ICML 2025 论文列表

With the rapid expansion in the scale of large language models (LLMs), enabling efficient distributed inference across multiple computing units has become increasingly critical. However, communication overheads from popular distributed inference techniques such as Tensor Parallelism pose a significant challenge to achieve scalability and low latency. Therefore, we introduce a novel optimization technique, Sync-Point Drop (SPD), to reduce communication overheads in tensor parallelism by selectively dropping synchronization on attention outputs. In detail, we first propose a block design that allows execution to proceed without communication through SPD. Second, we apply different SPD strategies to attention blocks based on their sensitivity to the model accuracy. The proposed methods effectively alleviate communication bottlenecks while minimizing accuracy degradation during LLM inference, offering a scalable solution for diverse distributed environments: SPD offered about 20\% overall inference latency reduction with < 1\% accuracy regression for LLaMA2-70B inference over 8 GPUs.

Large Language Models (LLMs) solve complex problems using training-free methods like prompt engineering and in-context learning, yet ensuring reasoning correctness remains challenging. While self-correction methods such as self-consistency and self-refinement aim to improve reliability, they often reinforce biases due to the lack of effective feedback mechanisms. Multi-Agent Debate (MAD) has emerged as an alternative, but we identify two key limitations: bias reinforcement, where debate amplifies model biases instead of correcting them, and lack of perspective diversity, as all agents share the same model and reasoning patterns, limiting true debate effectiveness. To systematically evaluate these issues, we introduce $\textit{MetaNIM Arena}$, a benchmark designed to assess LLMs in adversarial strategic decision-making, where dynamic interactions influence optimal decisions. To overcome MAD’s limitations, we propose $\texttt{\textbf{DReaMAD}}$ ($\textbf{D}$iverse $\textbf{Rea}$soning via $\textbf{M}$ulti-$\textbf{A}$gent $\textbf{D}$ebate with Refined Prompt), a novel framework that (1) refines LLMs’ strategic prior knowledge to improve reasoning quality and (2) promotes diverse viewpoints within a single model by systematically modifying prompts, reducing bias. Empirical results show that $\texttt{\textbf{DReaMAD}}$ significantly improves decision accuracy, reasoning diversity, and bias mitigation across multiple strategic tasks, establishing it as a more effective approach for LLM-based decision-making.

Individualized treatment rules/recommendations (ITRs) aim to improve patient outcomes by tailoring treatments to the characteristics of each individual. However, in high-dimensional treatment settings, existing methods face significant challenges due to data sparsity within treatment groups and highly unbalanced covariate distributions across groups. To address these challenges, we propose a novel calibration-weighted treatment fusion procedure that robustly balances covariates across treatment groups and fuses similar treatments using a penalized working model. The fusion procedure ensures the recovery of latent treatment group structures when either the calibration model or the outcome model is correctly specified. In the fused treatment space, practitioners can seamlessly apply state-of-the-art ITR learning methods with the flexibility to utilize a subset of covariates, thereby achieving robustness while addressing practical concerns such as fairness. We establish theoretical guarantees, including consistency, the oracle property of treatment fusion, and regret bounds when integrated with multi-armed ITR learning methods such as policy trees. Simulation studies show superior group recovery and policy value compared to existing approaches. We illustrate the practical utility of our method using EHR-derived data from patients with Chronic Lymphocytic Leukemia and Small Lymphocytic Lymphoma.

Solving time-dependent parametric partial differential equations (PDEs) is challenging for data-driven methods, as these models must adapt to variations in parameters such as coefficients, forcing terms, and initial conditions. State-of-the-art neural surrogates perform adaptation through gradient-based optimization and meta-learning to implicitly encode the variety of dynamics from observations. This often comes with increased inference complexity. Inspired by the in-context learning capabilities of large language models (LLMs), we introduce Zebra, a novel generative auto-regressive transformer designed to solve parametric PDEs without requiring gradient adaptation at inference. By leveraging in-context information during both pre-training and inference, Zebra dynamically adapts to new tasks by conditioning on input sequences that incorporate context example trajectories. As a generative model, Zebra can be used to generate new trajectories and allows quantifying the uncertainty of the predictions. We evaluate Zebra across a variety of challenging PDE scenarios, demonstrating its adaptability, robustness, and superior performance compared to existing approaches.

Preference learning is a central component for aligning current LLMs, but this process can be vulnerable to data poisoning attacks. To address this concern, we introduce PoisonBench, a benchmark for evaluating large language models' susceptibility to data poisoning during preference learning. Data poisoning attacks can manipulate large language model responses to include hidden malicious content or biases, potentially causing the model to generate harmful or unintended outputs while appearing to function normally. We deploy two distinct attack types across eight realistic scenarios, assessing 22 widely-used models. Our findings reveal concerning trends: (1) Scaling up parameter size does not always enhance resilience against poisoning attacks and the influence on model resilience varies among different model suites. (2) There exists a log-linear relationship between the effects of the attack and the data poison ratio; (3) The effect of data poisoning can generalize to extrapolated triggers that are not included in the poisoned data. These results expose weaknesses in current preference learning techniques, highlighting the urgent need for more robust defenses against malicious models and data manipulation.

Supervised fine-tuning (SFT) on domain-specific data is the dominant approach for adapting foundation models to specialized tasks. However, it has been observed that SFT models tend to forget knowledge acquired during pretraining. In vision models, ensembling a pretrained model with its fine-tuned counterpart has been shown to mitigate this issue. In this work, we demonstrate that the same holds for language models, and, more strikingly, we observe an overadaptation phenomenon: the ensemble model not only retains general knowledge from the foundation model but also outperforms the fine-tuned model even on the fine-tuning domain itself. Despite the empirical success of ensembling, a theoretical understanding of its benefits remains underexplored. We develop a formal theoretical analysis of the overadaptation phenomenon. Ensembling mitigates this by balancing two primary sources of error: bias, caused by insufficient fine-tuning, and variance, introduced by overfitting to fine-tuning data. While regularization techniques aim to address this trade-off, we show that ensembling provides a more effective solution. We analyze this phenomenon in over-parameterized linear settings and demonstrate that interpolating between pretrained and fine-tuned weights significantly improves performance. These findings offer theoretical justification for the observed advantages of model ensembling, supported by empirical experiments consistent with our analysis.

The local inductive bias of message-passing graph neural networks (GNNs) hampers their ability to exploit key structural information (e.g., connectivity and cycles). Positional encoding (PE) and Persistent Homology (PH) have emerged as two promising approaches to mitigate this issue. PE schemes endow GNNs with location-aware features, while PH methods enhance GNNs with multiresolution topological features. However, a rigorous theoretical characterization of the relative merits and shortcomings of PE and PH has remained elusive. We bridge this gap by establishing that neither paradigm is more expressive than the other, providing novel constructions where one approach fails but the other succeeds. Our insights inform the design of a novel learnable method, PiPE (Persistence-informed Positional Encoding), which is provably more expressive than both PH and PE. PiPE demonstrates strong performance across a variety of tasks (e.g., molecule property prediction, graph classification, and out-of-distribution generalization), thereby advancing the frontiers of graph representation learning. Code is available at https://github.com/Aalto-QuML/PIPE

We propose a computationally efficient alternative to generalized random forests (GRFs) for estimating heterogeneous effects in large dimensions. While GRFs rely on a gradient-based splitting criterion, which in large dimensions is computationally expensive and unstable, our method introduces a fixed-point approximation that eliminates the need for Jacobian estimation. This gradient-free approach preserves GRF’s theoretical guarantees of consistency and asymptotic normality while significantly improving computational efficiency. We demonstrate that our method achieves a speedup of multiple times over standard GRFs without compromising statistical accuracy. Experiments on both simulated and real-world data validate our approach. Our findings suggest that the proposed method is a scalable alternative for localized effect estimation in machine learning and causal inference applications.

In-context learning has become a standard approach for utilizing language models. However, selecting and processing suitable demonstration examples can be challenging and time-consuming, especially when dealing with large numbers of them. We propose Iterative Vectors (IVs), a technique that explores activation space to enhance in-context performance by simulating gradient updates during inference. IVs extract and iteratively refine activation-based meta-gradients, applying them during inference without requiring backpropagation at any stage. We evaluate IVs across various tasks using four popular models and observe significant improvements. Our findings suggest that in-context activation steering is a promising direction, opening new avenues for future research.

The Machine Learning (ML) community has wit- nessed explosive growth, with millions of ML models being published on the Web. Reusing ML model components has been prevalent nowadays. Developers are often required to choose a license to publish and govern the use of their models. Popular options include Apache-2.0, OpenRAIL (Responsible AI Licenses), Creative Commons Licenses (CCs), Llama2, and GPL-3.0. Currently, no standard or widely accepted best practices ex- ist for model licensing. But does this lack of standardization lead to undesired consequences? Our answer is Yes. After reviewing the clauses of the most widely adopted licenses, we take the position that current model licensing practices are dragging us into a quagmire of legal noncom- pliance. To support this view, we explore the cur- rent practices in model licensing and highlight the differences between various model licenses. We then identify potential legal risks associated with these licenses and demonstrate these risks using examples from real-world repositories on Hug- ging Face. To foster a more standardized future for model licensing, we also propose a new draft of model licenses, ModelGo Licenses (MGLs), to address these challenges and promote better compliance. https://www.modelgo.li/

We propose a distillation scaling law that estimates distilled model performance based on a compute budget and its allocation between the student and teacher. Our findings mitigate the risks associated with large-scale distillation by enabling compute-optimal allocation for both the teacher and student to maximize student performance. We provide compute-optimal distillation recipes for two key scenarios: when a teacher already exists, and when a teacher needs training. In settings involving many students or an existing teacher, distillation outperforms supervised learning up to a compute level that scales predictably with student size. Conversely, if only one student is to be distilled and a teacher also requires training, supervised learning is generally preferable. Additionally, our large-scale study of distillation increases our understanding of the process and helps inform experimental design.

Actor-critic algorithms have become a cornerstone in reinforcement learning (RL), leveraging the strengths of both policy-based and value-based methods. Despite recent progress in understanding their statistical efficiency, no existing work has successfully learned an $\epsilon$-optimal policy with a sample complexity of $O(1/\epsilon^2)$ trajectories with general function approximation when strategic exploration is necessary. We address this open problem by introducing a novel actor-critic algorithm that attains a sample-complexity of $O(dH^5 \log|\mathcal{A}|/\epsilon^2 + d H^4 \log|\mathcal{F}|/ \epsilon^2)$ trajectories, and accompanying $\sqrt{T}$ regret when the Bellman eluder dimension $d$ does not increase with $T$ at more than a $\log T$ rate. Here, $\mathcal{F}$ is the critic function class, and $\mathcal{A}$ is the action space. Our algorithm integrates optimism, off-policy critic estimation targeting the optimal Q-function, and rare-switching policy resets. We extend this to the setting of Hybrid RL, where we show that initializing the critic with offline data yields sample efficiency gains, and also provide a \textit{non-optimistic} provably efficient actor-critic algorithm, addressing another open problem in the literature. Numerical experiments support our theoretical findings.

Integer-order calculus fails to capture the long-range dependence (LRD) and memory effects found in many complex systems. Fractional calculus addresses these gaps through fractional-order integrals and derivatives, but fractional-order dynamical systems pose substantial challenges in system identification and optimal control tasks. In this paper, we theoretically derive the optimal control via linear quadratic regulator (LQR) for fractional-order linear time-invariant (FOLTI) systems and develop an end-to-end deep learning framework based on this theoretical foundation. Our approach establishes a rigorous mathematical model, derives analytical solutions, and incorporates deep learning to achieve data-driven optimal control of FOLTI systems. Our key contributions include: (i) proposing a novel method for system identification and optimal control strategy in FOLTI systems, (ii) developing the first end-to-end data-driven learning framework, Fractional-Order Learning for Optimal Control (FOLOC), that learns control policies from observed trajectories, and (iii) deriving theoretical bounds on the sample complexity for learning accurate control policies under fractional-order dynamics. Experimental results indicate that our method accurately approximates fractional-order system behaviors without relying on Gaussian noise assumptions, pointing to promising avenues for advanced optimal control.

In this work, we study offline reinforcement learning (RL) with zero-shot generalization property (ZSG), where the agent has access to an offline dataset including experiences from different environments, and the goal of the agent is to train a policy over the training environments which performs well on test environments without further interaction. Existing work showed that classical offline RL fails to generalize to new, unseen environments. We propose pessimistic empirical risk minimization (PERM) and pessimistic proximal policy optimization (PPPO), which leverage pessimistic policy evaluation to guide policy learning and enhance generalization. We show that both PERM and PPPO are capable of finding a near-optimal policy with ZSG. Our result serves as a first step in understanding the foundation of the generalization phenomenon in offline reinforcement learning.

Supervised fine-tuning (SFT) has become a crucial step for aligning pretrained large language models (LLMs) using supervised datasets of input-output pairs. However, despite being supervised, SFT is inherently limited by its generative training objective. To address its limitations, the existing common strategy is to follow SFT with a separate phase of preference optimization (PO), which relies on either human-labeled preference data or a strong reward model to guide the learning process. In this paper, we address the limitations of SFT by exploring one of the most successful techniques in conventional supervised learning: discriminative learning. We introduce **Discriminative Fine-Tuning (DFT)**, an improved variant of SFT, which mitigates the burden of collecting human-labeled preference data or training strong reward models. Unlike SFT that employs a generative approach and overlooks negative data, DFT adopts a **discriminative paradigm** that increases the probability of positive answers while suppressing potentially negative ones, aiming for **data prediction** instead of token prediction. Our contributions include: (i) a discriminative probabilistic framework for fine-tuning LLMs by explicitly modeling the discriminative likelihood of an answer among all possible outputs given an input; (ii) efficient algorithms to optimize this discriminative likelihood; and (iii) extensive experiments demonstrating DFT's effectiveness, achieving performance better than SFT and comparable to if not better than SFT→PO. The code can be found at https://github.com/Optimization-AI/DFT.

Segment Anything Model (SAM) has demonstrated remarkable zero-shot segmentation capabilities across various visual tasks. However, its performance degrades significantly when deployed in new target domains with substantial distribution shifts. While existing self-training methods based on fixed teacher-student architectures have shown improvements, they struggle to ensure that the teacher network consistently outperforms the student under severe domain shifts. To address this limitation, we propose a novel Collaborative Mutual Learning Framework for source-free SAM adaptation, leveraging dual-networks in a dynamic and cooperative manner. Unlike fixed teacher-student paradigms, our method dynamically assigns the teacher and student roles by evaluating the reliability of each collaborative network in each training iteration. Our framework incorporates a dynamic mutual learning mechanism with three key components: a direct alignment loss for knowledge transfer, a reverse distillation loss to encourage diversity, and a triplet relationship loss to refine feature representations. These components enhance the adaptation capabilities of the collaborative networks, enabling them to generalize effectively to target domains while preserving their pre-trained knowledge. Extensive experiments on diverse target domains demonstrate that our proposed framework achieves state-of-the-art adaptation performance.

While multimodal contrastive learning methods (e.g., CLIP) can achieve impressive zero-shot classification performance, recent research has revealed that these methods are vulnerable to backdoor attacks. To defend against backdoor attacks on CLIP, existing defense methods focus on either the pre-training stage or the fine-tuning stage, which would unfortunately cause high computational costs due to numerous parameter updates and are not applicable in black-box settings. In this paper, we provide the first attempt at a computationally efficient backdoor detection method to defend against backdoored CLIP in the inference stage. We empirically find that the visual representations of backdoored images are insensitive to benign and malignant changes in class description texts. Motivated by this observation, we propose BDetCLIP, a novel test-time backdoor detection method based on contrastive prompting. Specifically, we first prompt a language model (e.g., GPT-4) to produce class-related description texts (benign) and class-perturbed random texts (malignant) by specially designed instructions. Then, the distribution difference in cosine similarity between images and the two types of class description texts can be used as the criterion to detect backdoor samples. Extensive experiments validate that our proposed BDetCLIP is superior to state-of-the-art backdoor detection methods, in terms of both effectiveness and efficiency.

Large language models (LLMs) acquire information from pre-training corpora, but their stored knowledge can become inaccurate or outdated over time. Model editing addresses this challenge by modifying model parameters without retraining, and prevalent approaches leverage hypernetworks to generate these parameter updates. However, they face significant challenges in lifelong editing due to their incompatibility with LLM parameters that dynamically change during the editing process. To address this, we observed that hypernetwork-based lifelong editing aligns with reinforcement learning modeling and proposed **RLEdit**, an RL-based editing method. By treating editing losses as rewards and optimizing hypernetwork parameters at the full knowledge sequence level, we enable it to precisely capture LLM changes and generate appropriate parameter updates. Our extensive empirical evaluation across several LLMs demonstrates that RLEdit outperforms existing methods in lifelong editing with superior effectiveness and efficiency, achieving a **59.24%** improvement while requiring only **2.11%** of the time compared to most approaches.

Multi-objective optimization problems arise widely in various fields. In practice, multi-objective optimization is generally solved by heuristics with tunable parameters that are highly application-specific. Tuning parameters based on real-world instances (a.k.a. algorithm configuration) are generally empirical without theoretical guarantees. In this work, we establish the theoretical foundation of data-driven multi-objective optimization through the lens of machine learning theory. We provide generalization guarantees on selecting parameters for multi-objective optimization algorithms based on sampled problem instances. Moreover, if the performance metric of the algorithm is the Pareto volume, we can PAC-learn the approximately optimal configuration in polynomial time. We apply our framework to various algorithms, including approximation algorithms, local search, and linear programming. Experiments on multiple problems verify our theoretical findings.