NeurIPS 2025 论文列表

With the growing demand for deploying large language models (LLMs) across diverse applications, improving their inference efficiency is crucial for sustainable and democratized access. However, retraining LLMs to meet new user-specific requirements is prohibitively expensive and environmentally unsustainable. In this work, we propose a practical and scalable alternative: composing efficient hybrid language models from existing pre-trained models. Our approach, X-EcoMLA, introduces a family of 1B, 3B, and 8B hybrid models by combining State Space Models (SSMs) and Multi-head Latent Attention (MLA) layers, using a refined initialization and post-training pipeline to efficiently transfer knowledge from pre-trained Transformers. X-EcoMLA achieves Transformer-level accuracy with near-SSM efficiency using only 7–11 billion training tokens (compared to the trillions required for pre-training) and an 8B teacher. Moreover, it dramatically reduces KV cache size—down to 3.9%, 2%, and 2.73% of the original for the 1B, 3B, and 8B variants, respectively—while preserving 100%, 100%, and over 97% of average zero-shot performance on LM Harness tasks. Compared to models like MambaInLLaMA, X-EcoMLA, Minitron, and Llamba, our approach consistently delivers competitive or superior accuracy while using significantly fewer tokens, smaller teachers, and vastly reduced KV cache memory. Notably, X-EcoMLA-8B surpasses Minitron-8B in few-shot accuracy by 7%, while using 8× fewer training tokens, over 12× smaller KV cache, and a smaller teacher (8B vs. 15B). It also achieves 1.4x–3.3x higher throughput (tokens/s) than MambaInLlama. The source code is released at https://github.com/AMD-AGI/AMD-Hybrid-Models.

We demonstrate, for the first time, fully quantized training (FQT) of large language models (LLMs) using predominantly 4-bit floating-point (FP4) precision for weights, activations, and gradients on datasets up to 200 billion tokens. We extensively investigate key design choices for FP4, including block sizes, scaling formats, and rounding methods. Our analysis shows that the NVFP4 format, where each block of 16 FP4 values (E2M1) shares a scale represented in E4M3, provides optimal results. We use stochastic rounding for backward and update passes and round-to-nearest for the forward pass to enhance stability. Additionally, we identify a theoretical and empirical threshold for effective quantized training: when the gradient norm falls below approximately $\sqrt{3}$ times the quantization noise, quantized training becomes less effective. Leveraging these insights, we successfully train a 7-billion-parameter model on 256 Intel Gaudi2 accelerators. The resulting FP4-trained model achieves downstream task performance comparable to a standard BF16 baseline, confirming that FP4 training is a practical and highly efficient approach for large-scale LLM training. A reference implementation is supplied in https://github.com/Anonymous1252022/fp4-all-the-way.

Visual prompting (VP) has emerged as a promising parameter-efficient fine-tuning approach for adapting pre-trained vision models to downstream tasks without modifying model parameters. Despite offering advantages like negligible computational overhead and compatibility with black-box models, conventional VP methods typically achieve lower accuracy than other adaptation approaches. Our analysis reveals two critical limitations: the restricted expressivity of simple additive transformation and a tendency toward overfitting when the parameter count increases. To address these challenges, we propose ACAVP (Affine, Color, and Additive Visual Prompting), which enhances VP's expressive power by introducing complementary transformation operations: affine transformation for creating task-specific prompt regions while preserving original image information, and color transformation for emphasizing task-relevant visual features. Additionally, we identify that overfitting is a critical issue in VP training and introduce TrivialAugment as an effective data augmentation, which not only benefits our approach but also significantly improves existing VP methods, with performance gains of up to 12 percentage points on certain datasets. This demonstrates that appropriate data augmentation is universally beneficial for VP training. Extensive experiments across twelve diverse image classification datasets with two different model architectures demonstrate that ACAVP achieves state-of-the-art accuracy among VP methods, surpasses linear probing in average accuracy, and exhibits superior robustness to distribution shifts, all while maintaining minimal computational overhead during inference.

The Koopman operator provides a principled framework for analyzing nonlinear dynamical systems through linear operator theory. Recent advances in dynamic mode decomposition (DMD) have shown that trajectory data can be used to identify dominant modes of a system in a data-driven manner. Building on this idea, deep learning methods such as VAMPnet and DPNet have been proposed to learn the leading singular subspaces of the Koopman operator. However, these methods require backpropagation through potentially numerically unstable operations on empirical second moment matrices, such as singular value decomposition and matrix inversion, during objective computation, which can introduce biased gradient estimates and hinder scalability to large systems. In this work, we propose a scalable and conceptually simple method for learning the top-$k$ singular functions of the Koopman operator for stochastic dynamical systems based on the idea of low-rank approximation. Our approach eliminates the need for unstable linear-algebraic operations and integrates easily into modern deep learning pipelines. Empirical results demonstrate that the learned singular subspaces are both reliable and effective for downstream tasks such as eigen-analysis and multi-step prediction.

Recent advances in multimodal semantic segmentation show that incorporating auxiliary inputs—such as depth or thermal images—can significantly improve performance over single-modality (RGB-only) approaches. However, most existing solutions rely on parallel backbone networks and complex fusion modules, greatly increasing model size and computational demands. Inspired by prompt tuning in large language models, we introduce \textbf{MixPrompt}: a prompting-based framework that integrates auxiliary modalities into a pretrained RGB segmentation model without modifying its architecture. MixPrompt uses a lightweight prompting module to extract and fuse information from auxiliary inputs into the main RGB backbone. This module is initialized using the early layers of a pretrained RGB feature extractor, ensuring a strong starting point. At each backbone layer, MixPrompt aligns RGB and auxiliary features in multiple low-rank subspaces, maximizing information use with minimal parameter overhead. An information mixing scheme enables cross-subspace interaction for further performance gains. During training, only the prompting module and segmentation head are updated, keeping the RGB backbone frozen for parameter efficiency. Experiments across NYU Depth V2, SUN-RGBD, MFNet, and DELIVER datasets show that MixPrompt achieves improvements of 4.3, 1.1, 0.4, and 1.1 mIoU, respectively, over two-branch baselines, while using nearly half the parameters. MixPrompt also outperforms recent prompting-based methods under similar compute budgets.

Associative memory models based on Hopfield networks and self-attention based on key-value mechanisms have been popular approaches in the study of memory mechanisms in deep learning. It has been pointed out that the state update rule of the modern Hopfield network (MHN) in the adiabatic approximation is in agreement with the self-attention layer of Transformer. In this paper, we go beyond this approximation and investigate the relationship between MHN and self-attention. Our results show that the correspondence between Hopfield networks and Transformers can be established in a more generalized form by adding a new variable, the hidden state derived from the MHN, to self-attention. This new attention mechanism, modern Hopfield attention (MHA), allows the inheritance of attention scores from the input layer of the Transformer to the output layer, which greatly improves the nature of attention weights. In particular, we show both theoretically and empirically that MHA hidden states significantly improve serious problem of deep Transformers known as rank collapse and token uniformity. We also confirm that MHA can systematically improve accuracy without adding training parameters to the Vision Transformer or GPT. Our results provide a new case in which Hopfield networks can be a useful perspective for improving the Transformer architecture.

We show that low-rank adaptation of large-scale models suffers from a low stable rank that is well below the linear algebraic rank of the subspace, degrading fine-tuning performance. To mitigate the underutilization of the allocated subspace, we propose PoLAR, a parameterization inspired by the polar decomposition that factorizes the low-rank update into two direction matrices constrained to Stiefel manifolds and an unconstrained scale matrix. Our theory shows that PoLAR yields an exponentially faster convergence rate on a canonical low-rank adaptation problem. Pairing the parameterization with Riemannian optimization leads to consistent gains on three different benchmarks testing general language understanding, commonsense reasoning, and mathematical problem solving with base model sizes ranging from 350M to 27B.

Estimating the uncertainty of responses from Large Language Models (LLMs) remains a critical challenge. While recent Bayesian methods have demonstrated effectiveness in quantifying uncertainty through low-rank weight updates, they typically require complex fine-tuning or post-training procedures. In this paper, we propose **T**raining-**F**ree **B**ayesianization (**TFB**), a simple yet theoretically grounded framework that efficiently transforms trained low-rank adapters into Bayesian ones without additional training. TFB systematically searches for the maximally acceptable level of variance in the weight posterior, constrained within a family of low-rank isotropic Gaussian distributions. Our theoretical analysis shows that under mild conditions, this search process is equivalent to KL-regularized variational optimization, a generalized form of variational inference. Through comprehensive experiments, we show that TFB achieves superior uncertainty estimation and generalization compared to existing methods while eliminating the need for complex Bayesianization training procedures.

We study population convergence guarantees of stochastic gradient descent (SGD) for smooth convex objectives in the interpolation regime, where the noise at optimum is zero or near zero. The behavior of the last iterate of SGD in this setting---particularly with large (constant) stepsizes---has received growing attention in recent years due to implications for the training of over-parameterized models, as well as to analyzing forgetting in continual learning and to understanding the convergence of the randomized Kaczmarz method for solving linear systems. We establish that after $T$ steps of SGD on $\beta$-smooth convex loss functions with stepsize $0 < \eta < 2/\beta$, the last iterate exhibits expected excess risk $\widetilde{O}(\tfrac{1}{\eta (2-\beta \eta) T^{1-\beta\eta/2}} + \tfrac{\eta}{(2-\beta\eta)^2} T^{\beta\eta/2} \sigma_\star^2)$, where $\sigma_\star^2$ denotes the variance of the stochastic gradients at the optimum. In particular, for a well-tuned stepsize we obtain a near optimal $\widetilde{O}(1/T + \sigma_\star/\sqrt T)$ rate for the last iterate, extending the results of Varre et al. (2021) beyond least squares regression; and when $\sigma_\star=0$ we obtain a rate of $O(1/\sqrt T)$ with $\eta=1/\beta$, improving upon the best-known $O(T^{-1/4})$ rate recently established by Evron et al. (2025) in the special case of realizable linear regression.

On-device large models (LMs) reduce cloud dependency but expose proprietary model weights to the end-user, making them vulnerable to white-box model stealing (MS) attacks. A common defense is TEE-Shielded DNN Partition (TSDP), which places all trainable LoRA adapters (fine tuned on private data) inside a trusted execution environment (TEE). However, this design suffers from excessive host-to-TEE communication latency. We propose AegisGuard, a fine tuning and deployment framework that selectively shields the MS sensitive adapters while offloading the rest to the GPU, balancing security and efficiency. AegisGuard integrates two key components: i) RL-based Sensitivity Measurement (RSM), which injects Gaussian noise during training and applies a lightweight reinforcement learning to rank adapters based on their impact on model stealing; and (ii) Shielded-Adapter Compression (SAC), which structurally prunes the selected adapters to reduce both parameter size and intermediate feature maps, further lowering TEE computation and data transfer costs. Extensive experiments demonstrate that AegisGuard achieves black-box level MS resilience (surrogate accuracy around 39%, matching fully shielded baselines), while reducing end-to-end inference latency by 2–3× and cutting TEE memory usage by 4× compared to state-of-the-art TSDP methods.

Low-Rank Adaptation (LoRA) has become the de facto parameter-efficient fine-tuning (PEFT) method for large language models (LLMs) by constraining weight updates to low-rank matrices. Recent works such as Tied-LoRA, VeRA, and VB-LoRA push efficiency further by introducing additional constraints to reduce the trainable parameter space. In this paper, we show that the parameter space reduction strategies employed by these LoRA variants can be formulated within a unified framework, Uni-LoRA, where the LoRA parameter space, flattened as a high-dimensional vector space R^D, can be reconstructed through a projection from a subspace R^d, with d << D. We demonstrate that the fundamental difference among various LoRA methods lies in the choice of the projection matrix, P ∈ R^{D×d}. Most existing LoRA variants rely on layer-wise or structure-specific projections that limit cross-layer parameter sharing, thereby compromising parameter efficiency. In light of this, we introduce an efficient and theoretically grounded projection matrix that is isometric, enabling global parameter sharing and reducing computation overhead. Furthermore, under the unified view of Uni-LoRA, this design requires only a single trainable vector to reconstruct LoRA parameters for the entire LLM -- making Uni-LoRA both a unified framework and a “one-vector-only” solution. Extensive experiments on GLUE, mathematical reasoning, and instruction tuning benchmarks demonstrate that Uni-LoRA achieves state-of-the-art parameter efficiency while outperforming or matching prior approaches in predictive performance.

Variations in Magnetic resonance imaging (MRI) scanners and acquisition protocols cause distribution shifts that degrade reconstruction performance on unseen data. Test-time adaptation (TTA) offers a promising solution to address this discrepancies. However, previous single-shot TTA approaches are inefficient due to repeated training and suboptimal distributional models. Self-supervised learning methods may risk over-smoothing in scarce data scenarios. To address these challenges, we propose a novel Dual-Stage Distribution and Slice Adaptation (D2SA) via MRI implicit neural representation (MR-INR) to improve MRI reconstruction performance and efficiency, which features two stages. In the first stage, an MR-INR branch performs patient-wise distribution adaptation by learning shared representations across slices and modelling patient-specific shifts with mean and variance adjustments. In the second stage, single-slice adaptation refines the output from frozen convolutional layers with a learnable anisotropic diffusion module, preventing over-smoothing and reducing computation. Experiments across five MRI distribution shifts demonstrate that our method can integrate well with various self-supervised learning (SSL) framework, improving performance and accelerating convergence under diverse conditions.

In this paper, we explore the asymptotically optimal tuning parameter choice in ridge regression for estimating nuisance functions of a statistical functional that has recently gained prominence in conditional independence testing and causal inference. Given a sample of size $n$, we study estimators of the Expected Conditional Covariance (ECC) between variables $Y$ and $A$ given a high-dimensional covariate $X \in \mathbb{R}^p$. Under linear regression models for $Y$ and $A$ on $X$ and the proportional asymptotic regime $p/n \to c \in (0, \infty)$, we evaluate three existing ECC estimators and two sample splitting strategies for estimating the required nuisance functions. Since no consistent estimator of the nuisance functions exists in the proportional asymptotic regime without imposing further structure on the problem, we first derive debiased versions of the ECC estimators that utilize the ridge regression nuisance function estimators. We show that our bias correction strategy yields $\sqrt{n}$-consistent estimators of the ECC across different sample splitting strategies and estimator choices. We then derive the asymptotic variances of these debiased estimators to illustrate the nuanced interplay between the sample splitting strategy, estimator choice, and tuning parameters of the nuisance function estimators for optimally estimating the ECC. Our analysis reveals that prediction-optimal tuning parameters (i.e., those that optimally estimate the nuisance functions) may not lead to the lowest asymptotic variance of the ECC estimator -- thereby demonstrating the need to be careful in selecting tuning parameters based on the final goal of inference. Finally, we verify our theoretical results through extensive numerical experiments.

We introduce a new family of algorithms for detecting and estimating a rank-one signal from a noisy observation under prior information about that signal's direction, focusing on examples where the signal is known to have entries biased to be positive. Given a matrix observation $\mathbf{Y}$, our algorithms construct a *nonlinear Laplacian*, another matrix of the form $\mathbf{Y} + \mathrm{diag}(\sigma(\mathbf{Y1}))$ for a nonlinear $\sigma: \mathbb{R} \to \mathbb{R}$, and examine the top eigenvalue and eigenvector of this matrix. When $\mathbf{Y}$ is the (suitably normalized) adjacency matrix of a graph, our approach gives a class of algorithms that search for unusually dense subgraphs by computing a spectrum of the graph "deformed" by the degree profile $\mathbf{Y1}$. We study the performance of such algorithms compared to direct spectral algorithms (the case $\sigma = 0$) on models of sparse principal component analysis with biased signals, including the Gaussian planted submatrix problem. For such models, we rigorously characterize the strength of rank-one signal, as a function of the nonlinearity $\sigma$, required for an outlier eigenvalue to appear in the spectrum of a nonlinear Laplacian matrix. While identifying the $\sigma$ that minimizes the required signal strength in closed form seems intractable, we explore three approaches to design $\sigma$ numerically: exhaustively searching over simple classes of $\sigma$, learning $\sigma$ from datasets of problem instances, and tuning $\sigma$ using black-box optimization of the critical signal strength. We find both theoretically and empirically that, if $\sigma$ is chosen appropriately, then nonlinear Laplacian spectral algorithms substantially outperform direct spectral algorithms, while retaining the conceptual simplicity of spectral methods compared to broader classes of computations like approximate message passing or general first order methods.

Large Language Models (LLMs) have yet to effectively leverage the vast amounts of edge-device data, and Federated Learning (FL) offers a promising paradigm to collaboratively fine-tune LLMs without transferring private edge data to the cloud. To operate within the computational and communication constraints of edge devices, recent literature on federated fine-tuning of LLMs proposes the use of low-rank adaptation (LoRA) and similar parameter-efficient methods. However, LoRA-based methods suffer from accuracy degradation in FL settings, primarily because of data and computational heterogeneity across clients. We propose Ravan, an adaptive multi-head LoRA method that balances parameter efficiency and model expressivity by reparameterizing the weight updates as the sum of multiple LoRA heads, $s_i\textbf{B}_i\textbf{H}_i\textbf{A}_i$, in which only the $\textbf{H}_i$ parameters and their lightweight scaling factors $s_i$ are trained. These trainable scaling factors let the optimization focus on the most useful heads, recovering a higher-rank approximation of the full update without increasing the number of communicated parameters since clients upload $s_i\textbf{H}_i$ directly. Experiments on vision and language benchmarks show that Ravan improves test accuracy by 2–8\% over prior parameter-efficient baselines, making it a robust and scalable solution for federated fine-tuning of LLMs.

The scaling of model and data sizes has reshaped the AI landscape, establishing finetuning pretrained models as the standard paradigm for solving downstream tasks. However, dominant finetuning methods typically rely on weight adaptation, often lack interpretability, and depend on heuristically chosen hyperparameters. In this paper, we take a different perspective and shift the focus from weights to activation functions, viewing them through the lens of spline operators. We propose Curvature Tuning (CT), an interpretable and principled steering method that modulates a model's decision boundary by injecting a single hyperparameter into its activation functions. We show that CT provably adjusts model decision boundary curvature and, more fundamentally, projects a model onto a space of smooth functions—thereby complementing current finetuning methods, whose effect lies primarily in feature adaptation. Making this hyperparameter trainable gives rise to a novel and highly parameter-efficient finetuning method. Empirically, CT improves both generalization and robustness. For example, it boosts downstream accuracy of ResNet-50/152 by 8.59\%/8.34\% over linear probing and 4.64\%/1.70\% over LoRA across 12 datasets, and improves robust accuracy on the $\ell_{\infty}$ benchmark from RobustBench by 1032.64\%/1494.46\%. Our code is available at https://github.com/Leon-Leyang/curvature-tuning.

Although transformer-based models have shown exceptional empirical performance, the fundamental principles governing their training dynamics are inadequately characterized beyond configuration-specific studies. Inspired by empirical evidence showing improved reasoning capabilities under small initialization scales in language models, we employ the gradient flow analytical framework established in \cite{zhou2022towards} to systematically investigate linearized Transformer training dynamics. Our theoretical analysis dissects the dynamics of attention modules into two distinct stages. In the first stage, asymmetric weight perturbations from random initialization sustain non-degenerate gradient dynamics in parameter matrices, facilitating systematic escape from small initialization regimes. Subsequently, these matrices undergo condensation, progressively aligning toward the target orientation. In the second stage, the previously static key-query matrices actively participate in training, driving the normalized matrices toward asymptotic rank collapse. This two-stage framework generalizes classical directional convergence results.

Adapting large-scale foundation models in multi-task scenarios often suffers from task conflict and oblivion. To mitigate such issues, we propose a novel "model MoE-ization" strategy that leads to a conflict- and oblivion-resistant multi-task adaptation method. Given a weight matrix of a pre-trained model, our method applies SVD to it and introduces a learnable router to adjust its singular values based on tasks and samples. Accordingly, the weight matrix becomes a Mixture of Orthogonal Rank-one Experts (MoORE), in which each expert corresponds to the outer product of a left singular vector and the corresponding right one. We can improve the model capacity by imposing a learnable orthogonal transform on the right singular vectors. Unlike low-rank adaptation (LoRA) and its MoE-driven variants, MoORE guarantees the experts' orthogonality and maintains the column space of the original weight matrix. These two properties make the adapted model resistant to the conflicts among the new tasks and the oblivion of its original tasks, respectively. Experiments on various datasets demonstrate that MoORE outperforms existing multi-task adaptation methods consistently, showing its superiority in terms of conflict- and oblivion-resistance. The code is available at https://github.com/DaShenZi721/MoORE.

Modern Parameter-Efficient Fine-Tuning (PEFT) methods such as low-rank adaptation (LoRA) reduce the cost of customizing large language models (LLMs), yet still require a separate optimization run for every downstream dataset. We introduce \textbf{Drag-and-Drop LLMs (\textit{DnD})}, a prompt-conditioned parameter generator that eliminates per-task training by mapping a handful of unlabeled task prompts directly to LoRA weight updates. A lightweight text encoder distills each prompt batch into condition embeddings, which are then transformed by a cascaded hyper-convolutional decoder into the full set of LoRA matrices. Once trained in a diverse collection of prompt-checkpoint pairs, DnD produces task-specific parameters in seconds, yielding i) up to \textbf{12,000$\times$} lower overhead than full fine-tuning, ii) average gains up to \textbf{30\%} in performance over the strongest training LoRAs on unseen common-sense reasoning, math, coding, and multimodal benchmarks, and iii) robust cross-domain generalization improving \textbf{40\%} performance without access to the target data or labels. Our results demonstrate that prompt-conditioned parameter generation is a viable alternative to gradient-based adaptation for rapidly specializing LLMs. We open source \href{https://jerryliang24.github.io/DnD}{our project} in support of future research.