlestone dimension and Level-constrained $(\mathrm{L+1})$-Branching dimension, if the list size is $\mathrm{L} \in \mathbb{N}$. Eventually, we conclude our work by raising an open question regarding eliminating the factor list size, which seems to be a crucial step, as it has consistently appeared in previous works on this subject.

Cross-domain few-shot learning (CDFSL) aims to transfer knowledge from a data-sufficient source domain to data-scarce target domains. Although Vision Transformer (ViT) has shown superior capability in many vision tasks, its transferability against huge domain gaps in CDFSL is still under-explored. In this paper, we find an intriguing phenomenon: during the source-domain training, prompt tuning, as a common way to train ViT, could be harmful for the generalization of ViT in target domains, but setting them to random noises (i.e., random registers) could consistently improve target-domain performance. We then delve into this phenomenon for an interpretation. We find that learnable prompts capture domain information during the training on the source dataset, which views irrelevant visual patterns as vital cues for recognition. This can be viewed as a kind of overfitting and increases the sharpness of the loss landscapes. In contrast, random registers are essentially a novel way of perturbing attention for the sharpness-aware minimization, which helps the model find a flattened minimum in loss landscapes, increasing the transferability. Based on this phenomenon and interpretation, we further propose a simple but effective approach for CDFSL to enhance the perturbation on attention maps by adding random registers on the semantic regions of image tokens, improving the effectiveness and efficiency of random registers. Extensive experiments on four benchmarks validate our rationale and state-of-the-art performance. Codes and models are available at https://github.com/shuaiyi308/REAP.

Many generative applications, such as synthesis-based 3D molecular design, involve constructing compositional objects with continuous features. Here, we introduce Compositional Generative Flows (CGFlow), a novel framework that extends flow matching to generate objects in compositional steps while modeling continuous states. Our key insight is that modeling compositional state transitions can be formulated as a straightforward extension of the flow matching interpolation process. We further build upon the theoretical foundations of generative flow networks (GFlowNets), enabling reward-guided sampling of compositional structures. We apply CGFlow to synthesizable drug design by jointly designing the molecule's synthetic pathway with its 3D binding pose. Our approach achieves state-of-the-art binding affinity and synthesizability on all 15 targets from the LIT-PCBA benchmark, and 4.2x improvement in sampling efficiency compared to 2D synthesis-based baseline. To our best knowledge, our method is also the first to achieve state of-art-performance in both Vina Dock (-9.42) and AiZynth success rate (36.1\%) on the CrossDocked2020 benchmark.

Large language models (LLMs) have demonstrated the capability to perform in-context learning (ICL) for completely unseen tasks in classification or language completion. Sequence to sequence (seq2seq) is another popular task category with several applications seeking quick adaptation with ICL. We present a systematic analysis of the ICL capability of LLMs on Seq2Seq tasks using a formal structured language-pair. Our study reveals a critical limitation: except for very short input sequences, ICL fails to achieve consistent learning across all output positions. This exposes a fundamental weakness of modern LLMs — their inability to effectively uncover the alignment between input and output sequences. Consequently, this limitation results in incomplete induction heads, which are essential for in-context learning of new discrete mappings. To address these limitations, we propose ICA-Tune, a method for focused fine-tuning of an LLM using in-context examples. We present a mechanistic evaluation with two accuracy probes to show how input-output alignment emerges in middle layers of an LLM without direct supervision. This alignment leads to an abrupt jump in the completeness of the induction heads in higher layers. We show that, compared to standard fine-tuning, ICA-Tune enables more sample efficient learning and better generalization to OOD instances.

Recent advancements in computer vision have seen Implicit Neural Representations (INR) becoming a dominant representation form for data due to their compactness and expressive power. To solve various vision tasks with INR data, vision networks can either be purely INR-based, but are thereby limited by simplistic operations and performance constraints, or include raster-based methods, which then tend to lose crucial structural information of the INR during the conversion process. To address these issues, we propose DVI, a novel Derivative-based Vision network for INR, capable of handling a variety of vision tasks across various data modalities, while achieving the best performance among the existing methods by incorporating state of the art raster-based methods into a INR based architecture. DVI excels by extracting semantic information from the high order derivative map of the INR, then seamlessly fusing it into a pre-existing raster-based vision network, enhancing its performance with deeper, task-relevant semantic insights. Extensive experiments on five vision tasks across three data modalities demonstrate DVI's superiority over existing methods. Additionally, our study encompasses comprehensive ablation studies to affirm the efficacy of each element of DVI, the influence of different derivative computation techniques and the impact of derivative orders. Reproducible codes are provided in the supplementary materials.