NeurIPS 2025 论文列表

Large-scale datasets are usually required to train deep neural networks; however, they increase computational complexity, hindering practical applications. Recently, dataset distillation for images and texts has attracted considerable attention, as it reduces the original dataset to a small synthetic one to alleviate the computational burden of training while preserving essential task-relevant information. However, dataset distillation for 3D point clouds remains largely unexplored, as point clouds exhibit fundamentally different characteristics from those of images, making this task more challenging. In this paper, we propose a distribution-matching-based distillation framework for 3D point clouds that jointly optimizes the geometric structures and orientations of synthetic 3D objects. To address the semantic misalignment caused by the unordered nature of point clouds, we introduce a Semantically Aligned Distribution Matching (SADM) loss, which is computed on the sorted features within each channel. Moreover, to handle rotational variations, we jointly learn optimal rotation angles while updating the synthetic dataset to better align with the original feature distribution. Extensive experiments on widely used benchmark datasets demonstrate that the proposed method consistently outperforms existing dataset distillation approaches, achieving higher accuracy and strong cross-architecture generalization.

Engineering design optimization seeks to automatically determine the shapes, topologies, or parameters of components that maximize performance under given conditions. This process often depends on physics-based simulations, which are difficult to install, computationally expensive, and require domain-specific expertise. To mitigate these challenges, we introduce EngiBench, the first open‐source library and datasets spanning diverse domains for data‐driven engineering design. EngiBench provides a unified API and a curated set of benchmarks---covering aeronautics, heat conduction, photonics, and more---that enable fair, reproducible comparisons of optimization and machine learning algorithms, such as generative or surrogate models. We also release EngiOpt, a companion library offering a collection of such algorithms compatible with the EngiBench interface. Both libraries are modular, letting users plug in novel algorithms or problems, automate end-to-end experiment workflows, and leverage built-in utilities for visualization, dataset generation, feasibility checks, and performance analysis. We demonstrate their versatility through experiments comparing state-of-the-art techniques across multiple engineering design problems, an undertaking that was previously prohibitively time-consuming to perform. Finally, we show that these problems pose significant challenges for standard machine learning methods due to highly sensitive and constrained design manifolds.

Accurately tracking camera intrinsics is crucial for achieving 3D understanding from 2D video. However, most 3D algorithms assume that camera intrinsics stay constant throughout a video, which is often not true for many real-world in-the-wild videos. A major obstacle in this field is a lack of dynamic camera intrinsics benchmarks--existing benchmarks typically offer limited diversity in scene content and intrinsics variation, and none provide per-frame intrinsic changes for consecutive video frames. In this paper, we present Intrinsics in Flux (InFlux), a real-world benchmark that provides per-frame ground truth intrinsics annotations for videos with dynamic intrinsics. Compared to prior benchmarks, InFlux captures a wider range of intrinsic variations and scene diversity, featuring 143K+ annotated frames from 386 high-resolution indoor and outdoor videos with dynamic camera intrinsics. To ensure accurate per-frame intrinsics, we build a comprehensive lookup table of calibration experiments and extend the Kalibr toolbox to improve its accuracy and robustness. Using our benchmark, we evaluate existing baseline methods for predicting camera intrinsics and find that most struggle to achieve accurate predictions on videos with dynamic intrinsics. For the dataset, code, videos, and submission, please visit https://influx.cs.princeton.edu/.

The augmentation of algorithms with predictions of the optimal solution, such as from a machine-learning algorithm, has garnered significant attention in recent years, particularly in facility location problems. Moving beyond the traditional focus on utilitarian and egalitarian objectives, we design learning-augmented facility location mechanisms for the envy ratio objective, a fairness metric defined as the maximum ratio between the utilities of any two agents. For the deterministic setting, we propose a mechanism which utilizes predictions to achieve $\alpha$-consistency and $\frac{\alpha}{\alpha - 1}$-robustness for a selected parameter $\alpha \in [1,2]$, and prove its optimality. We also resolve open questions raised by Ding et al. [2020], devising a randomized mechanism without predictions to improve upon the best-known approximation ratio from $2$ to $1.8944$. Building upon these advancements, we construct a novel randomized mechanism which incorporates predictions to achieve improved performance guarantees.

Large Language Models (LLMs) have demonstrated remarkable capabilities but typically require extensive computational resources and memory for inference. Post-training quantization (PTQ) can effectively reduce these demands by storing weights in lower bit-width formats. However, standard uniform quantization often leads to notable performance degradation, particularly in low-bit scenarios. In this work, we introduce a Grouped Lattice Vector Quantization (GLVQ) framework that assigns each group of weights a customized lattice codebook, defined by a learnable generation matrix. To address the non-differentiability of the quantization process, we adopt Babai rounding to approximate nearest-lattice-point search during training, which enables stable optimization of the generation matrices. Once trained, decoding reduces to a simple matrix-vector multiplication, yielding an efficient and practical quantization pipeline. Experiments on multiple benchmarks show that our approach achieves a better trade-off between model size and accuracy compared to existing post-training quantization baselines, highlighting its effectiveness in deploying large models under stringent resource constraints. Our source code is available on GitHub repository: https://github.com/xzhang9308/GLVQ.

The recent boom of large pre-trained models witnesses remarkable success in developing foundation models (FMs) for time series forecasting. Despite impressive performance across diverse downstream forecasting tasks, existing time series FMs possess massive network architectures and require substantial pre-training on large-scale datasets, which significantly hinders their deployment in resource-constrained environments. In response to this growing tension between versatility and affordability, we propose **SEMPO**, a novel lightweight foundation model that requires pretraining on relatively small-scale data, yet exhibits strong general time series forecasting. Concretely, SEMPO comprises two key modules: 1) _energy-aware **S**p**E**ctral decomposition module_, that substantially improves the utilization of pre-training data by modeling not only the high-energy frequency signals but also the low-energy yet informative frequency signals that are ignored in current methods; and 2) _**M**ixture-of-**P**r**O**mpts enabled Transformer_, that learns heterogeneous temporal patterns through small dataset-specific prompts and adaptively routes time series tokens to prompt-based experts for parameter-efficient model adaptation across different datasets and domains. Equipped with these modules, SEMPO significantly reduces both pre-training data scale and model size, while achieving strong generalization. Extensive experiments on two large-scale benchmarks covering 16 datasets demonstrate the superior performance of SEMPO in both zero-shot and few-shot forecasting scenarios compared with state-of-the-art methods. Code and data are available at https://github.com/mala-lab/SEMPO.

The constrained version of the standard online convex optimization (OCO) framework, called COCO is considered, where on every round, a convex cost function and a convex constraint function are revealed to the learner after it chooses the action for that round. The objective is to simultaneously minimize the static regret and cumulative constraint violation (CCV). An algorithm is proposed that guarantees a static regret of $O(\sqrt{T})$ and a CCV of $\min\{{\cal V}, O(\sqrt{T}\log T) \}$, where ${\cal V}$ depends on the distance between the consecutively revealed constraint sets, the shape of constraint sets, dimension of action space and the diameter of the action space. When constraint sets have additional structure, ${\cal V}=O(1)$. Compared to the state of the art results, static regret of $O(\sqrt{T})$ and CCV of $O(\sqrt{T}\log T)$, that were universal, the new result on CCV is instance dependent, which is derived by exploiting the geometric properties of the constraint sets.

Reinforcement learning (RL) has become an indispensable post-training step for unlocking the full potential of Large Language Models (LLMs). Its core motivation is to incentivize the model’s inference trajectory via a reward model, effectively balancing the exploration–exploitation trade-off in scenarios where collecting exhaustive input–output ground-truth pairs is infeasible. This motivation naturally extends to visual generation, where perfect alignment between an image and a textual prompt is inherently ambiguous and often unattainable. However, existing visual generative models are not yet ready for RL due to the following two fundamental drawbacks that undermine the foundations of RL: 1) For diffusion-based models, the actual generation trajectories of sampled images cannot be reliably rewarded, as diffusion inversion is notoriously difficult. 2) For autoregressive (AR) models, we show that the widely used spatial visual tokens do not satisfy the Bellman equation and thus violate the policy improvement theorem of RL. To this end, we propose to use Selftok (Self-consistency Tokenizer), which represents each image as a sequential 1D stream of discrete, autoregressive tokens. Together with language, we train a pure AR vision-language model (VLM) for visual generation. Impressively, without using any text-image training pairs, a simple policy gradient algorithm applied to Selftok tokens significantly boosts visual generation performance, surpassing existing models by a large margin. Implementation details are provided in the Appendix.

In large-scale recommendation systems, multimodal (MM) content is increasingly introduced to enhance the generalization of ID features. The rise of Multimodal Large Language Models (MLLMs) enables the construction of unified user and item representations. However, the semantic distribution gap between MM and ID representations leads to \textit{convergence inconsistency} during joint training: the ID branch converges quickly, while the MM branch requires more epochs, thus limiting overall performance. To address this, we propose a two-stage framework including MM representation learning and joint training optimization. First, we fine-tune the MLLM to generate unified user and item representations, and introduce collaborative signals by post-aligning user ID representations to alleviate semantic differences. Then, we propose an Adaptive Gradient Masking (AGM) training strategy to dynamically regulate parameter updates between ID and MLLM branches. AGM estimates the contribution of each representation with mutual information, and applies non-uniform gradient masking at the sub-network level to balance optimization. We provide theoretical analysis of AGM's effectiveness and further introduce an unbiased variant, AGM*, to enhance training stability. Experiments on offline and online A/B tests validate the effectiveness of our approach in mitigating convergence inconsistency and improving performance.

Recent advances in sign language research have benefited from CNN-based backbones, which are primarily transferred from traditional computer vision tasks (\eg object detection, image recognition). However, these CNN-based backbones usually excel at extracting features like contours and texture, but may struggle with capturing sign-related features. To capture such sign-related features, SignGraph model extracts the cross-region sign features by building the Local Sign Graph (LSG) module and the Temporal Sign Graph (TSG) module. However, we emphasize that although capturing cross-region dependencies can improve sign language performance, it may degrade the representation quality of local regions. To mitigate this, we introduce MixSignGraph, which represents sign sequences as a group of mixed graphs for feature extraction. Specifically, besides the LSG module and TSG module that model the intra-frame and inter-frame cross-regions features, we design a simple yet effective Hierarchical Sign Graph (HSG) module, which enhances local region representations following the extraction of cross-region features, by aggregating the same-region features from different-granularity feature maps of a frame, \ie to boost discriminative local features. In addition, to further improve the performance of gloss-free sign language task, we propose a simple yet counter-intuitive Text-based CTC Pre-training (TCTC) method, which generates pseudo gloss labels from text sequences for model pre-training. Extensive experiments conducted on the current five sign language datasets demonstrate that MixSignGraph surpasses the most current models on multiple sign language tasks across several datasets, without relying on any additional cues. Code and models are available at: \href{https://github.com/gswycf/SignLanguage}{\textcolor{blue}{https://github.com/gswycf/SignLanguage}}.

We study the robust geometric median problem in Euclidean space $\mathbb{R}^d$, with a focus on coreset construction. A coreset is a compact summary of a dataset $P$ of size $n$ that approximates the robust cost for all centers $c$ within a multiplicative error $\varepsilon$. Given an outlier count $m$, we construct a coreset of size $\tilde{O}(\varepsilon^{-2} \cdot \min \\{ \varepsilon^{-2}, d \\})$ when $n \geq 4m$, eliminating the $O(m)$ dependency present in prior work [Huang et al., 2022 & 2023]. For the special case of $d = 1$, we achieve an optimal coreset size of $\tilde{\Theta}(\varepsilon^{-1/2} + \frac{m}{n} \varepsilon^{-1})$, revealing a clear separation from the vanilla case studied in [Huang et al., 2023; Afshani and Chris, 2024]. Our results further extend to robust $(k,z)$-clustering in various metric spaces, eliminating the $m$-dependence under mild data assumptions. The key technical contribution is a novel non-component-wise error analysis, enabling substantial reduction of outlier influence, unlike prior methods that retain them. Empirically, our algorithms consistently outperform existing baselines in terms of size-accuracy tradeoffs and runtime, even when data assumptions are violated across a wide range of datasets.

Recent generations of frontier language models have introduced Large Reasoning Models (LRMs) that generate detailed thinking processes before providing answers. While these models demonstrate improved performance on reasoning benchmarks, their fundamental capabilities, scaling properties, and limitations remain insufficiently understood. Current evaluations primarily focus on established mathematical and coding benchmarks, emphasizing final answer accuracy. However, this evaluation paradigm often suffers from data contamination and does not provide insights into the reasoning traces' structure and quality. In this work, we systematically investigate these gaps with the help of controllable puzzle environments that allow precise manipulation of compositional complexity while maintaining consistent logical structures. This setup enables the analysis of not only final answers but also the internal reasoning traces, offering insights into how LRMs ``think''. Through extensive experimentation across diverse puzzles, we show that frontier LRMs face a complete accuracy collapse beyond certain complexities. Moreover, they exhibit a counterintuitive scaling limit: their reasoning effort increases with problem complexity up to a point, then declines despite having an adequate token budget. By comparing LRMs with their standard LLM counterparts under equivalent inference compute, we identify three performance regimes: (1) low-complexity tasks where standard models surprisingly outperform LRMs, (2) medium-complexity tasks where additional thinking in LRMs demonstrates advantage, and (3) high-complexity tasks where both models experience complete collapse. We found that LRMs have limitations in exact computation: they fail to use explicit algorithms and reason inconsistently across scales and problems. We also investigate the reasoning traces in more depth, studying the patterns of explored solutions and analyzing the models' computational behavior, shedding light on their strengths, limitations, and ultimately raising questions about the nature for their reasoning capabilities.

The growing body of work on Physics-Informed Neural Networks (PINNs) seeks to use machine learning strategies to improve methods for solution discovery of Partial Differential Equations (PDEs). While classical solvers may remain the preferred tool of choice in the short-term, PINNs can be viewed as complementary. The expectation is that in some specific use cases, they can be effective, standalone. A key step in training PINNs is selecting domain points for loss evaluation, where Monte Carlo sampling remains the dominant but often suboptimal in low dimension settings, common in physics. We leverage recent advances in asymptotic expansions of quadrature nodes and weights (for weight functions belonging to the modified Gauss-Jacobi family) together with suitable adjustments for parameterization towards a data-driven framework for learnable quadrature rules. A direct benefit is a performance improvement of PINNs, relative to existing alternatives, on a wide range of problems studied in the literature. Beyond finding a standard solution for an instance of a single PDE, our construction enables learning rules to predict solutions for a given family of PDEs via hyper-networks, a useful capability for PINNs.

This paper investigates the role of explicit thinking process in rule-based reinforcement fine-tuning (RFT) for multi-modal large language models (MLLMs). We first extend \textit{Thinking-RFT} to image classification task, using verifiable rewards for fine-tuning~(FT). Experiments show {Thinking-RFT} significantly outperforms supervised FT and yields a cross-dataset generalization effect. We then rethink and question whether explicit thinking in RFT is always necessary and beneficial. Challenging the convention that explicit thinking is crucial for the success of RFT, we introduce \textit{No-Thinking-RFT}, exploring RFT without thinking by introducing a simple equality accuracy reward. We evaluate No-Thinking-RFT on six diverse tasks across different model sizes and types. Experiment results reveal four key findings: \textbf{(1).} Visual perception tasks do not require thinking during RFT, as No-Thinking-RFT consistently outperforms or matches Thinking-RFT across model sizes and types. \textbf{(2).} Models with limited capabilities struggle to generate high-quality CoT for RFT, making Thinking-RFT less effective than No-Thinking-RFT. \textbf{(3).} There are inconsistencies between the answers in the thinking tags and answer tags for some responses of Thinking-RFT, which show lower average accuracy than the overall accuracy. \textbf{(4).} The performance gain of No-Thinking-RFT mainly stems from improved learning during no thinking FT and the avoidance of inference overthinking, as evidenced by the partial gains from appending empty thinking tags at inference time of Thinking-RFT. We hypothesize that explicit thinking before verifiable answers may hinder reward convergence and reduce performance in certain scenarios. To test this, we propose \textit{Think-After-Answer}, which places thinking after the answer to mitigate this effect for experimental verification. Lastly, we conduct a pilot study to explore whether MLLMs can learn when to think during RFT, introducing an \textit{Adaptive-Thinking} method. Experiments show that model converges to either thinking or not depending on model capability, achieving comparable or better performance than both Thinking and No-Thinking-RFT. Our findings suggest MLLMs can adaptively decide to think or not based on their capabilities and task complexity, offering insights into the thinking process in RFT.

We propose Metric Automata Theory, an elegant generalisation of classic Automata Theory to continuous dynamical systems, that constitutes a unifying theory of all kinds of Recurrent Neural Networks (RNNs), including widely-adopted architectures such as xLSTM and State Space Models (SSMs). The theory allows one to analyse RNNs both in the finite and unbounded precision settings seamlessly, while utilising fundamental results of Automata Theory. It also provides a novel notion of robustness that guarantees numerical stability and contributes to stability of learning. We employ the theory to prove a comprehensive set of expressivity results for widely-adopted RNNs, with a focus on robustness and finite-precision. Notably, we contrast the capabilities of xLSTM and SSMs for robustly modelling all star-free regular languages—xLSTM can do so, while SSMs cannot robustly recognize the FLIP-FLOP language. Thus we give a novel perspective on the importance of non-linear recurrences, giving insight for why xLSTM shows superior performance to SSMs on several tasks. We provide an improved understanding of the capabilities of Mamba, a popular SSM model. We show that Mamba is not generally capable of recognising the star-free languages under finite-precision, which is seemingly in contrast with the existing theoretical and empirical results for SSMs. We clarify the picture, by showing that Mamba admits a piecewise-linearly separable state space that allows it to approximate star-free languages, with some length-generalisation abilities. At the same time, Mamba does not admit such state spaces for languages like Parity. This explains why empirically Mamba performs well on star-free languages, and fails on Parity.

Real-world applications are stretching context windows to hundreds of thousand of tokens while Large Language Models (LLMs) swell from billions to trillions of parameters. This dual expansion send compute and memory costs skyrocketing, making $\textit{token compression}$ indispensable. We introduce Vision Centric Token Compression ($\textbf{Vist}$), a $\textit{slow–fast}$ compression framework that mirrors human reading: the $\textit{fast}$ path renders distant tokens into images, letting a $\textbf{frozen, lightweight vision encoder}$ skim the low-salience context; the $\textit{slow}$ path feeds the proximal window into the LLM for fine-grained reasoning. A Probability-Informed Visual Enhancement (PVE) objective masks high-frequency tokens during training, steering the Resampler to concentrate on semantically rich regions—just as skilled reader gloss over function words. On eleven in-context learning benchmarks, $\textbf{Vist}$ achieves the same accuracy with 2.3$\times$ fewer tokens, cutting FLOPs by 16\% and memory by 50\%. This method delivers remarkable results, outperforming the strongest text encoder-based compression method CEPE by $\textbf{7.6}$\% on average over benchmarks like TriviaQA, NQ, PopQA, NLUI, and CLIN, setting a new standard for token efficiency in LLMs. The source code will be released.

4D Gaussian Splatting (4DGS) has recently gained considerable attention as a method for reconstructing dynamic scenes. Despite achieving superior quality, 4DGS typically requires substantial storage and suffers from slow rendering speed. In this work, we delve into these issues and identify two key sources of temporal redundancy. (Q1) \textbf{Short-Lifespan Gaussians}: 4DGS uses a large portion of Gaussians with short temporal span to represent scene dynamics, leading to an excessive number of Gaussians. (Q2) \textbf{Inactive Gaussians}: When rendering, only a small subset of Gaussians contributes to each frame. Despite this, all Gaussians are processed during rasterization, resulting in redundant computation overhead. To address these redundancies, we present \textbf{4DGS-1K}, which runs at over 1000 FPS on modern GPUs. For Q1, we introduce the Spatial-Temporal Variation Score, a new pruning criterion that effectively removes short-lifespan Gaussians while encouraging 4DGS to capture scene dynamics using Gaussians with longer temporal spans. For Q2, we store a mask for active Gaussians across consecutive frames, significantly reducing redundant computations. Compared to vanilla 4DGS, our method achieves a $41\times$ reduction in storage and $9\times$ faster rasterization on complex dynamic scenes, while maintaining comparable visual quality.

Language models (LMs) and their extension, vision-language models (VLMs), have achieved remarkable performance across various tasks. However, they still struggle with complex reasoning tasks that require multimodal or multilingual real world knowledge. To support such capabilities, an external memory system that can efficiently provide relevant multimodal information is essential. Existing approaches generally concatenate image and text tokens into a long sequence as memory, which, however, may drastically increase context length and even degrade performance. In contrast, we propose using continuous memory-a compact set of dense embeddings-to more effectively and efficiently represent multimodal and multilingual knowledge. Our key insight is that a VLM can serve as its own continuous memory encoder. We empirically show that this design improves performance on complex multimodal reasoning tasks. Building on this, we introduce a data-efficient and parameter-efficient method to fine-tune the VLM into a memory encoder, requiring only 1.2\% of the model’s parameters and a small corpus of 15.6K self-synthesized samples. Our approach CoMEM utilizes VLM's original capabilities to encode arbitrary multimodal and multilingual knowledge into just 8 continuous embeddings. Since the inference-time VLM remains frozen, our memory module is plug-and-play and can be flexibly integrated as needed. Extensive experiments across eight multimodal reasoning benchmarks demonstrate the effectiveness of our approach. Code and data is publicly released here https://github.com/WenyiWU0111/CoMEM.

Spiking Neural Networks (SNNs) offer a biologically plausible and energy-efficient framework for temporal information processing. However, existing studies overlook a fundamental property widely observed in biological neurons—synaptic heterogeneity, which plays a crucial role in temporal processing and cognitive capabilities. To bridge this gap, we introduce HetSyn, a generalized framework that models synaptic heterogeneity with synapse-specific time constants. This design shifts temporal integration from the membrane potential to the synaptic current, enabling versatile timescale integration and allowing the model to capture diverse synaptic dynamics. We implement HetSyn as HetSynLIF, an extended form of the leaky integrate-and-fire (LIF) model equipped with synapse-specific decay dynamics. By adjusting the parameter configuration, HetSynLIF can be specialized into vanilla LIF neurons, neurons with threshold adaptation, and neuron-level heterogeneous models. We demonstrate that HetSynLIF not only improves the performance of SNNs across a variety of tasks—including pattern generation, delayed match-to-sample, speech recognition, and visual recognition—but also exhibits strong robustness to noise, enhanced working memory performance, efficiency under limited neuron resources, and generalization across timescales. In addition, analysis of the learned synaptic time constants reveals trends consistent with empirical observations in biological synapses. These findings underscore the significance of synaptic heterogeneity in enabling efficient neural computation, offering new insights into brain-inspired temporal modeling. Code available at: https://github.com/dzcgood/HetSyn.