NeurIPS 2025 Papers

Attention is the cornerstone of modern Large Language Models (LLMs). Yet its quadratic complexity hinders efficiency and scalability, especially for long-context processing. A promising approach is to leverage sparsity in attention. However, existing sparsity-based solutions predominantly rely on predefined patterns or heuristics at the attention head level, struggling to adapt dynamically to different contexts efficiently. We propose SeerAttention, a simple yet effective attention mechanism that directly learns the block-level attention sparsity from the LLM itself. Inspired by the gating mechanism in Mixture of Experts (MoE), SeerAttention augments the conventional attention with a **learnable gate** that **selectively activates important blocks** within the attention map. Specifically, the gate first pools the query (Q) and key (K) tensors along the sequence dimension and processes them through learnable linear layers. The resulting matrices are then multiplied together to produce the gating scores, which are used to predict block-level attention sparsity. Combined with our block-sparse FlashAttention kernel, SeerAttention can achieve significant speedup on GPUs. When applied to pre-trained LLMs, SeerAttention only requires training the gate parameters in a lightweight self-distillation manner, allowing rapid convergence. Our evaluation results demonstrate that SeerAttention achieves better model accuracy and lower latency for long-context pre-filling compared to prior methods. Code is available at: https://github.com/microsoft/SeerAttention.

Multi-modal large language models (MLLMs) have emerged as a transformative approach for aligning visual and textual understanding. They typically require extremely high computational resources (e.g., thousands of GPUs) for training to achieve cross-modal alignment at multi-granularity levels. We argue that a key source of this inefficiency lies in the vision encoders they widely equip with, e.g., CLIP and SAM, which lack the alignment with language at multi-granularity levels. To address this issue, in this paper, we leverage hyperbolic space, which inherently models hierarchical levels and thus provides a principled framework for bridging the granularity gap between visual and textual modalities at an arbitrary granularity level. Concretely, we propose an efficient training paradigm for MLLMs, dubbed as \blg, which can optimize visual representations to align with their textual counterparts at an arbitrary granularity level through dynamic hyperbolic radius adjustment in hyperbolic space. \alg employs learnable matrices with M\"{o}bius multiplication operations, implemented via three effective configurations: diagonal scaling matrices, block-diagonal matrices, and banded matrices, providing a flexible yet efficient parametrization strategy. Comprehensive experiments across multiple MLLM benchmarks demonstrate that \alg consistently improves both existing pre-training and fine-tuning MLLMs clearly with less than 1\% additional parameters.

Visual Autoregressive (VAR) modeling has garnered significant attention for its innovative next-scale prediction approach, which yields substantial improvements in efficiency, scalability, and zero-shot generalization. Nevertheless, the coarse-to-fine methodology inherent in VAR results in exponential growth of the KV cache during inference, causing considerable memory consumption and computational redundancy. To address these bottlenecks, we introduce ScaleKV, a novel KV cache compression framework tailored for VAR architectures. ScaleKV leverages two critical observations: varying cache demands across transformer layers and distinct attention patterns at different scales. Based on these insights, ScaleKV categorizes transformer layers into two functional groups: drafters and refiners. Drafters exhibit dispersed attention across multiple scales, thereby requiring greater cache capacity. Conversely, refiners focus attention on the current token map to process local details, consequently necessitating substantially reduced cache capacity. ScaleKV optimizes the multi-scale inference pipeline by identifying scale-specific drafters and refiners, facilitating differentiated cache management tailored to each scale. Evaluation on the state-of-the-art text-to-image VAR model family, Infinity, demonstrates that our approach effectively reduces the required KV cache memory to 10% while preserving pixel-level fidelity.

Linear attention methods offer Transformers $O(N)$ complexity but typically underperform standard softmax attention. We identify two fundamental limitations affecting these approaches: the restriction to convex combinations that only permits additive information blending, and uniform accumulated weight bias that dilutes attention in long contexts. We propose Zero-Sum Linear Attention (ZeroS), which addresses these limitations by removing the constant zero-order term $1/t$ and reweighting the remaining zero-sum softmax residuals. This modification creates mathematically stable weights, enabling both positive and negative values and allowing a single attention layer to perform contrastive operations. While maintaining $O(N)$ complexity, ZeroS theoretically expands the set of representable functions compared to convex combinations. Empirically, it matches or exceeds standard softmax attention across various sequence modeling benchmarks.

Achieving seamless viewing of high-fidelity volumetric video, comparable to 2D video experiences, remains an open challenge. Existing volumetric video compression methods either lack the flexibility to adjust quality and bitrate within a single model for efficient streaming across diverse networks and devices, or struggle with real-time decoding and rendering on lightweight mobile platforms. To address these challenges, we introduce 4DGCPro, a novel hierarchical 4D Gaussian compression framework that facilitates real-time mobile decoding and high-quality rendering via progressive volumetric video streaming in a single bitstream. Specifically, we propose a perceptually-weighted and compression-friendly hierarchical 4D Gaussian representation with motion-aware adaptive grouping to reduce temporal redundancy, preserve coherence, and enable scalable multi-level detail streaming. Furthermore, we present an end-to-end entropy-optimized training scheme, which incorporates layer-wise rate-distortion (RD) supervision and attribute-specific entropy modeling for efficient bitstream generation. Extensive experiments show that 4DGCPro enables flexible quality and variable bitrate within a single model, achieving real-time decoding and rendering on mobile devices while outperforming existing methods in RD performance across multiple datasets.

Despite their proficiency in various language tasks, Large Language Models (LLMs) struggle with combinatorial problems like Satisfiability, Traveling Salesman Problem, or even basic arithmetic. We address this gap through a novel approach for solving problems in the class NP. We focus on the paradigmatic task of Sudoku and achieve state-of-the-art accuracy (99\%) compared to prior neuro-symbolic approaches. Unlike prior work that used custom architectures, our method employs a vanilla decoder-only Transformer (GPT-2) without external tools or function calling. Our method integrates imitation learning of simple Sudoku rules with an explicit Depth-First Search (DFS) exploration strategy involving informed guessing and backtracking. Moving beyond imitation learning, we seek to minimize the number of guesses until reaching a solution. We provide a rigorous analysis of this setup by formalizing its connection to a contextual variant of $\textit{Min-Sum Set Cover}$, a well-studied problem in algorithms and stochastic optimization.

Graph Neural Networks (GNNs) are pivotal in graph classification but often struggle with generalization and overfitting. We introduce a unified and efficient Graph Multi-View (GMV) learning framework that integrates multi-view learning into GNNs to enhance robustness and efficiency. Leveraging the lottery ticket hypothesis, GMV activates diverse sub-networks within a single GNN through a novel training pipeline, which includes mixed-view generation, and multi-view decomposition and learning. This approach simultaneously broadens "views" from the data, model, and optimization perspectives during training to enhance the generalization capabilities of GNNs. During inference, GMV only incorporates additional prediction heads into standard GNNs, thereby achieving multi-view learning at minimal cost. Our experiments demonstrate that GMV surpasses other augmentation and ensemble techniques for GNNs and Graph Transformers across various graph classification scenarios.

Quantization plays a crucial role in accelerating the inference of large-scale models, and rotational matrices have been shown to effectively improve quantization performance by smoothing outliers. However, end-to-end fine-tuning of rotational optimization algorithms incurs high computational costs and is prone to overfitting. To address this challenge, we propose an efficient distribution-aware rotational calibration method, DartQuant, which reduces the complexity of rotational optimization by constraining the distribution of the activations after rotation. This approach also effectively reduces reliance on task-specific losses, thereby mitigating the risk of overfitting. Additionally, we introduce the QR-Orth optimization scheme, which replaces expensive alternating optimization with a more efficient solution. In a variety of model quantization experiments, DartQuant demonstrates superior performance. Compared to existing methods, it achieves 47$\times$ acceleration and 10$\times$ memory savings for rotational optimization on a 70B model. Furthermore, it is the first to successfully complete rotational calibration for a 70B model on a single 3090 GPU, making quantization of large language models feasible in resource-constrained environments.

The attention mechanism of a transformer has a quadratic complexity, leading to high inference costs and latency for long sequences. However, attention matrices are mostly sparse, which implies that many entries may be omitted from computation for efficient inference. Sparse attention inference methods aim to reduce this computational burden; however, they also come with a troublesome performance degradation. We discover that one reason for this degradation is that the sparse calculation induces a distributional shift in the attention outputs. The distributional shift causes decoding-time queries to fail to align well with the appropriate keys from the prefill stage, leading to a drop in performance. We propose a simple, novel, and effective procedure for correcting this distributional shift, bringing the distribution of sparse attention outputs closer to that of quadratic attention. Our method can be applied on top of any sparse attention method, and results in an average 36\%pt performance increase, recovering 88\% of quadratic attention accuracy on the 131K RULER benchmark when applied on top of sliding window attention with sink tokens while only adding a small overhead. Our method can maintain approximately 98.5\% sparsity over full quadratic attention, making our model 32 times faster than Flash Attention 2 when processing 1M token prefills.

Residual connections are pivotal for deep neural networks, enabling greater depth by mitigating vanishing gradients. However, in standard residual updates, the module’s output is directly added to the input stream. This can lead to updates that predominantly reinforce or modulate the existing stream direction, potentially underutilizing the module’s capacity for learning entirely novel features. In this work, we introduce _Orthogonal Residual Update_: we decompose the module’s output relative to the input stream and add only the component orthogonal to this stream. This design aims to guide modules to contribute primarily new representa-tional directions, fostering richer feature learning while promoting more efficient training. We demonstrate that our orthogonal update strategy improves generalization accuracy and training stability across diverse architectures (ResNetV2, Vision Transformers) and datasets (CIFARs, TinyImageNet, ImageNet-1k), achieving, for instance, a +3.78 pp Acc@1 gain for ViT-B on ImageNet-1k. Code and models are available at https://github.com/BootsofLagrangian/ortho-residual.

Training high-performing Small Language Models (SLMs) remains computationally expensive, even with knowledge distillation and pruning from larger teacher models. Existing approaches often face three key challenges: (1) information loss from hard pruning, (2) inefficient alignment of representations, and (3) underutilization of informative activations, particularly from Feed-Forward Networks (FFNs). To address these challenges, we introduce \textbf{Low-Rank Clone (LRC)}, an efficient pre-training method that constructs SLMs aspiring to behavioral equivalence with strong teacher models. LRC trains a set of low-rank projection matrices that jointly enable soft pruning by compressing teacher weights, and activation clone by aligning student activations, including FFN signals, with those of the teacher. This unified design maximizes knowledge transfer while removing the need for explicit alignment modules. Extensive experiments with open-source teachers such as Llama-3.2-3B-Instruct and Qwen2.5-3B/7B-Instruct show that LRC matches or surpasses the performance of state-of-the-art models trained on trillions of tokens--using only 20B tokens, achieving over \textbf{1,000$\times$} greater training efficiency. Our codes and model checkpoints are available at https://github.com/CURRENTF/LowRankClone and https://huggingface.co/JitaiHao/LRC-4B-Base.

As Large Language Models (LLMs) are frequently updated, LoRA weights trained on earlier versions quickly become obsolete. The conventional practice of retraining LoRA weights from scratch on the latest model is costly, time-consuming, and environmentally detrimental, particularly as the diversity of LLMs and downstream tasks expands. This motivates a critical question: "How can we efficiently leverage existing LoRA weights to adapt to newer model versions?" To address this, we propose LoRASuite, a modular approach tailored specifically to various types of LLM updates. First, we compute a transfer matrix utilizing known parameters from both old and new LLMs. Next, we allocate corresponding layers and attention heads based on centered kernel alignment and cosine similarity metrics, respectively. A subsequent small-scale, skillful fine-tuning step ensures numerical stability. Experimental evaluations demonstrate that LoRASuite consistently surpasses small-scale vanilla LoRA methods. Notably, on backbone LLMs such as MiniCPM and Qwen, LoRASuite even exceeds the performance of full-scale LoRA retraining, with average improvements of +1.4 and +6.6 points on math tasks, respectively. Additionally, LoRASuite significantly reduces memory consumption by 5.5 GB and computational time by 78.23%.

Flow and diffusion-based models have emerged as powerful tools for scientific applications, particularly for sampling non-normalized probability distributions, as exemplified by Boltzmann Generators (BGs). A critical challenge in deploying these models is their reliance on sample likelihood computations, which scale prohibitively with system size $n$, often rendering them infeasible for large-scale problems. To address this, we introduce $\textit{HollowFlow}$, a flow-based generative model leveraging a novel non-backtracking graph neural network (NoBGNN). By enforcing a block-diagonal Jacobian structure, HollowFlow likelihoods are evaluated with a constant number of backward passes in $n$, yielding speed-ups of up to $\mathcal{O}(n^2)$: a significant step towards scaling BGs to larger systems. Crucially, our framework generalizes: $\textbf{any equivariant GNN or attention-based architecture}$ can be adapted into a NoBGNN. We validate HollowFlow by training BGs on two different systems of increasing size. For both systems, the sampling and likelihood evaluation time decreases dramatically, following our theoretical scaling laws. For the larger system we obtain a $10^2\times$ speed-up, clearly illustrating the potential of HollowFlow-based approaches for high-dimensional scientific problems previously hindered by computational bottlenecks.

This paper introduces the Convex Potential Mirror Langevin Algorithm (CPMLA), a novel method to improve sampling efficiency for Energy-Based Models (EBMs). CPMLA uses mirror Langevin dynamics with a convex potential flow as a dynamic mirror map for EBM sampling. This dynamic mirror map enables targeted geometric exploration on the data manifold, accelerating convergence to the target distribution. Theoretical analysis proves that CPMLA achieves exponential convergence with vanishing bias under relaxed log-concave conditions, supporting its efficiency in adapting to complex data distributions. Experiments on benchmarks like CIFAR-10, SVHN, and CelebA demonstrate CPMLA's improved sampling quality and inference efficiency over existing techniques.

Efficient deployment of small language models (SLMs) is essential for numerous real-world applications with stringent latency constraints.While previous work on SLM design has primarily focused on reducing the number of parameters to achieve parameter-optimal SLMs, parameter efficiency does not necessarily translate into proportional real-device speed-ups. This work aims to identify the key determinants of SLMs' real-device latency and offer generalizable principles and methodologies for SLM design and training when real-device latency is the primary consideration. Specifically, we identify two central architectural factors: depth–width ratios and operator choices. The former is crucial for small-batch-size latency, while the latter affects both latency and large-batch-size throughput. In light of this, we first study latency-optimal depth–width ratios, with the key finding that although deep–thin models generally achieve better accuracy under the same parameter budget, they may not lie on the accuracy–latency trade-off frontier. Next, we explore emerging efficient attention alternatives to evaluate their potential as candidate building operators. Using the identified promising operators, we construct an evolutionary search framework to automatically discover latency-optimal combinations of these operators within hybrid SLMs, thereby advancing the accuracy–latency frontier. In addition to architectural improvements, we further enhance SLM training using a weight normalization technique that enables more effective weight updates and improves final convergence. This technique can serve as a generalizable component for future SLMs. Combining these methods, we introduce a new family of hybrid SLMs, called Nemotron-Flash, which significantly advances the accuracy–efficiency frontier of state-of-the-art SLMs, e.g., achieving over +5.5\% average accuracy, 1.3$\times$/1.9$\times$ lower latency, and 18.7$\times$/45.6$\times$ higher throughput compared to Qwen3-1.7B/0.6B, respectively.

Shapley values have emerged as a critical tool for explaining which features impact the decisions made by machine learning models. However, computing exact Shapley values is difficult, generally requiring an exponential (in the feature dimension) number of model evaluations. To address this, many model-agnostic randomized estimators have been developed, the most influential and widely used being the KernelSHAP method (Lundberg & Lee, 2017). While related estimators such as unbiased KernelSHAP (Covert & Lee, 2021) and LeverageSHAP (Musco & Witter, 2025) are known to satisfy theoretical guarantees, bounds for KernelSHAP have remained elusive. We describe a broad and unified framework that encompasses KernelSHAP and related estimators constructed using both with and without replacement sampling strategies. We then prove strong non-asymptotic theoretical guarantees that apply to all estimators from our framework. This provides, to the best of our knowledge, the first theoretical guarantees for KernelSHAP and sheds further light on tradeoffs between existing estimators. Through comprehensive benchmarking on small and medium dimensional datasets for Decision-Tree models, we validate our approach against exact Shapley values, consistently achieving low mean squared error with modest sample sizes. Furthermore, we make specific implementation improvements to enable scalability of our methods to high-dimensional datasets. Our methods, tested on datasets such MNIST and CIFAR10, provide consistently better results compared to the KernelSHAP library.

Current Reinforcement Fine-tuning (RFT) paradigms for Large Language Models (LLMs) suffer from sample inefficiency due to the redundant exposure of identical queries under uniform data sampling. While previous work has explored curriculum learning via heuristic difficulty metrics, these strategies exhibit limitations by neglecting the intrinsic learning signals generated by the model itself, thus leading to suboptimal training regimes. In this paper, we identify a model-inherent signal termed *angle concentration* that effectively reflects an LLM's capacity to learn from specific data. We theoretically and empirically demonstrate a correlation between the angular distribution of token hidden state vectors and the resulting gradient, revealing a learning preference for data exhibiting higher angle concentration. Inspired by this finding, we propose GAIN-RL, a Gradient-driven Angle-Informed Navigated RL framework. By leveraging the model's intrinsic angle concentration signal, GAIN-RL dynamically selects training data in each epoch, ensuring consistently impactful gradient updates and thus significantly enhancing overall training efficiency. Empirical evaluations show that GAIN-RL (GRPO) achieves over a 2.5$\times$ acceleration in training efficiency across diverse mathematical and coding tasks and varying model scales. Furthermore, GAIN-RL (GRPO)'s efficient sampling yields data-efficient training, achieving better performance with half the original data compared to vanilla GRPO with full training data.

To establish last-iterate convergence for Counterfactual Regret Minimization (CFR) algorithms in learning a Nash equilibrium (NE) of extensive-form games (EFGs), recent studies reformulate learning an NE of the original EFG as learning the NEs of a sequence of (perturbed) regularized EFGs. Hence, proving last-iterate convergence in solving the original EFG reduces to proving last-iterate convergence in solving (perturbed) regularized EFGs. However, these studies only establish last-iterate convergence for Online Mirror Descent (OMD)-based CFR algorithms instead of Regret Matching (RM)-based CFR algorithms in solving perturbed regularized EFGs, resulting in a poor empirical convergence rate, as RM-based CFR algorithms typically outperform OMD-based CFR algorithms. In addition, as solving multiple perturbed regularized EFGs is required, fine-tuning across multiple perturbed regularized EFGs is infeasible, making parameter-free algorithms highly desirable. This paper show that CFR$^+$, a classical parameter-free RM-based CFR algorithm, achieves last-iterate convergence in learning an NE of perturbed regularized EFGs. This is the first parameter-free last-iterate convergence for RM-based CFR algorithms in perturbed regularized EFGs. Leveraging CFR$^+$ to solve perturbed regularized EFGs, we get Reward Transformation CFR$^+$ (RTCFR$^+$). Importantly, we extend prior work on the parameter-free property of CFR$^+$, enhancing its stability, which is vital for the empirical convergence of RTCFR$^+$. Experiments show that RTCFR$^+$ exhibits a significantly faster empirical convergence rate than existing algorithms that achieve theoretical last-iterate convergence. Interestingly, RTCFR$^+$ show performance no worse than average-iterate convergence CFR algorithms. It is the first last-iterate convergence algorithm to achieve such performance. Our code is available at https://github.com/menglinjian/NeurIPS-2025-RTCFR.

Accurately estimating *semantic* aleatoric and epistemic uncertainties in large language models (LLMs) is particularly challenging in free-form question answering (QA), where obtaining stable estimates often requires many expensive generations. We introduce a **diversity-steered sampler** that discourages semantically redundant outputs during decoding, covers both autoregressive and masked diffusion paradigms, and yields substantial sample-efficiency gains. The key idea is to inject a continuous semantic-similarity penalty into the model’s proposal distribution using a natural language inference (NLI) model lightly finetuned on partial prefixes or intermediate diffusion states. We debias downstream uncertainty estimates with importance reweighting and shrink their variance with control variates. Across four QA benchmarks, our method matches or surpasses baselines while covering more semantic clusters with the same number of samples. Being modular and requiring no gradient access to the base LLM, the framework promises to serve as a drop-in enhancement for uncertainty estimation in risk-sensitive model deployments.