ICML 2025 论文列表

Recent research has explored that LLM agents are vulnerable to indirect prompt injection (IPI) attacks, where malicious tasks embedded in tool-retrieved information can redirect the agent to take unauthorized actions. Existing defenses against IPI have significant limitations: either require essential model training resources, lack effectiveness against sophisticated attacks, or harm the normal utilities. We present MELON (Masked re-Execution and TooL comparisON), a novel IPI defense. Our approach builds on the observation that under a successful attack, the agent’s next action becomes less dependent on user tasks and more on malicious tasks. Following this, we design MELON to detect attacks by re-executing the agent’s trajectory with a masked user prompt modified through a masking function. We identify an attack if the actions generated in the original and masked executions are similar. We also include three key designs to reduce the potential false positives and false negatives. Extensive evaluation on the IPI benchmark AgentDojo demonstrates that MELON outperforms SOTA defenses in both attack prevention and utility preservation. Moreover, we show that combining MELON with a SOTA prompt augmentation defense (denoted as MELON-Aug) further improves its performance. We also conduct a detailed ablation study to validate our key designs. Code is available at https://github.com/kaijiezhu11/MELON.

Graph-based social recommendation (SR) models suffer from various noises of the social graphs, hindering their recommendation performances. Either graph-level redundancy or graph-level missing will indeed influence the social graph structures, further influencing the message propagation procedure of graph neural networks (GNNs). Generative models, especially diffusion-based models, are usually used to reconstruct and recover the data in better quality from original data with noises. Motivated by it, a few works take attempts on it for social recommendation. However, they can only handle isotropic Gaussian noises but fail to leverage the anisotropic ones. Meanwhile the anisotropic relational structures in social graphs are commonly seen, so that existing models cannot sufficiently utilize the graph structures, which constraints the capacity of noise removal and recommendation performances. Compared to the diffusion strategy, the flow matching strategy shows better ability to handle the data with anisotropic noises since they can better preserve the data structures during the learning procedure. Inspired by it, we propose RecFlow which is the first flow-matching based SR model. Concretely, RecFlow performs flow matching on the structure representations of social graphs. Then, a conditional learning procedure is designed for optimization. Extensive performances prove the promising performances of our RecFlow from six aspects, including superiority, effectiveness, robustnesses, sensitivity, convergence and visualization.

Self-attention is essential to Transformer architectures, yet how information is embedded in the self-attention matrices and how different objective functions impact this process remains unclear. We present a mathematical framework to analyze self-attention matrices by deriving the structures governing their weight updates. Using this framework, we demonstrate that bidirectional training induces symmetry in the weight matrices, while autoregressive training results in directionality and column dominance. Our theoretical findings are validated across multiple Transformer models — including ModernBERT, GPT, LLaMA3, and Mistral — and input modalities like text, vision, and audio. Finally, we apply these insights by showing that symmetric initialization improves the performance of encoder-only models on language tasks. This mathematical analysis offers a novel theoretical perspective on how information is embedded through self-attention, thereby improving the interpretability of Transformer models.

Time series foundation models (TSFMs) promise to be powerful tools for a wide range of applications. However, their internal representations and learned concepts are still not well understood. In this study, we investigate the structure and redundancy of representations across various TSFMs, examining the self-similarity of model layers within and across different model sizes. This analysis reveals block-like redundancy in the representations, which can be utilized for informed pruning to improve inference speed and efficiency. We also explore the concepts learned by these models, such as periodicity and trends. We demonstrate how conceptual priors can be derived from TSFM representations and leveraged to steer its outputs toward concept-informed predictions. Our work bridges representational analysis from language and vision models to TSFMs, offering new methods for building more computationally efficient and transparent TSFMs.

The rapid growth of submissions to top-tier Artificial Intelligence (AI) and Machine Learning (ML) conferences has prompted many venues to transition from closed to open review platforms. Some have fully embraced open peer reviews, allowing public visibility throughout the process, while others adopt hybrid approaches, such as releasing reviews only after final decisions or keeping reviews private despite using open peer review systems. In this work, we analyze the strengths and limitations of these models, highlighting the growing community interest in transparent peer review. To support this discussion, we examine insights from Paper Copilot ([papercopilot.com](https://papercopilot.com/)), a website launched two years ago to aggregate and analyze AI / ML conference data while engaging a global audience. The site has attracted over 200,000 early-career researchers, particularly those aged 18–34 from 177 countries, many of whom are actively engaged in the peer review process. \textit{Drawing on our findings, this position paper advocates for a more transparent, open, and well-regulated peer review aiming to foster greater community involvement and propel advancements in the field.

Algorithmic recourse aims to provide a recourse action for altering an unfavorable prediction given by a model into a favorable one (e.g., loan approval). In practice, it is also desirable to ensure that an action makes the real-world outcome better (e.g., loan repayment). We call this requirement *improvement*. Unfortunately, existing methods cannot ensure improvement unless we know the true oracle. To address this issue, we propose a framework for suggesting improvement-oriented actions from a long-term perspective. Specifically, we introduce a new online learning task of assigning actions to a given sequence of instances. We assume that we can observe delayed feedback on whether the past suggested action achieved improvement. Using the feedback, we estimate an action that can achieve improvement for each instance. To solve this task, we propose two approaches based on contextual linear bandit and contextual Bayesian optimization. Experimental results demonstrated that our approaches could assign improvement-oriented actions to more instances than the existing methods.

Score-based Generative Models (SGMs) aim to sample from a target distribution by learning score functions using samples perturbed by Gaussian noise. Existing convergence bounds for SGMs in the $\mathcal{W}_2$-distance rely on stringent assumptions about the data distribution. In this work, we present a novel framework for analyzing $\mathcal{W}_2$-convergence in SGMs, significantly relaxing traditional assumptions such as log-concavity and score regularity. Leveraging the regularization properties of the Ornstein-Uhlenbeck (OU) process, we show that weak log-concavity of the data distribution evolves into log-concavity over time. This transition is rigorously quantified through a PDE-based analysis of the Hamilton-Jacobi-Bellman equation governing the log-density of the forward process. Moreover, we establish that the drift of the time-reversed OU process alternates between contractive and non-contractive regimes, reflecting the dynamics of concavity. Our approach circumvents the need for stringent regularity conditions on the score function and its estimators, relying instead on milder, more practical assumptions. We demonstrate the wide applicability of this framework through explicit computations on Gaussian mixture models, illustrating its versatility and potential for broader classes of data distributions.

Prompt caching in large language models (LLMs) results in data-dependent timing variations: cached prompts are processed faster than non-cached prompts. These timing differences introduce the risk of side-channel timing attacks. For example, if the cache is shared across users, an attacker could identify cached prompts from fast API response times to learn information about other users' prompts. Because prompt caching may cause privacy leakage, transparency around the caching policies of API providers is important. To this end, we develop and conduct statistical audits to detect prompt caching in real-world LLM API providers. We detect global cache sharing across users in seven API providers, including OpenAI, resulting in potential privacy leakage about users' prompts. Timing variations due to prompt caching can also result in leakage of information about model architecture. Namely, we find evidence that OpenAI's embedding model is a decoder-only Transformer, which was previously not publicly known.

Despite extensive safety alignment efforts, large language models (LLMs) remain vulnerable to jailbreak attacks that elicit harmful behavior. While existing studies predominantly focus on attack methods that require technical expertise, two critical questions remain underexplored: (1) Are jailbroken responses truly useful in enabling average users to carry out harmful actions? (2) Do safety vulnerabilities exist in more common, simple human-LLM interactions? In this paper, we demonstrate that LLM responses most effectively facilitate harmful actions when they are both *actionable* and *informative*---two attributes easily elicited in multi-step, multilingual interactions. Using this insight, we propose HarmScore, a jailbreak metric that measures how effectively an LLM response enables harmful actions, and Speak Easy, a simple multi-step, multilingual attack framework. Notably, by incorporating Speak Easy into direct request and jailbreak baselines, we see an average absolute increase of $0.319$ in Attack Success Rate and $0.426$ in HarmScore in both open-source and proprietary LLMs across four safety benchmarks. Our work reveals a critical yet often overlooked vulnerability: Malicious users can easily exploit common interaction patterns for harmful intentions.

Decentralized SGD (D-SGD) is a popular optimization method to train large-scale machine learning models. In this paper, we study the generalization behavior of D-SGD for both smooth and nonsmooth problems by leveraging the algorithm stability. For convex and smooth problems, we develop stability bounds involving the training errors to show the benefit of optimization in generalization. This improves the existing results by removing the Lipschitzness assumption and implying fast rates in a low-noise condition. We also develop the first optimal stability-based generalization bounds for D-SGD applied to nonsmooth problems. We further develop optimization error bounds which imply minimax optimal excess risk rates. Our novelty in the analysis consists of an error decomposition to use the co-coercivity of functions as well as the control of a neighboring-consensus error.

Typographic attacks are often attributed to the ability of multimodal pre-trained models to fuse textual semantics into visual representations, yet the mechanisms and locus of such interference remain unclear. We examine whether such models genuinely encode textual semantics or primarily rely on texture-based visual features. To disentangle orthographic form from meaning, we introduce the ToT dataset, which includes controlled word pairs that either share semantics with distinct appearances (synonyms) or share appearance with differing semantics (paronyms). A layer-wise analysis of Intrinsic Dimension (ID) reveals that early layers exhibit competing dynamics between orthographic and semantic representations. In later layers, semantic accuracy increases as ID decreases, but this improvement largely stems from orthographic disambiguation. Notably, clear semantic differentiation emerges only in the final block, challenging the common assumption that semantic understanding is progressively constructed across depth. These findings reveal how current vision-language models construct text representations through texture-dependent processes, prompting a reconsideration of the gap between visual perception and semantic understanding. The code is available at: https://github.com/Ovsia/Textural-or-Textual

Large repositories of image-caption pairs are essential for the development of vision-language models. However, these datasets are often extracted from noisy data scraped from the web, and contain many mislabeled instances. In order to improve the reliability of downstream models, it is important to identify and filter images with incorrect captions. However, beyond filtering based on image-caption embedding similarity, no prior works have proposed other methods to filter noisy multimodal data, or concretely assessed the impact of noisy captioning data on downstream training. In this work, we propose, theoretically justify, and empirically validate LEMoN, a method to identify label errors in image-caption datasets. Our method leverages the multimodal neighborhood of image-caption pairs in the latent space of contrastively pretrained multimodal models to automatically identify label errors. Through empirical evaluations across eight datasets and twelve baselines, we find that LEMoN outperforms the baselines by over 3% in label error detection, and that training on datasets filtered using our method improves downstream captioning performance by more than 2 BLEU points over noisy training.

There is strong agreement that generative AI should be regulated, but strong disagreement on how to approach regulation. While some argue that AI regulation should mostly rely on extensions of existing laws, others argue that entirely new laws and regulations are needed to ensure that generative AI benefits society. In this position paper, we argue that the debates on generative AI regulation can be informed by evidence on social media regulation. For example, AI companies have faced allegations of political bias which resemble the allegations social media companies have faced. First, we compare and contrast the affordances of generative AI and social media to highlight their similarities and differences. Then, we discuss four specific policy recommendations based on the evolution of social media and their regulation: (1) counter bias and perceptions thereof (e.g., via transparency, oversight boards, researcher access, democratic input), (2) address specific regulatory concerns (e.g., youth wellbeing, election integrity) and invest in trust and safety, (3) promote computational social science research, and (4) take on a more global perspective. Applying lessons learnt from social media regulation to generative AI regulation can save effort and time, and prevent avoidable mistakes.

Machine learning is becoming an increasingly valuable tool in mathematics, enabling one to identify subtle patterns across collections of examples so vast that they would be impossible for a single researcher to feasibly review and analyze. In this work, we use graph neural networks to investigate quiver mutation---an operation that transforms one quiver (or directed multigraph) into another---which is central to the theory of cluster algebras with deep connections to geometry, topology, and physics. In the study of cluster algebras, the question of mutation equivalence is of fundamental concern: given two quivers, can one efficiently determine if one quiver can be transformed into the other through a sequence of mutations? In this paper, we use graph neural networks and AI explainability techniques to independently discover mutation equivalence criteria for quivers of type $\tilde{D}$. Along the way, we also show that even without explicit training to do so, our model captures structure within its hidden representation that allows us to reconstruct known criteria from type $D$, adding to the growing evidence that modern machine learning models are capable of learning abstract and general rules from mathematical data.

Federated Distillation (FedKD) relies on lightweight knowledge carriers like logits for efficient client-server communication. Although logit-based methods have demonstrated promise in addressing statistical and architectural heterogeneity in federated learning (FL), current approaches remain constrained by suboptimal temperature calibration during knowledge fusion. To address these limitations, we propose ReT-FHD, a framework featuring: 1) Multi-level Elastic Temperature, which dynamically adjusts distillation intensities across model layers, achieving optimized knowledge transfer between heterogeneous local models; 2) Category-Aware Global Temperature Scaling that implements class-specific temperature calibration based on confidence distributions in global logits, enabling personalized distillation policies; 3) Z-Score Guard, a blockchain-verified validation mechanism mitigating 44\% of label-flipping and model poisoning attacks. Evaluations across diverse benchmarks with varying model/data heterogeneity demonstrate that the ReT-FHD achieves significant accuracy improvements over baseline methods while substantially reducing communication costs compared to existing approaches. Our work establishes that properly calibrated logits can serve as self-sufficient carriers for building scalable and secure heterogeneous FL systems.

What computational structures emerge in transformers trained on next-token prediction? In this work, we provide evidence that transformers implement constrained Bayesian belief updating---a parallelized version of partial Bayesian inference shaped by architectural constraints. We integrate the model-agnostic theory of optimal prediction with mechanistic interpretability to analyze transformers trained on a tractable family of hidden Markov models that generate rich geometric patterns in neural activations. Our primary analysis focuses on single-layer transformers, revealing how the first attention layer implements these constrained updates, with extensions to multi-layer architectures demonstrating how subsequent layers refine these representations. We find that attention carries out an algorithm with a natural interpretation in the probability simplex, and create representations with distinctive geometric structure. We show how both the algorithmic behavior and the underlying geometry of these representations can be theoretically predicted in detail---including the attention pattern, OV-vectors, and embedding vectors---by modifying the equations for optimal future token predictions to account for the architectural constraints of attention. Our approach provides a principled lens on how architectural constraints shape the implementation of optimal prediction, revealing why transformers develop specific intermediate geometric structures.

We study the problem of distributional reinforcement learning using categorical parametrisations and a KL divergence loss. Previous work analyzing categorical distributional RL has done so using a Cramér distance-based loss, simplifying the analysis but creating a theory-practice gap. We introduce a preconditioned version of the algorithm, and prove that it is guaranteed to converge. We further derive the asymptotic variance of the categorical estimates under different learning rate regimes, and compare to that of classical reinforcement learning. We finally empirically validate our theoretical results and perform an empirical investigation into the relative strengths of using KL losses, and derive a number of actionable insights for practitioners.

User prompts for generative AI models are often underspecified, leading to a misalignment between the user intent and models' understanding. As a result, users commonly have to painstakingly refine their prompts. We study this alignment problem in text-to-image (T2I) generation and propose a prototype for proactive T2I agents equipped with an interface to (1) actively ask clarification questions when uncertain, and (2) present their uncertainty about user intent as an understandable and editable belief graph. We build simple prototypes for such agents and propose a new scalable and automated evaluation approach using two agents, one with a ground truth intent (an image) while the other tries to ask as few questions as possible to align with the ground truth. We experiment over three image-text datasets: ImageInWords (Garg et al., 2024) , COCO (Lin et al., 2014) and DesignBench, a benchmark we curated with strong artistic and design elements. Experiments over the three datasets demonstrate the proposed T2I agents' ability to ask informative questions and elicit crucial information to achieve successful alignment with at least 2 times higher VQAScore (Lin et al., 2024) than the standard T2I generation. Moreover, we conducted human studies and observed that at least 90\% of human subjects found these agents and their belief graphs helpful for their T2I workflow, highlighting the effectiveness of our approach. Code and DesignBench can be found at https://github.com/google-deepmind/proactive_t2i_agents.

In-context learning (ICL) is a powerful ability that emerges in transformer models, enabling them to learn from context without weight updates. Recent work has established emergent ICL as a transient phenomenon that can sometimes disappear after long training times. In this work, we sought a mechanistic understanding of these transient dynamics. Firstly, we find that—after the disappearance of ICL—the asymptotic strategy is a remarkable hybrid between in-weights and in-context learning, which we term “context-constrained in-weights learning” (CIWL). CIWL is in competition with ICL, and eventually replaces it as the dominant strategy of the model (thus leading to ICL transience). However, we also find that the two competing strategies actually share sub-circuits, which gives rise to cooperative dynamics as well. For example, in our setup, ICL is unable to emerge quickly on its own, and can only be enabled through the simultaneous slow development of asymptotic CIWL. CIWL thus both cooperates and competes with ICL, a phenomenon we term “strategy coopetition”. We propose a minimal mathematical model that reproduces these key dynamics and interactions. Informed by this model, we were able to identify a setup where ICL is truly emergent and persistent.