ICML 2025 论文列表

Existing methods for inferring latent relational structures struggle to integrate partial prior knowledge, such as known edges, node-degree constraints, and global sparsity, without destabilizing training or conflicting with probabilistic assumptions. We propose Soft-Gated Structural Inference (SGSI), a VAE framework that seamlessly incorporates domain constraints via (1) soft gating with learnable edge masks to preserve gradients, (2) cloning-clamping of deterministic edges to avoid distributional conflicts, and (3) adaptive regularization to balance data-driven learning with domain constraints. By excluding known edges from stochastic inference, SGSI reallocates capacity to uncertain interactions, optimizing the information bottleneck trade-off. Experiments on 16 datasets show SGSI improves edge recovery by up to $9$\% AUROC over baselines, scales to larger graphs ($94.2$\% AUROC), and maintains stable training. SGSI bridges domain expertise with data-driven learning, enabling interpretable and robust structural discovery in dynamical systems.

In recent years there have been remarkable breakthroughs in image-to-video generation. However, the 3D consistency and camera controllability of generated frames have remained unsolved. Recent studies have attempted to incorporate camera control into the generation process, but their results are often limited to simple trajectories or lack the ability to generate consistent videos from multiple distinct camera paths for the same scene. To address these limitations, we introduce Cavia, a novel framework for camera-controllable, multi-view video generation, capable of converting an input image into multiple spatiotemporally consistent videos. Our framework extends the spatial and temporal attention modules into view-integrated attention modules, improving both viewpoint and temporal consistency. This flexible design allows for joint training with diverse curated data sources, including scene-level static videos, object-level synthetic multi-view dynamic videos, and real-world monocular dynamic videos. To the best of our knowledge, Cavia is the first framework that enables users to generate multiple videos of the same scene with precise control over camera motion, while simultaneously preserving object motion. Extensive experiments demonstrate that Cavia surpasses state-of-the-art methods in terms of geometric consistency and perceptual quality.

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.

The deployment of mixture-of-experts (MoE) large language models (LLMs) presents significant challenges due to their high memory demands. These challenges become even more pronounced in multi-tenant environments, where shared resources must accommodate multiple models, limiting the effectiveness of conventional virtualization techniques. This paper addresses the problem of efficiently serving multiple fine-tuned MoE-LLMs on a single GPU. We propose a serving system that employs \textit{similarity-based expert consolidation} to reduce the overall memory footprint by sharing similar experts across models. To ensure output quality, we introduce \textit{runtime partial reconfiguration}, dynamically replacing non-expert layers when processing requests from different models. As a result, our approach achieves competitive output quality while maintaining throughput comparable to serving a single model, and incurs only a negligible increase in time-to-first-token (TTFT). Experiments on a server with a single NVIDIA A100 GPU (80GB) using Mixtral-8x7B models demonstrate an 85\% average reduction in turnaround time compared to NVIDIA's multi-instance GPU (MIG). Furthermore, experiments on Google's Switch Transformer Base-8 model with up to four variants demonstrate the scalability and resilience of our approach in maintaining output quality compared to other model merging baselines, highlighting its effectiveness.

Selecting high-quality candidates from large datasets is critical in applications such as drug discovery, precision medicine, and alignment of large language models (LLMs). While Conformal Selection (CS) provides rigorous uncertainty quantification, it is limited to univariate responses and scalar criteria. To address this, we propose Multivariate Conformal Selection (mCS), a generalization of CS designed for multivariate response settings. Our method introduces regional monotonicity and employs multivariate nonconformity scores to construct conformal $p$-values, enabling finite-sample False Discovery Rate (FDR) control. We present two variants: $\texttt{mCS-dist}$, using distance-based scores, and $\texttt{mCS-learn}$, which learns optimal scores via differentiable optimization. Experiments on simulated and real-world datasets demonstrate that mCS significantly improves selection power while maintaining FDR control, establishing it as a robust framework for multivariate selection tasks.

Graph Contrastive Learning (GCL) improves Graph Neural Network (GNN)-based protein representation learning by enhancing its generalization and robustness. Existing GCL approaches for protein representation learning rely on 2D topology, where graph augmentation is solely based on topological features, ignoring the intrinsic biological properties of proteins. Besides, 3D structure-based protein graph augmentation remains unexplored, despite proteins inherently exhibiting 3D structures. To bridge this gap, we propose novel biology-aware graph augmentation strategies from the perspective of invariance and integrate them into the protein GCL framework. Specifically, we introduce Functional Community Invariance (FCI)-based graph augmentation, which employs spectral constraints to preserve topology-driven community structures while incorporating residue-level chemical similarity as edge weights to guide edge sampling and maintain functional communities. Furthermore, we propose 3D Protein Structure Invariance (3-PSI)-based graph augmentation, leveraging dihedral angle perturbations and secondary structure rotations to retain critical 3D structural information of proteins while diversifying graph views. Extensive experiments on four different protein-related tasks demonstrate the superiority of our proposed GCL protein representation learning framework.

Language model (LLM) post-training can refine behaviors and unlock new skills, but the open science supporting these post-training techniques is still in its infancy. One limiting factor has been the difficulty of conducting large-scale comparative analyses of synthetic data generating models and LLM judges. To close this gap, we introduce WildChat-50M, the largest public chat dataset to date. We extend the existing WildChat dataset to include responses not only from GPT, but from over 50 different open-weight models, ranging in size from 0.5B to 104B parameters. We conduct an extensive comparative analysis and demonstrate the potential of this dataset by creating Re-Wild, our own public SFT mix, which outperforms the recent Tulu-3 SFT mixture from Allen AI with only 40% as many samples.

Future advanced AI systems may learn sophisticated strategies through reinforcement learning (RL) that humans cannot understand well enough to safely evaluate. We propose a training method which avoids agents learning undesired multi-step plans that receive high reward (multi-step "reward hacks") even if humans are not able to detect that the behavior is undesired. The method, Myopic Optimization with Non-myopic Approval (MONA), works by combining short-sighted optimization with far-sighted reward. We demonstrate that MONA can prevent multi-step reward hacking that ordinary RL causes, even without being able to detect the reward hacking and without any extra information that ordinary RL does not get access to. We study MONA empirically in three settings which model different misalignment failure modes including 2-step environments with LLMs representing delegated oversight and encoded reasoning and longer-horizon gridworld environments representing sensor tampering.

Differentiable structure learning of causal directed acyclic graphs (DAGs) is an emerging field in causal discovery, leveraging powerful neural learners. However, the incorporation of ancestral constraints, essential for representing abstract prior causal knowledge, remains an open research challenge. This paper addresses this gap by introducing a generalized framework for integrating ancestral constraints. Specifically, we identify two key issues: the non-equivalence of relaxed characterizations for representing path existence and order violations among paths during optimization. In response, we propose a binary-masked characterization method and an order-guided optimization strategy, tailored to address these challenges. We provide theoretical justification for the correctness of our approach, complemented by experimental evaluations on both synthetic and real-world datasets.

We study the problem of learning the topology of a directed Gaussian Graphical Model under the equal-variance assumption, where the graph has $n$ nodes and maximum in-degree $d$. Prior work has established that $O(d \log n)$ samples are sufficient for this task. However, an important factor that is often overlooked in these analyses is the dependence on the condition number of the covariance matrix of the model. Indeed, all algorithms from prior work require a number of samples that grows polynomially with this condition number. In many cases this is unsatisfactory, since the condition number could grow polynomially with $n$, rendering these prior approaches impractical in high-dimensional settings. In this work, we provide an algorithm that recovers the underlying graph and prove that the number of samples required is independent of the condition number. Furthermore, we establish lower bounds that nearly match the upper bound up to a $d$-factor, thus providing an almost tight characterization of the true sample complexity of the problem. Moreover, under a further assumption that all the variances of the variables are bounded, we design a polynomial-time algorithm that recovers the underlying graph, at the cost of an additional polynomial dependence of the sample complexity on $d$. We complement our theoretical findings with simulations on synthetic datasets that confirm our predictions.

Dense depth prediction deep neural networks (DNN) have achieved impressive results for both monocular and binocular data but they are limited by high computational complexity, restricting their use on low-end devices. For better on-device efficiency and hardware utilization, weights and activations of the DNN should be converted to low-bit precision. However, this precision is not sufficient for representing high dynamic range depth. In this paper, we aim to overcome this limitation and restore high-precision depth from low-bit precision predictions. To achieve this, we propose to represent high dynamic range depth as two low dynamic range components of a Hilbert curve, and to train the full precision DNN to directly predict the latter. For on-device deployment, we use standard quantization methods and add a post-processing step that reconstructs depth from the Hilbert curve components predicted in low-bit precision. Extensive experiments demonstrate that our method increases bit precision of predicted depth by up to three bits with little computational overhead. We also observe a positive side effect of quantization error reduction by up to five times. Our method enables effective and accurate depth prediction with DNN weights and activations quantized to eight bit precision.

The progress in materials science and drug discovery is impeded by the availability of labeled data and the high costs of manual annotation, driving the need for efficient strategies to capture molecular representations and enable accurate predictions. Pretrained Graph Neural Networks have shown promise in capturing universal molecular representations, but adapting them to task-specific applications remains challenging. In this paper, we propose Multilevel Informed Prompt-Tuning (MIPT), a novel framework for effectively tailoring pretrained models to molecule-related tasks. MIPT utilizes a lightweight, multi-level prompt learning module to capture node-level and graph-level task-specific knowledge, ensuring adaptable and efficient tuning. Additionally, a noise penalty mechanism is introduced to address mismatches between pretrained representations and downstream tasks, reducing irrelevant or noisy information. Experimental results show that MIPT surpasses all baselines, aligning graph space and task space while achieving significant improvements in molecule-related tasks, demonstrating its scalability and versatility for molecular tasks.

Cascade Ranking is a prevalent architecture in large-scale top-k selection systems like recommendation and advertising platforms. Traditional training methods focus on single-stage optimization, neglecting interactions between stages. Recent advances have introduced interaction-aware training paradigms, but still struggle to 1) align training objectives with the goal of the entire cascade ranking (i.e., end-to-end recall of ground-truth items) and 2) learn effective collaboration patterns for different stages. To address these challenges, we propose LCRON, which introduces a novel surrogate loss function derived from the lower bound probability that ground truth items are selected by cascade ranking, ensuring alignment with the overall objective of the system. According to the properties of the derived bound, we further design an auxiliary loss for each stage to drive the reduction of this bound, leading to a more robust and effective top-k selection. LCRON enables end-to-end training of the entire cascade ranking system as a unified network. Experimental results demonstrate that LCRON achieves significant improvement over existing methods on public benchmarks and industrial applications, addressing key limitations in cascade ranking training and significantly enhancing system performance.

Neural operators have emerged as a promising approach for solving high-dimensional partial differential equations (PDEs). However, existing neural operators often have difficulty in dealing with constrained PDEs, where the solution must satisfy additional equality or inequality constraints beyond the governing equations. To close this gap, we propose a novel neural operator, Hyper Extended Adaptive PDHG (HEAP) for constrained high-dim PDEs, where the learned operator evolves in the parameter space of PDEs. We first show that the evolution operator learning can be formulated as a quadratic programming (QP) problem, then unroll the adaptive primal-dual hybrid gradient (APDHG) algorithm as the QP-solver into the neural operator architecture. This allows us to improve efficiency while retaining theoretical guarantees of the constrained optimization. Empirical results on a variety of high-dim PDEs show that HEAP outperforms the state-of-the-art neural operator model.

Despite the tremendous success of deep learning in computer vision, models still fall behind humans in generalizing to new input distributions. Existing benchmarks do not investigate the specific failure points of models by analyzing performance under many controlled conditions. Our study systematically dissects where and why models struggle with contour integration - a hallmark of human vision -- by designing an experiment that tests object recognition under various levels of object fragmentation. Humans (n=50) perform at high accuracy, even with few object contours present. This is in contrast to models which exhibit substantially lower sensitivity to increasing object contours, with most of the over 1,000 models we tested barely performing above chance. Only at very large scales ($\sim5B$ training dataset size) do models begin to approach human performance. Importantly, humans exhibit an integration bias - a preference towards recognizing objects made up of directional fragments over directionless fragments. We find that not only do models that share this property perform better at our task, but that this bias also increases with model training dataset size, and training models to exhibit contour integration leads to high shape bias. Taken together, our results suggest that contour integration is a hallmark of object vision that underlies object recognition performance, and may be a mechanism learned from data at scale.

Large Language Models (LLMs) matrices can often be expressed in low-rank format with potential to relax memory and compute resource requirements. Unlike previous works which pivot around developing novel matrix decomposition algorithms, in this work we focus to study the emerging non-uniform low-rank properties across weight matrices in LLMs through the lens of stabilizing gradient subspace. \textit{Firstly,} we provide a theoretical framework to understand the stabilization of gradient subspaces through Hessian analysis. \textit{Secondly,} we empirically establish a consequential relationship between the gradient dynamics and low-rank expressiveness of weight matrices. Our findings reveal that different LLM components exhibit varying levels of converged low-rank structure, necessitating a non-uniform rank reduction across them to minimize performance drop due to compression. In view of that, we present \textit{Weight Low-Rank Projection} \textbf{(WeLore)} that unifies weight compression and memory-efficient fine-tuning as ONE, in a data-agnostic and one-shot way. Going beyond only as a compression technique, WeLore categorizes weight matrices into Low-rank Components (LRCs) and Non-Low-rank Components (N-LRCs) based on their ability to express themselves as low-rank. Our gradient dynamics perspective illustrate that \textit{LRCs tend to have better finetuning capabilities} and their standalone finetuning can closely mimic (sometimes outperform) the training loss trajectory and performance of full-finetuning with notable memory and compute footprint reduction. All codes and checkpoints will be released.

While Large Language Models (LLMs) demonstrate impressive performance in mathematics, existing math benchmarks come with significant limitations. Many focus on problems with fixed ground-truth answers, and are often saturated due to problem simplicity or the viability of guessing or memorization. Crucially, they capture only a narrow subset of relevant math problems. To address this research gap, we introduce MathConstruct, a new benchmark of 127 challenging problems sourced from various math competitions, which targets *constructive proofs*, a widely encountered problem type requiring the construction of mathematical objects with specific properties. These proofs are particularly suitable for LLM evaluation, as solution correctness can be easily verified. Our automated verifiers also enable MathConstruct to generate problem variations, used to evaluate robustness. State-of-the-art LLMs solve only 41\% of MathConstruct problems, highlighting its complexity and importance for LLM evaluation.

Mathematical reasoning tasks pose significant challenges for large language models (LLMs) because they require precise logical deduction and sequence analysis. In this work, we introduce the concept of critical tokens -- elements within reasoning trajectories that significantly influence incorrect outcomes. We present a novel framework for identifying these tokens through rollout sampling and demonstrate their substantial divergence from traditional error tokens. Through extensive experiments on datasets such as GSM8K and MATH500, we show that identifying and replacing critical tokens significantly improves model accuracy. We propose an efficient methodology for pinpointing these tokens in large-scale datasets using contrastive estimation and extend this framework to enhance model training processes with direct preference optimization (DPO). Experimental results on GSM8K and MATH500 benchmarks with the widely used models Llama-3 (8B and 70B) and Deepseek-math (7B) demonstrate the effectiveness of the proposed approach, cDPO. Our results underscore the potential of leveraging critical tokens to reduce errors in reasoning tasks, advancing the development of AI systems capable of robust logical deduction.

Semi-supervised Federated Learning (SSFL) is a promising approach that allows clients to collaboratively train a global model in the absence of their local data labels. The key step of SSFL is the re-labeling where each client adopts two types of available models, namely global and local models, to re-label the local data. While various technologies such as using the global model or the average of two models have been proposed to conduct the re-labeling step, little literature delves deeply into the performance dominance and limitations of the two models. In this paper, we first theoretically and empirically demonstrate that the local model achieves higher re-labeling accuracy over local data while the global model can progressively improve the re-labeling performance by introducing the extra data knowledge of other clients. Based on these findings, we propose BSemiFL which re-labels the local data through the collaboration between the local and global model in a Bayesian approach. Specifically, to re-label any given local sample, BSemiFL first uses Bayesian inference to assess the closeness of the local/global model to the sample. Then, it applies a weighted combination of their pseudo labels, using the closeness as the weights. Theoretical analysis shows that the labeling error of our method is smaller than that of simply using the global model, the local model, or their simple average. Experimental results show that BSemiFL improves the performance by up to $9.8\%$ as compared to state-of-the-art methods.