NeurIPS 2025 论文列表

In both neuroscience and artificial intelligence (AI), it is well-established that neural “coupling” gives rise to dynamically distributed systems. These systems exhibit self-organized spatiotemporal patterns of synchronized neural oscillations, enabling the representation of abstract concepts. By capitalizing on the unprecedented amount of human neuroimaging data, we propose that advancing the theoretical understanding of rhythmic coordination in neural circuits can offer powerful design principles for the next generation of machine learning models with improved efficiency and robustness. To this end, we introduce a physics-informed deep learning framework for \underline{B}rain \underline{R}hythm \underline{I}dentification by \underline{K}uramoto and \underline{C}ontrol (coined \modelname{}) to characterize the synchronization of neural oscillations that shapes the dynamics of evolving cognitive states. Recognizing that brain networks are structurally connected yet behaviorally dynamic, we further conceptualize rhythmic neural activity as an artificial dynamical system of coupled oscillators, offering a shared mechanistic bridge to brain-inspired machine intelligence. By treating each node as an oscillator interacting with its neighbors, this approach moves beyond the conventional paradigm of graph heat diffusion and establishes a new regime of representation compression through oscillatory synchronization. Empirical evaluations demonstrate that this synchronization-driven mechanism not only mitigates over-smoothing in deep GNNs but also enhances the model’s capacity for reasoning and solving complex graph-based problems.

This paper initiates the study of data-dependent regret bounds in constrained MAB settings. These are bounds that depend on the sequence of losses that characterize the problem instance. Thus, in principle they can be much smaller than classical $\widetilde{\mathcal{O}}(\sqrt{T})$ regret bounds, while being equivalent to them in the worst case. Despite this, data-dependent regret bounds have been completely overlooked in constrained MABs. The goal of this paper is to answer the question: Can data-dependent regret bounds be derived in the presence of constraints? We provide an affirmative answer in constrained MABs with adversarial losses and stochastic constraints. Specifically, our main focus is on the most challenging and natural settings with hard constraints, where the learner must ensure that the constraints are always satisfied with high probability. We design an algorithm with a regret bound consisting of two data-dependent terms. The first one captures the difficulty of satisfying the constraints, while the second one encodes the complexity of learning independently of their presence. We also prove a lower bound showing that these two terms are not artifacts of our specific approach and analysis, but rather the fundamental components that inherently characterize the problem complexity. Finally, in designing our algorithm, we also derive some novel results in the related (and easier) soft constraints settings, which may be of independent interest.

Large foundation models are typically trained on data from multiple domains, with the data mixture--the proportion of each domain used--playing a critical role in model performance. The standard approach to selecting this mixture relies on trial and error, which becomes impractical for large-scale pretraining. We propose a systematic method to determine the optimal data mixture for any target domain using scaling laws. Our approach accurately predicts the loss of a model of size $N$ trained with $D$ tokens and a specific domain weight vector $h$. We validate the universality of these scaling laws by demonstrating their predictive power in three distinct and large-scale settings: large language model (LLM), native multimodal model (NMM), and large vision models (LVM) pretraining. We further show that these scaling laws can extrapolate to new data mixtures and across scales: their parameters can be accurately estimated using a few small-scale training runs, and used to estimate the performance at larger scales and unseen domain weights. The scaling laws allow to derive the optimal domain weights for any target domain under a given training budget ($N$,$D$), providing a principled alternative to costly trial-and-error methods.

Bilevel optimization has garnered significant attention in the machine learning community recently, particularly regarding the development of efficient numerical methods. While substantial progress has been made in developing efficient algorithms for optimistic bilevel optimization, the study of methods for solving Pessimistic Bilevel Optimization (PBO) remains relatively less explored, especially the design of fully first-order, single-loop gradient-based algorithms. This paper aims to bridge this research gap. We first propose a novel smooth approximation to the PBO problem, using penalization and regularization techniques. Building upon this approximation, we then propose SiPBA (Single-loop Pessimistic Bilevel Algorithm), a new gradient-based method specifically designed for PBO which avoids second-order derivative information or inner-loop iterations for subproblem solving. We provide theoretical validation for the proposed smooth approximation scheme and establish theoretical convergence for the algorithm SiPBA. Numerical experiments on synthetic examples and practical applications demonstrate the effectiveness and efficiency of SiPBA.

Offline safe reinforcement learning aims to learn policies that satisfy predefined safety constraints from static datasets. Existing sequence-model-based methods condition action generation on symmetric input tokens for return-to-go and cost-to-go, neglecting their intrinsic asymmetry: RTG serves as a flexible performance target, while CTG should represent a rigid safety boundary. This symmetric conditioning leads to unreliable constraint satisfaction, especially when encountering out-of-distribution cost trajectories. To address this, we propose Boundary-to-Region (B2R), a framework that enables asymmetric conditioning through cost signal realignment . B2R redefines CTG as a boundary constraint under a fixed safety budget, unifying the cost distribution of all feasible trajectories while preserving reward structures. Combined with rotary positional embeddings , it enhances exploration within the safe region. Experimental results show that B2R satisfies safety constraints in 35 out of 38 safety-critical tasks while achieving superior reward performance over baseline methods. This work highlights the limitations of symmetric token conditioning and establishes a new theoretical and practical approach for applying sequence models to safe RL.

Gradient clipping is increasingly important in centralized learning (CL) and federated learning (FL). Many works focus on its optimization properties under strong assumptions involving Gaussian noise and standard smoothness. However, practical machine learning tasks often only satisfy weaker conditions, such as heavy-tailed noise and $(L_0, L_1)$-smoothness. To bridge this gap, we propose a high-probability analysis for clipped Stochastic Gradient Descent (SGD) under these weaker assumptions. Our findings show a better convergence rate than existing ones can be achieved, and our high-probability analysis does not rely on the bounded gradient assumption. Moreover, we extend our analysis to FL, where a gap remains between expected and high-probability convergence, which the naive clipped SGD cannot bridge. Thus, we design a new \underline{Fed}erated \underline{C}lipped \underline{B}atched \underline{G}radient (FedCBG) algorithm, and prove the convergence and generalization bounds with high probability for the first time. Our analysis reveals the trade-offs between the optimization and generalization performance. Extensive experiments demonstrate that \methodname{} can generalize better to unseen client distributions than state-of-the-art baselines.

The challenge of learning with noisy labels is significant in machine learning, as it can severely degrade the performance of prediction models if not addressed properly. This paper introduces a novel framework that conceptualizes noisy label correction as a reinforcement learning (RL) problem. The proposed approach, Reinforcement Learning for Noisy Label Correction (RLNLC), defines a comprehensive state space representing data and their associated labels, an action space that indicates possible label corrections, and a reward mechanism that evaluates the efficacy of label corrections. RLNLC learns a deep feature representation based policy network to perform label correction through reinforcement learning, utilizing an actor-critic method. The learned policy is subsequently deployed to iteratively correct noisy training labels and facilitate the training of the prediction model. The effectiveness of RLNLC is demonstrated through extensive experiments on multiple benchmark datasets, where it consistently outperforms existing state-of-the-art techniques for learning with noisy labels.

We study stochastic linear bandits with heavy-tailed rewards, where the rewards have a finite $(1+\epsilon)$-absolute central moment bounded by $\upsilon$ for some $\epsilon \in (0,1]$. We improve both upper and lower bounds on the minimax regret compared to prior work. When $\upsilon = \mathcal{O}(1)$, the best prior known regret upper bound is $\tilde{O}(d T^{\frac{1}{1+\epsilon}})$. While a lower with the same scaling has been given, it relies on a construction using $\upsilon = d$, and adapting the construction to the bounded-moment regime with $\upsilon = \mathcal{O}(1)$ yields only a $\Omega(d^{\frac{\epsilon}{1+\epsilon}} T^{\frac{1}{1+\epsilon}})$ lower bound. This matches the known rate for multi-armed bandits and is generally loose for linear bandits, in particular being $\sqrt{d}$ below the optimal rate in the finite-variance case ($\epsilon = 1$). We propose a new elimination-based algorithm guided by experimental design, which achieves regret $\tilde{\mathcal{O}}(d^{\frac{1+3\epsilon}{2(1+\epsilon)}} T^{\frac{1}{1+\epsilon}})$, thus improving the dependence on $d$ for all $\epsilon \in (0,1)$ and recovering a known optimal result for $\epsilon = 1$. We also establish a lower bound of $\Omega(d^{\frac{2\epsilon}{1+\epsilon}} T^{\frac{1}{1+\epsilon}})$, which strictly improves upon the multi-armed bandit rate and highlights the hardness of heavy-tailed linear bandit problems. For finite action sets of size $n$, we derive upper and lower bounds of $\tilde{\mathcal{O}}(\sqrt d (\log n)^{\frac{\epsilon}{1+\epsilon}}T^{\frac{1}{1+\epsilon}})$ and $\tilde\Omega(d^{\frac{\epsilon}{1+\epsilon}}(\log n)^{\frac{\epsilon}{1+\epsilon}} T^{\frac{1}{1+\epsilon}})$, respectively. Finally, we provide action-set-dependent regret upper bounds and show that for some geometries, such as $l_p$-norm balls for $p \le 1 + \epsilon$, we can further reduce the dependence on $d$.

The efficiency and quality of language model pretraining are largely determined by the way pretraining data are selected. In this paper, we introduce *Group-MATES*, an efficient group-level data selection approach to optimize the speed-quality frontier of language model pretraining. Specifically, Group-MATES parameterizes costly group-level selection with a relational data influence model. To train this model, we sample training trajectories of the language model and collect oracle data influences alongside. The relational data influence model approximates the oracle data influence by weighting individual influence with relationships among training data. To enable efficient selection with our relational data influence model, we partition the dataset into small clusters using relationship weights and select data within each cluster independently. Experiments on DCLM 400M-4x, 1B-1x, and 3B-1x show that Group-MATES achieves 3.5\%-9.4\% relative performance gains over random selection across 22 downstream tasks, nearly doubling the improvements achieved by state-of-the-art individual data selection baselines. Furthermore, Group-MATES reduces the number of tokens required to reach a certain downstream performance by up to 1.75x, substantially elevating the speed-quality frontier. Further analyses highlight the critical role of relationship weights in the relational data influence model and the effectiveness of our cluster-based inference. Our code is open-sourced at https://github.com/facebookresearch/Group-MATES.

Understanding the evolving dependence between two sets of multivariate signals is fundamental in neuroscience and other domains where sub-networks in a system interact dynamically over time. Despite the growing interest in multivariate time series analysis, existing methods for between-clusters dependence typically rely on the assumption of stationarity and lack the temporal resolution to capture transient, frequency-specific interactions. To overcome this limitation, we propose scale-specific wavelet canonical coherence (WaveCanCoh), a novel framework that extends canonical coherence analysis to the nonstationary setting by leveraging the multivariate locally stationary wavelet model. The proposed WaveCanCoh enables the estimation of time-varying canonical coherence between clusters, providing interpretable insight into scale-specific time-varying interactions between clusters. Through extensive simulation studies, we demonstrate that WaveCanCoh accurately recovers true coherence structures under both locally stationary and general nonstationary conditions. Application to local field potential (LFP) activity data recorded from the hippocampus reveals distinct dynamic coherence patterns between correct and incorrect memory-guided decisions, illustrating capacity of the method to detect behaviorally relevant neural coordination. These results highlight WaveCanCoh as a flexible and principled tool for modeling complex cross-group dependencies in nonstationary multivariate systems. Code for implementing WaveCanCoh is available at https://github.com/mhaibo/WaveCanCoh.git.

Large Language Models (LLMs) enable a new form of digital experimentation where treatments combine human and model-generated content in increasingly sophisticated ways. The main methodological challenge in this setting is representing these high-dimensional treatments without losing their semantic meaning or rendering analysis intractable. Here we address this problem by focusing on learning low-dimensional representations that capture the underlying structure of such treatments. These representations enable downstream applications such as guiding generative models to produce meaningful treatment variants and facilitating adaptive assignment in online experiments. We propose double kernel representation learning, which models the causal effect through the inner product of kernel-based representations of treatments and user covariates. We develop an alternating-minimization algorithm that learns these representations efficiently from data and provide convergence guarantees under a low-rank factor model. As an application of this framework, we introduce an adaptive design strategy for online experimentation and demonstrate the method's effectiveness through numerical experiments.

Polymers are large macromolecules composed of repeating structural units known as monomers and are widely applied in fields such as energy storage, construction, medicine, and aerospace. However, existing graph neural network methods, though effective for small molecules, only model the single unit of polymers and fail to produce consistent vector representations for the true polymer structure with varying numbers of units. To address this challenge, we introduce Graph Repetition Invariance (GRIN), a novel method to learn polymer representations that are invariant to the number of repeating units in their graph representations. GRIN integrates a graph-based maximum spanning tree alignment with repeat-unit augmentation to ensure structural consistency. We provide theoretical guarantees for repetition‐invariance from both model and data perspectives, demonstrating that three repeating units are the minimal augmentation required for optimal invariant representation learning. GRIN outperforms state-of-the-art baselines on both homopolymer and copolymer benchmarks, learning stable, repetition-invariant representations that generalize effectively to polymer chains of unseen sizes.

We demonstrate that geometrically distinctive keys during LLM inference tend to have high attention scores. Based on the phenomenon we propose KeyDiff, a training-free KV cache eviction method based solely on key similarity. Unlike other KV cache eviction methods, KeyDiff can process arbitrarily long prompts within strict resource constraints and efficiently generate responses. We provide a theoretical basis for KeyDiff by relating key diversity with attention scores. These results imply KeyDiff can efficiently identify the most important tokens to retain. Notably KeyDiff does not rely on attention scores, allowing the use of optimized attention mechanisms like FlashAttention. Under a strict memory allowance, we demonstrate the effectiveness of KeyDiff for the Llama and Qwen model families by observing a performance gap of less than 0.04\% with 8K cache budget (~23\% KV cache reduction) from the non-evicting baseline on LongBench for Llama 3.1-8B and Llama 3.2-3B. We also observe near baseline performance for Deepseek-R1-Distill-Llama-8B on the Math500 reasoning benchmark and decrease end-to-end inference latency by up to 30\% compared to the other token-eviction methods.

A major problem of kernel-based methods (e.g., least squares support vector machines, LS-SVMs) for solving linear/nonlinear ordinary differential equations (ODEs) is the prohibitive $O(an^3)$ ($a=1$ for linear ODEs and 27 for nonlinear ODEs) part of their computational complexity with increasing temporal discretization points $n$. We propose a novel Nyström-accelerated LS-SVMs framework that breaks this bottleneck by reformulating ODEs as primal-space constraints. Specifically, we derive for the first time an explicit Nyström-based mapping and its derivatives from one-dimensional temporal discretization points to a higher $m$-dimensional feature space ($1< m\le n$), enabling the learning process to solve linear/nonlinear equation systems with $m$-dependent complexity. Numerical experiments on sixteen benchmark ODEs demonstrate: 1) $10-6000$ times faster computation than classical LS-SVMs and physics-informed neural networks (PINNs), 2) comparable accuracy to LS-SVMs ($<0.13\%$ relative MAE, RMSE, and $\left \| y-\hat{y} \right \| _{\infty } $difference) while maximum surpassing PINNs by 72\% in RMSE, and 3) scalability to $n=10^4$ time steps with $m=50$ features. This work establishes a new paradigm for efficient kernel-based ODEs learning without significantly sacrificing the accuracy of the solution.

Large Language Models have excelled in various domains but face efficiency challenges due to the growing Key-Value (KV) cache required for long-sequence inference. Recent efforts aim to reduce KV cache size by evicting vast non-critical cache elements during runtime while preserving generation quality. However, these methods typically allocate compression budgets uniformly across all attention heads, ignoring the unique attention patterns of each head. In this paper, we establish a theoretical loss upper bound between pre- and post-eviction attention output, explaining the optimization target of prior cache eviction methods, while guiding the optimization of adaptive budget allocation. Base on this, we propose {\it Ada-KV}, the first head-wise adaptive budget allocation strategy. It offers plug-and-play benefits, enabling seamless integration with prior cache eviction methods. Extensive evaluations on 13 datasets from Ruler and 16 datasets from LongBench, all conducted under both question-aware and question-agnostic scenarios, demonstrate substantial quality improvements over existing methods. Our code is available at https://github.com/FFY0/AdaKV.

Zeroth-order Optimization (ZO) has received wide attention in machine learning, especially when computing full gradient is expensive or even impossible. Recently, ZO has emerged as an important paradigm for memory-efficient fine-tuning of large language models (LLMs), circumventing the memory overhead of backpropagation. However, existing ZO gradient estimators exhibit dimension-dependent variance scaling as $\Theta(d)$, leading to dimension-dependent convergence rates without further assumptions on the objective function, which is prohibitive for large-scale LLM parameters. To address this problem, we present a Pseudo-Zeroth-Order (PseuZO) framework for optimizing composite objective functions, especially large-scale models: $ \min_{\mathbf{x} \in \mathcal{X}} \mathcal{F}(\mathbf{x})= \bbE_{\mathbf{z}} g\circ h(\mathbf{x};\mathbf{z}) $, where $h$ represents complex, high-dimensional representations and $g$ is a task-specific loss. While existing zeroth-order methods estimate gradients with final loss functions, our PseuZO algorithm estimate the Jacobian matrix of $h(\mathbf{x})$ with the model output $\mathbf{o}= h(\mathbf{x})$, and the gradient of the loss function on model output $\mathbf{e} = \nabla_{\mathbf{o}} g(\mathbf{o})$, and apply exponential moving average on Jacobian estimators to reduce the variance. Moreover, we use the sliding window technique to reduce memory costs. Our algorithm achieves an $O( \max \lbrace \alpha_1 L\epsilon^{-2}, \alpha_1 L \sigma_2^2\epsilon^{-4} \rbrace )$ convergence rate, where $\alpha_1$ is the effective dimension of $\mathcal{F}$. Experimental results demonstrate that PseuZO outperforms MeZO and MeZO-SVRG in classification, multiple choice and generation tasks in both full-parameter and PEFT fine-tuning settings by boosting convergence in the early stages of training. For instance, under the same computation time, with respect to SST2 task, PesuZO gets 9.8\% higher accuracy than MeZO (91.2\% v.s. 82.4\%). With the sliding window technique, our PseuZO achieves $70\%\sim80\%$ memory reduction compared to FO-SGD for different model sizes as PseuZO only introduced a small dimension-independent memory overhead, which enables efficient scaling of the model size. The code is available at https://github.com/YangBigMn/PseuZO. $\newcommand{\bbE}{\mathbb{E}}$

Learning dynamics is essential for model-based control and Reinforcement Learning in systems operating in changing environments, such as robotics, autonomous vehicles, and power systems. However, limited system measurements, such as those from low-resolution meters, demand sample-efficient learning. Symmetry provides a powerful inductive bias by characterizing equivalent relations in system behavior to improve sample efficiency. While recent methods attempt to discover symmetries from data, they typically assume a single global symmetry group and treat symmetry discovery and dynamic learning as separate tasks, leading to limited expressiveness and error accumulation. In this paper, we propose the Latent Mixture of Symmetries (Latent MoS), an expressive model that captures symmetry-governed latent factors from complex dynamical measurements. Latent MoS focuses on dynamic learning while locally preserving the underlying symmetric transformations. To further capture long-range temporal equivalence, we introduce a hierarchical architecture that stacks Latent MoS blocks across multiple time scales. Numerical experiments across diverse physical systems demonstrate that Latent MoS significantly outperforms state-of-the-art baselines in interpolation and extrapolation tasks while offering interpretable latent representations suitable for future geometric and safety-critical analysis.

We consider off-policy evaluation (OPE) in Partially Observable Markov Decision Processes (POMDPs) with unobserved confounding. Recent advances have introduced bridge-function to circumvent unmeasured confounding and develop estimators for the policy value, yet the statistical error bounds of them related to the length of horizon $T$ and the size of the state-action space $|\mathcal{O}||\mathcal{A}|$ remain largely unexplored. In this paper, we systematically investigate the finite-sample error bounds of OPE estimators in finite-horizon tabular confounded POMDPs. Specifically, we show that under certain rank conditions, the estimation error for policy value can achieve a rate of $\mathcal{O}(T^{1.5}/\sqrt{n})$, excluding the cardinality of the observation space $|\mathcal{O}|$ and the action space $|\mathcal{A}|$. With an additional mild condition on the concentrability coefficients in confounded POMDPs, the rate of estimation error can be improved to $\mathcal{O}(T/\sqrt{n})$. We also show that for a fully history-dependent policy, the estimation error scales as $\mathcal{O}\big(T/\sqrt{n}(|\mathcal{O}| |\mathcal{A}|)^{\frac{T}{2}}\big)$, highlighting the exponential error dependence introduced by history-based proxies to infer hidden states. Furthermore, when the target policy is memoryless policy, the error bound improves to $\mathcal{O}\big(T/\sqrt{n}\sqrt{|\mathcal{O}| |\mathcal{A}|}\big)$, which matches the optimal rate known for tabular MDPs. To the best of our knowledge, this is the first work to provide a comprehensive finite-sample analysis of OPE in confounded POMDPs.

We introduce Neural Green’s Function, a neural solution operator for linear partial differential equations (PDEs) whose differential operators admit eigendecompositions. Inspired by Green’s functions, the solution operators of linear PDEs that depend exclusively on the domain geometry, we design Neural Green’s Function to imitate their behavior, achieving superior generalization across diverse irregular geometries and source and boundary functions. Specifically, Neural Green’s Function extracts per-point features from a volumetric point cloud representing the problem domain and uses them to predict a decomposition of the solution operator, which is subsequently applied to evaluate solutions via numerical integration. Unlike recent learning-based solution operators, which often struggle to generalize to unseen source or boundary functions, our framework is, by design, agnostic to the specific functions used during training, enabling robust and efficient generalization. In the steady-state thermal analysis of mechanical part geometries from the MCB dataset, Neural Green’s Function outperforms state-of-the-art neural operators, achieving an average error reduction of 13.9% across five shape categories, while being up to 350 times faster than a numerical solver that requires computationally expensive meshing.