ICML 2025 论文列表

We consider the problem of selecting instrumental variables from observational data, a fundamental challenge in causal inference. Existing methods mostly focus on additive linear, constant effects models, limiting their applicability in complex real-world scenarios. In this paper, we tackle a more general and challenging setting: the additive non-linear, constant effects model. We first propose a novel testable condition, termed the Cross Auxiliary-based independent Test (CAT) condition, for selecting the valid IV set. We show that this condition is both necessary and sufficient for identifying valid instrumental variable sets within such a model under milder assumptions. Building on this condition, we develop a practical algorithm for selecting the set of valid instrumental variables. Extensive experiments on both synthetic and two real-world datasets demonstrate the effectiveness and robustness of our proposed approach, highlighting its potential for broader applications in causal analysis.

Feature learning in neural networks is crucial for their expressive power and inductive biases, motivating various theoretical approaches. Some approaches describe network behavior after training through a change in kernel scale from initialization, resulting in a generalization power comparable to a Gaussian process. Conversely, in other approaches training results in the adaptation of the kernel to the data, involving directional changes to the kernel. The relationship and respective strengths of these two views have so far remained unresolved. This work presents a theoretical framework of multi-scale adaptive feature learning bridging these two views. Using methods from statistical mechanics, we derive analytical expressions for network output statistics which are valid across scaling regimes and in the continuum between them. A systematic expansion of the network's probability distribution reveals that mean-field scaling requires only a saddle-point approximation, while standard scaling necessitates additional correction terms. Remarkably, we find across regimes that kernel adaptation can be reduced to an effective kernel rescaling when predicting the mean network output in the special case of a linear network. However, for linear and non-linear networks, the multi-scale adaptive approach captures directional feature learning effects, providing richer insights than what could be recovered from a rescaling of the kernel alone.

Training data quality is one of the most important drivers of final model quality. In this work, we introduce a method for evaluating data integrity based on the assumption that low-quality input prompts result in high variance and low quality responses. This is achieved by measuring the rejected response quality and the reward gap between the chosen and rejected preference pair. Our method, Rejecting Instruction Preferences (RIP) can be used to filter prompts from existing training sets, or to make high quality synthetic datasets, yielding large performance gains across various benchmarks compared to unfiltered data. Using Llama 3.1-8B-Instruct, RIP improves AlpacaEval2 LC Win Rate by 9.4%, Arena-Hard by 8.7%, and WildBench by 9.9%. Using Llama 3.3-70B-Instruct, RIP improves Arena-Hard from 67.5 to 82.9, from 18th place to 6th overall in the leaderboard.

Different modalities hold considerable gaps in optimization trajectories, including speeds and paths, which lead to *modality laziness* and *modality clash* when jointly training multimodal models, resulting in insufficient and imbalanced multimodal learning. Existing methods focus on enforcing the weak modality by adding modality-specific optimization objectives, aligning their optimization speeds, or decomposing multimodal learning to enhance unimodal learning. These methods fail to achieve both unimodal sufficiency and multimodal balance. In this paper, we, for the first time, address both concerns by proposing multimodal Data Remixing, including decoupling multimodal data and filtering hard samples for each modality to mitigate modality imbalance; and then batch-level reassembling to align the gradient directions and avoid cross-modal interference, thus enhancing unimodal learning sufficiency. Experimental results demonstrate that our method can be seamlessly integrated with existing approaches, improving accuracy by approximately **6.50\%$\uparrow$** on CREMAD and **3.41\%$\uparrow$** on Kinetic-Sounds, without training set expansion or additional computational overhead during inference. The source code is available at Data Remixing.

Decentralized federated learning (DFL) is a collaborative machine learning framework for training a model across participants without a central server or raw data exchange. DFL faces challenges due to statistical heterogeneity, as participants often possess data of different distributions reflecting local environments and user behaviors. Recent work has shown that the neural tangent kernel (NTK) approach, when applied to federated learning in a centralized framework, can lead to improved performance. We propose an approach leveraging the NTK to train client models in the decentralized setting, while introducing a synergy between NTK-based evolution and model averaging. This synergy exploits inter-client model deviation and improves both accuracy and convergence in heterogeneous settings. Empirical results demonstrate that our approach consistently achieves higher accuracy than baselines in highly heterogeneous settings, where other approaches often underperform. Additionally, it reaches target performance in 4.6 times fewer communication rounds. We validate our approach across multiple datasets, network topologies, and heterogeneity settings to ensure robustness and generalization. Source code for NTK-DFL is available at https://github.com/Gabe-Thomp/ntk-dfl}{https://github.com/Gabe-Thomp/ntk-dfl

This paper focuses on the fundamental challenge of partitioning input variables in attribution methods for Explainable AI, particularly in Shapley value-based approaches. Previous methods always compute attributions given a predefined partition but lack theoretical guidance on how to form meaningful variable partitions. We identify that attribution conflicts arise when the attribution of a coalition differs from the sum of its individual variables' attributions. To address this, we analyze the numerical effects of AND-OR interactions in AI models and extend the Shapley value to a new attribution metric for variable coalitions. Our theoretical findings reveal that specific interactions cause attribution conflicts, and we propose three metrics to evaluate coalition faithfulness. Experiments on synthetic data, NLP, image classification, and the game of Go validate our approach, demonstrating consistency with human intuition and practical applicability.

Recent advances in visual generation have made significant strides in producing content of exceptional quality. However, most methods suffer from a fundamental problem - a bottleneck of inference computational efficiency. Most of these algorithms involve multiple passes over a transformer model to generate tokens or denoise inputs. However, the model size is kept consistent throughout all iterations, which makes it computationally expensive. In this work, we aim to address this issue primarily through two key ideas - (a) not all parts of the generation process need equal compute, and we design a decode time model scaling schedule to utilize compute effectively, and (b) we can cache and reuse some of the intermediate computation. Combining these two ideas leads to using smaller models to process more tokens while large models process fewer tokens. These different-sized models do not increase the parameter size, as they share parameters. We rigorously experiment with ImageNet256$\times$256 , UCF101, and Kinetics600 to showcase the efficacy of the proposed method for image/video generation and frame prediction. Our experiments show that with almost $3\times$ less compute than baseline, our model obtains competitive performance.

Two-stage Learning-to-Defer (L2D) enables optimal task delegation by assigning each input to either a fixed main model or one of several offline experts, supporting reliable decision-making in complex, multi-agent environments. However, existing L2D frameworks assume clean inputs and are vulnerable to adversarial perturbations that can manipulate query allocation—causing costly misrouting or expert overload. We present the first comprehensive study of adversarial robustness in two-stage L2D systems. We introduce two novel attack strategies—untargeted and targeted—which respectively disrupt optimal allocations or force queries to specific agents. To defend against such threats, we propose SARD, a convex learning algorithm built on a family of surrogate losses that are provably Bayes-consistent and $(\mathcal{R}, \mathcal{G})$-consistent. These guarantees hold across classification, regression, and multi-task settings. Empirical results demonstrate that SARD significantly improves robustness under adversarial attacks while maintaining strong clean performance, marking a critical step toward secure and trustworthy L2D deployment.

Over 40\% of the global population lives within 100 kilometers of the coast, which contributes more than \$8 trillion annually to the global economy. Unfortunately, coastal ecosystems are increasingly vulnerable to more frequent and intense extreme weather events and rising sea levels. Coastal scientists use numerical models to simulate complex physical processes, but these models are often slow and expensive. In recent years, deep learning has become a promising alternative to reduce the cost of numerical models. However, progress has been hindered by the lack of a large-scale, high-resolution coastal simulation dataset to train and validate deep learning models. Existing studies often focus on relatively small datasets and simple processes. To fill this gap, we introduce a decade-long, high-resolution (<100m) coastal circulation modeling dataset on a real-world 3D mesh in southwest Florida with around 6 million cells. The dataset contains key oceanography variables (e.g., current velocities, free surface level, temperature, salinity) alongside external atmospheric and river forcings. We evaluated a customized Vision Transformer model that takes initial and boundary conditions and external forcings and predicts ocean variables at varying lead times. The dataset provides an opportunity to benchmark novel deep learning models for high-resolution coastal simulations (e.g., physics-informed machine learning, neural operator learning). The code and dataset can be accessed at https://github.com/spatialdatasciencegroup/CoastalBench.

Supervised fine-tuning is a standard method for adapting pre-trained large language models (LLMs) to downstream tasks. Quantization has been recently studied as a post-training technique for efficient LLM deployment. To obtain quantized fine-tuned LLMs, conventional pipelines would first fine-tune the pre-trained models, followed by post-training quantization. This often yields suboptimal performance as it fails to leverage the synergy between fine-tuning and quantization. To effectively realize low-bit quantization of weights, activations and KV caches in LLMs, we propose an algorithm named Rotated Straight-Through-Estimator (RoSTE), which combines quantization-aware supervised fine-tuning (QA-SFT) with an adaptive rotation strategy that identifies an effective rotation configuration to reduce activation outliers. We provide theoretical insights on RoSTE by analyzing its prediction error when applied to an overparameterized least square quantized training problem. Our findings reveal that the prediction error is directly proportional to the quantization error of the converged weights, which can be effectively managed through an optimized rotation configuration. Experiments on Pythia, Qwen and Llama models of different sizes demonstrate the effectiveness of RoSTE. Compared to existing post-SFT quantization baselines, our method consistently achieves superior performances across various tasks and different LLM architectures. Our code is available at https://github.com/OptimAI-Lab/RoSTE.

Generative recommendation (GR) is an emerging paradigm where user actions are tokenized into discrete token patterns and autoregressively generated as predictions. However, existing GR models tokenize each action independently, assigning the same fixed tokens to identical actions across all sequences without considering contextual relationships. This lack of context-awareness can lead to suboptimal performance, as the same action may hold different meanings depending on its surrounding context. To address this issue, we propose ActionPiece to explicitly incorporate context when tokenizing action sequences. In ActionPiece, each action is represented as a *set* of item features. Given the action sequence corpora, we construct the vocabulary by merging feature patterns as new tokens, based on their co-occurrence frequency both within individual sets and across adjacent sets. Considering the unordered nature of feature sets, we further introduce set permutation regularization, which produces multiple segmentations of action sequences with the same semantics. Our code is available at: https://github.com/google-deepmind/action_piece.

We consider the question of learnability of distribution classes in the presence of adaptive adversaries -- that is, adversaries capable of intercepting the samples requested by a learner and applying manipulations with full knowledge of the samples before passing it on to the learner. This stands in contrast to oblivious adversaries, who can only modify the underlying distribution the samples come from but not their i.i.d.\ nature. We formulate a general notion of learnability with respect to adaptive adversaries, taking into account the budget of the adversary. We show that learnability with respect to additive adaptive adversaries is a strictly stronger condition than learnability with respect to additive oblivious adversaries.

This paper studies the long-tailed semi-supervised learning (LTSSL) with distribution mismatch, where the class distribution of the labeled training data follows a long-tailed distribution and mismatches with that of the unlabeled training data. Most existing methods introduce auxiliary classifiers (experts) to model various unlabeled data distributions and produce pseudo-labels, but the expertises of various experts are not fully utilized. We observe that different experts are good at predicting different intervals of samples, e.g., long-tailed expert is skilled in samples located in the head interval and uniform expert excels in samples located in the medium interval. Therefore, we propose a dynamic expert assignment module that can estimate the class membership (i.e., head, medium, or tail class) of samples, and dynamically assigns suitable expert to each sample based on the estimated membership to produce high-quality pseudo-label in the training phase and produce prediction in the testing phase. We also theoretically reveal that integrating different experts' strengths will lead to a smaller generalization error bound. Moreover, we find that the deeper features are more biased toward the head class but with more discriminative ability, while the shallower features are less biased but also with less discriminative ability. We, therefore, propose a multi-depth feature fusion module to utilize different depth features to mitigate the model bias. Our method demonstrates its effectiveness through comprehensive experiments on the CIFAR-10-LT, STL-10-LT, and SVHN-LT datasets across various settings. The code is available at https://github.com/yaxinhou/Meta-Expert.

Fake images, created by recently advanced generative models, have become increasingly indistinguishable from real ones, making their detection crucial, urgent, and challenging. This paper introduces PiD (Pixelwise Decomposition Residuals), a novel detection method that focuses on residual signals within images. Generative models are designed to optimize high-level semantic content (principal components), often overlooking low-level signals (residual components). PiD leverages this observation by disentangling residual components from images, encouraging the model to uncover more underlying and general forgery clues independent of semantic content. Compared to prior approaches that rely on reconstruction techniques or high-frequency information, PiD is computationally efficient and does not rely on any generative models for reconstruction. Specifically, PiD operates at the pixel level, mapping the pixel vector to another color space (e.g., YUV) and then quantizing the vector. The pixel vector is mapped back to the RGB space and the quantization loss is taken as the residual for AIGC detection. Our experiment results are striking and highly surprising: PiD achieves 98% accuracy on the widely used GenImage benchmark, highlighting the effectiveness and generalization performance.

Multi-solid systems are foundational to a wide range of real-world applications, yet modeling their complex interactions remains challenging. Existing deep learning methods predominantly rely on implicit modeling, where the factors influencing solid deformation are not explicitly represented but are instead indirectly learned. However, as the number of solids increases, these methods struggle to accurately capture intricate physical interactions. In this paper, we introduce a novel explicit modeling paradigm that incorporates factors influencing solid deformation through structured modules. Specifically, we present Unisoma, a unified and flexible Transformer-based model capable of handling variable numbers of solids. Unisoma directly captures physical interactions using contact modules and adaptive interaction allocation mechanism, and learns the deformation through a triplet relationship. Compared to implicit modeling techniques, explicit modeling is more well-suited for multi-solid systems with diverse coupling patterns, as it enables detailed treatment of each solid while preventing information blending and confusion. Experimentally, Unisoma achieves consistent state-of-the-art performance across seven well-established datasets and two complex multi-solid tasks. Code is avaiable at [https://github.com/therontau0054/Unisoma](https://github.com/therontau0054/Unisoma).

Temporal sequences, even after stationarization, often exhibit leptokurtic distributions with fat tails and persistent distribution shifts. These properties destabilize feature dynamics, amplify model variance, and hinder model convergence in time series forecasting. To address this, we propose Morphing-Flow (MoF), a framework that combines a spline-based transform layer (Flow) and a test-time-trained method (Morph), which adaptively normalizes non-stationary, fat-tailed distributions while preserving critical extreme features. MoF ensures that inputs remain within a network’s effective activation space—a structured, normal-like distribution—even under distributional drift. Experiments across eight datasets show that MoF achieves state-of-the-art performance: With a simple linear backbone architecture, it matches the performance of state-of-the-art models on datasets such as Electricity and ETTh2. When paired with a patch-based Mamba architecture, MoF outperforms its closest competitor by 6.3% on average and reduces forecasting errors in fat-tailed datasets such as Exchange by 21.7%. Moreover, MoF acts as a plug-and-play module, boosting performance in existing models without architectural changes.

**In this position paper, we argue that human baselines in foundation model evaluations must be more rigorous and more transparent to enable meaningful comparisons of human vs. AI performance, and we provide recommendations and a reporting checklist towards this end.** Human performance baselines are vital for the machine learning community, downstream users, and policymakers to interpret AI evaluations. Models are often claimed to achieve "super-human" performance, but existing baselining methods are neither sufficiently rigorous nor sufficiently well-documented to robustly measure and assess performance differences. Based on a meta-review of the measurement theory and AI evaluation literatures, we derive a framework with recommendations for designing, executing, and reporting human baselines. We synthesize our recommendations into a checklist that we use to systematically review 115 human baselines (studies) in foundation model evaluations and thus identify shortcomings in existing baselining methods; our checklist can also assist researchers in conducting human baselines and reporting results. We hope our work can advance more rigorous AI evaluation practices that can better serve both the research community and policymakers. Data is available at: [https://github.com/kevinlwei/human-baselines](https://github.com/kevinlwei/human-baselines).

Federated learning harnesses the power of distributed optimization to train a unified machine learning model across separate clients. However, heterogeneous data distributions and computational workloads can lead to inconsistent updates and limit model performance. This work tackles these challenges by proposing FedECADO, a new algorithm inspired by a dynamical system representation of the federated learning process. FedECADO addresses non-IID data distribution through an aggregate sensitivity model that reflects the amount of data processed by each client. To tackle heterogeneous computing, we design a multi-rate integration method with adaptive step-size selections that synchronizes active client updates in continuous time. Compared to prominent techniques, including FedProx, FedExp, and FedNova, FedECADO achieves higher classification accuracies in numerous heterogeneous scenarios.

Deep clustering, an unsupervised technique independent of labels, necessitates tailored supervision for model training. Prior methods explore supervision like similarity and pseudo labels, yet overlook individual sample training analysis. Our study correlates sample stability during unsupervised training with clustering accuracy and network memorization on a per-sample basis. Unstable representations across epochs often lead to mispredictions, indicating difficulty in memorization and atypicality. Leveraging these findings, we introduce supervision signals for the first time based on sample stability at the representation level. Our proposed strategy serves as a versatile tool to enhance various deep clustering techniques. Experiments across benchmark datasets showcase that incorporating sample stability into training can improve the performance of deep clustering. The code is available at https://github.com/LZX-001/LFSS.