NeurIPS 2025 论文列表

Large language models are increasingly customized through fine-tuning and other adaptations, creating challenges in enforcing licensing terms and managing downstream impacts such as protecting intellectual property or identifying vulnerabilities. We address this challenge by developing a framework for testing model provenance. Our approach is based on the key observation that real-world model derivations preserve significant similarities in model outputs that can be detected through statistical analysis. Using only black-box access to models, we employ multiple hypothesis testing to compare model similarities against a baseline established by unrelated models. On two comprehensive real-world benchmarks spanning models from 30M to 4B parameters and comprising over 600 models, our tester achieves 90-95% precision and 80-90% recall in identifying derived models. These results demonstrate the viability of systematic provenance verification in production environments even when only API access is available.

Spike sorting is an essential process in neural recording, which identifies and separates electrical signals from individual neurons recorded by electrodes in the brain, enabling researchers to study how specific neurons communicate and process information. Although there exist a number of spike sorting methods which have contributed to significant neuroscientific breakthroughs, many are heuristically designed, making it challenging to verify their correctness due to the difficulty of obtaining ground truth labels from real-world neural recordings. In this work, we explore a data-driven, deep learning-based approach. We begin by creating a large-scale dataset through electrophysiology simulations using biologically realistic computational models. We then present SimSort, a pretraining framework for spike sorting. Trained solely on simulated data, SimSort demonstrates zero-shot generalizability to real-world spike sorting tasks, yielding consistent improvements over existing methods across multiple benchmarks. These results highlight the potential of simulation-driven pretraining to enhance the robustness and scalability of spike sorting in experimental neuroscience.

Multimodal models have demonstrated powerful capabilities in complex tasks requiring multimodal alignment, including zero-shot classification and cross-modal retrieval. However, existing models typically rely on millions of paired multimodal samples, which are prohibitively expensive or infeasible to obtain in many domains. In this work, we explore the feasibility of building multimodal models with limited amount of paired data by aligning pretrained unimodal foundation models. We show that high-quality alignment is possible with as few as tens of thousands of paired samples$\unicode{x2013}$less than 1\% of the data typically used in the field. To achieve this, we introduce STRUCTURE, an effective regularization technique that preserves the neighborhood geometry of the latent space of unimodal encoders. Additionally, we show that aligning last layers is often suboptimal and demonstrate the benefits of aligning the layers with the highest representational similarity across modalities. These two components can be readily incorporated into existing alignment methods, yielding substantial gains across 24 zero-shot image classification and retrieval benchmarks, with average relative improvement of 51.6\% in classification and 91.8\% in retrieval tasks. Our results highlight the effectiveness and broad applicability of our framework for limited-sample multimodal learning and offer a promising path forward for resource-constrained domains.

Measuring similarity between training examples is critical for curating high-quality and diverse pretraining datasets for language models. However, similarity is typically computed with a generic off-the-shelf embedding model that has been trained for tasks such as retrieval. Whether these embedding-based similarity metrics are well-suited for pretraining data selection remains largely unexplored. In this paper, we propose a new framework to assess the suitability of a similarity metric specifically for data curation in language model pretraining applications. Our framework's first evaluation criterion captures how well distances reflect generalization in pretraining loss between different training examples. Next, we use each embedding model to guide a standard diversity-based data curation algorithm and measure its utility by pretraining a language model on the selected data and evaluating downstream task performance. Finally, we evaluate the capabilities of embeddings to distinguish between examples from different data sources. With these evaluations, we demonstrate that standard off-the-shelf embedding models are not well-suited for the pretraining data curation setting, underperforming even remarkably simple embeddings that are extracted from models trained on the same pretraining corpus. Our experiments are performed on the Pile, for pretraining a 1.7B parameter language model on 200B tokens. We believe our analysis and evaluation framework serves as a foundation for the future design of embeddings that specifically reason about similarity in pretraining datasets.

The performance of learning-based object detection algorithms, which attempt to both classify and locate objects within images, is determined largely by the quality of the annotated dataset used for training. Two types of labelling noises are prevalent: objects that are incorrectly classified (categorization noise) and inaccurate bounding boxes (localization noise); both noises typically occur together in large-scale datasets. In this paper we propose a distillation-based method to train object detectors that takes into account both categorization and localization noise. The key insight underpinning our method is that the early-learning phenomenon - in which models trained on noisy data with mixed clean and false labels tend to first fit to the clean data, and memorize the false labels later -- manifests earlier for localization noise than for categorization noise. We propose a method that uses models from the early-learning phase (before overfitting to noisy data occurs) as a teacher network. A plug-in module implementation compatible with general object detection architectures is developed, and its performance is validated against the state-of-the-art using PASCAL VOC, MS COCO and VinDr-CXR medical detection datasets.

We study bilevel optimization problems where the lower-level problems are strongly convex and have coupled linear constraints. To overcome the potential non-smoothness of the hyper-objective and the computational challenges associated with the Hessian matrix, we utilize penalty and augmented Lagrangian methods to reformulate the original problem as a single-level one. Especially, we establish a strong theoretical connection between the reformulated function and the original hyper-objective by characterizing the closeness of their values and derivatives. Based on this reformulation, we propose a single-loop, first-order algorithm for linearly constrained bilevel optimization (SFLCB). We provide rigorous analyses of its non-asymptotic convergence rates, showing an improvement over prior double-loop algorithms -- form $O(\epsilon^{-3}\log(\epsilon^{-1}))$ to $O(\epsilon^{-3})$. The experiments corroborate our theoretical findings and demonstrate the practical efficiency of the proposed SFLCB algorithm. Simulation code is provided at https://github.com/ShenGroup/SFLCB.

Transformers have demonstrated exceptional performance across a wide range of domains. While their ability to perform reinforcement learning in-context has been established both theoretically and empirically, their behavior in non-stationary environments remains less understood. In this study, we address this gap by showing that transformers can achieve nearly optimal dynamic regret bounds in non-stationary settings. We prove that transformers are capable of approximating strategies used to handle non-stationary environment, and can learn the approximator in the in-context learning setup. Our experiments further show that transformers can match or even outperform existing expert algorithms in such environments.

Diverse decoding of large language models is crucial for applications requiring multiple semantically distinct responses, yet existing methods primarily achieve lexical rather than semantic diversity. This limitation significantly constrains Best-of-N strategies, group-based reinforcement learning, and data synthesis. While temperature sampling and diverse beam search modify token distributions or apply n-gram penalties, they fail to ensure meaningful semantic differentiation. We introduce Semantic-guided Diverse Decoding (SemDiD), operating directly in embedding space that balances quality with diversity through three complementary mechanisms: orthogonal directional guidance, dynamic inter-group repulsion, and position-debiased probability assessment. SemDiD harmonizes these competing objectives using adaptive gain functions and constraint optimization, ensuring both quality thresholds and maximal semantic differentiation. Experiments show SemDiD consistently outperforms existing methods, improving Best-of-N coverage by 1.4-5.2% across diverse tasks and accelerating RLHF training convergence by 15% while increasing accuracy by up to 2.1%.

Uncertainty quantification (UQ) in scientific machine learning is increasingly critical as neural networks are widely adopted to tackle complex problems across diverse scientific disciplines. For physics-informed neural networks (PINNs), a prominent model in scientific machine learning, uncertainty is typically quantified using Bayesian or dropout methods. However, both approaches suffer from a fundamental limitation: the prior distribution or dropout rate required to construct honest confidence sets cannot be determined without additional information. In this paper, we propose a novel method within the framework of extended fiducial inference (EFI) to provide rigorous uncertainty quantification for PINNs. The proposed method leverages a narrow-neck hyper-network to learn the parameters of the PINN and quantify their uncertainty based on imputed random errors in the observations. This approach overcomes the limitations of Bayesian and dropout methods, enabling the construction of honest confidence sets based solely on observed data. This advancement represents a significant breakthrough for PINNs, greatly enhancing their reliability, interpretability, and applicability to real-world scientific and engineering challenges. Moreover, it establishes a new theoretical framework for EFI, extending its application to large-scale models, eliminating the need for sparse hyper-networks, and significantly improving the automaticity and robustness of statistical inference.

Large reconstruction model has remarkable progress, which can directly predict 3D or 4D representations for unseen scenes and objects. However, current work has not systematically explored the potential of large reconstruction models in the field of autonomous driving. To achieve this, we introduce the Large 4D Gaussian Reconstruction Model (DrivingRecon). With an elaborate and simple framework design, it not only ensures efficient and high-quality reconstruction, but also provides potential for downstream tasks. There are two core contributions: firstly, the Prune and Dilate Block (PD-Block) is proposed to prune redundant and overlapping Gaussian points and dilate Gaussian points for complex objects. Then, dynamic and static decoupling is tailored to better learn the temporary-consistent geometry across different time. Experimental results demonstrate that DrivingRecon significantly improves scene reconstruction quality compared to existing methods. Furthermore, we explore applications of DrivingRecon in model pre-training, vehicle type adaptation, and scene editing. Our code will be available.

Brain-computer interfaces have shown great potential in motor and speech rehabilitation, but still suffer from low performance stability across days, mostly due to the instabilities in neural signals. These instabilities, partially caused by neuron deaths and electrode shifts, leading to channel-level variabilities among different recording days. Previous studies mostly focused on aligning multi-day neural signals of onto a low-dimensional latent manifold to reduce the variabilities, while faced with difficulties when neural signals exhibit significant drift. Here, we propose to learn a channel-level invariant neural representation to address the variabilities in channels across days. It contains a channel-rearrangement module to learn stable representations against electrode shifts, and a channel reconstruction module to handle the missing neurons. The proposed method achieved the state-of-the-art performance with cross-day decoding tasks over two months, on multiple benchmark BCI datasets. The proposed approach showed good generalization ability that can be incorporated to different neural networks.

The energy paradigm, exemplified by Hopfield networks, offers a principled framework for memory in neural systems by interpreting dynamics as descent on an energy surface. While powerful for static associative memories, it falls short in modeling sequential memory, where transitions between memories are essential. We introduce the Exponential Dynamic Energy Network (EDEN), a novel architecture that extends the energy paradigm to temporal domains by evolving the energy function over multiple timescales. EDEN combines a static high-capacity energy network with a slow, asymmetrically interacting modulatory population, enabling robust and controlled memory transitions. We formally derive short-timescale energy functions that govern local dynamics and use them to analytically compute memory escape times, revealing a phase transition between static and dynamic regimes. The analysis of capacity, defined as the number of memories that can be stored with minimal error rate as a function of the dimensions of the state space (number of feature neurons), for EDEN shows that it achieves exponential sequence memory capacity $\mathcal{O}(\gamma^N)$, outperforming the linear capacity $\mathcal{O}(N)$ of conventional models. Furthermore, EDEN's dynamics resemble the activity of time and ramping cells observed in the human brain during episodic memory tasks, grounding its biological relevance. By unifying static and sequential memory within a dynamic energy framework, EDEN offers a scalable and interpretable model for high-capacity temporal memory in both artificial and biological systems.

Existing GFlowNet-based methods for vehicle routing problems (VRPs) typically employ Trajectory Balance (TB) to achieve global optimization but often neglect important aspects of local optimization. While Detailed Balance (DB) addresses local optimization more effectively, it alone falls short in solving VRPs, which inherently require holistic trajectory optimization. To address these limitations, we introduce the Hybrid-Balance GFlowNet (HBG) framework, which uniquely integrates TB and DB in a principled and adaptive manner by aligning their intrinsically complementary strengths. Additionally, we propose a specialized inference strategy for depot-centric scenarios like the Capacitated Vehicle Routing Problem (CVRP), leveraging the depot node's greater flexibility in selecting successors. Despite this specialization, HBG maintains broad applicability, extending effectively to problems without explicit depots, such as the Traveling Salesman Problem (TSP). We evaluate HBG by integrating it into two established GFlowNet-based solvers, i.e., AGFN and GFACS, and demonstrate consistent and significant improvements across both CVRP and TSP, underscoring the enhanced solution quality and generalization afforded by our approach.

Gradients of neural networks encode valuable information for optimization, editing, and analysis of models. Therefore, practitioners often treat gradients as inputs to task-specific algorithms, e.g., using gradient statistics for pruning or optimization. Recent works explore *learning* algorithms that operate directly on gradients but use architectures that are not specifically designed for gradient processing, hindering their applicability. In this paper, we present a principled approach for designing architectures that process gradients. Our approach is guided by three principles: (1) equivariant design that preserves neuron permutation symmetries, (2) processing sets of gradients across multiple data points to capture curvature information, and (3) efficient gradient representation through rank-1 decomposition. Based on these principles, we introduce GradMetaNet, a novel architecture for learning on gradients, constructed from simple equivariant blocks. We prove universality results for GradMetaNet, and show that previous approaches cannot approximate natural gradient-based functions that GradMetaNet can. We then demonstrate GradMetaNet's effectiveness on a diverse set of gradient-based tasks for *MLPs* and *transformers*, such as learned optimization, INR editing, and loss landscape curvature estimation.

With the development of large language models (LLMs), efficient inference through Key-Value (KV) cache compression has attracted considerable attention, especially for long-context generation. To compress the KV cache, recent methods identify critical KV tokens through static modeling of attention scores. However, these methods often struggle to accurately determine critical tokens as they neglect the *temporal patterns* in attention scores, resulting in a noticeable degradation in LLM performance. To address this challenge, we propose **AttentionPredictor**, which is the **first learning-based method to directly predict attention patterns for KV cache compression and critical token identification**. Specifically, AttentionPredictor learns a lightweight, unified convolution model to dynamically capture spatiotemporal patterns and predict the next-token attention scores. An appealing feature of AttentionPredictor is that it accurately predicts the attention score and shares the unified prediction model, which consumes negligible memory, among all transformer layers. Moreover, we propose a cross-token critical cache prefetching framework that hides the token estimation time overhead to accelerate the decoding stage. By retaining most of the attention information, AttentionPredictor achieves **13$\times$** KV cache compression and **5.6$\times$** speedup in a cache offloading scenario with comparable LLM performance, significantly outperforming the state-of-the-arts. The code is available at https://github.com/MIRALab-USTC/LLM-AttentionPredictor.

Direct Preference Optimization (DPO) has emerged as a simple and effective approach for aligning large language models (LLMs) with human preferences, bypassing the need for a learned reward model. Despite its growing adoption, a fundamental question remains open: what characteristics of preference data are most critical for DPO performance? In this work, we provide a systematic study of how preference data distribution influences DPO, from both theoretical and empirical perspectives. We show that the quality of chosen responses plays a dominant role in optimizing the DPO objective, while the quality of rejected responses may have relatively limited impact. Our theoretical analysis characterizes the optimal response distribution under DPO and reveals how contrastiveness between responses helps primarily by improving the chosen samples. We further study an online DPO setting and show it effectively reduces to supervised fine-tuning on the chosen responses. Extensive experiments across diverse tasks confirm our findings: improving the quality of chosen responses consistently boosts performance regardless of the quality of the rejected responses. We also investigate the benefit of mixing the on-policy data. Our results interpret the mechanism behind some widely adopted strategies and offer practical insights for constructing high-impact preference datasets for LLM alignment.

Diffusion Language Models (DLMs) have been seen as a promising competitor for autoregressive language models (ARs). However, diffusion language models have long been constrained by slow inference. A core challenge is that their non‑autoregressive architecture and bidirectional attention preclude the key–value cache that accelerates decoding. We address this bottleneck by proposing a KV-cache-like mechanism, **d**elayed **KV-Cache**, for the denoising process of DLMs. Our approach is motivated by the observation that different tokens have distinct representation dynamics throughout the diffusion process. Accordingly, we propose a delayed and conditioned caching strategy for key and value states. We design two complementary variants to cache key and value step‑by‑step: (1) dKV-Cache-Decode, which provides almost lossless acceleration, and even improves performance on long sequences, suggesting that existing DLMs may under‑utilise contextual information during inference. (2) dKV-Cache‑Greedy, which has aggressive caching with reduced lifespan, achieving higher speed-ups with quadratic time complexity at the cost of some performance degradation. dKV-Cache, in final, achieves from 2-10$\times$ speedup in inference, largely narrowing the gap between ARs and DLMs. We evaluate our dKV-Cache on several benchmarks, delivering acceleration across general language understanding, mathematical, and code‑generation benchmarks. Experiments demonstrate that cache can also be used in DLMs, even in a training-free manner from current DLMs.

Diffusion-based generative models are emerging as powerful tools for long-horizon planning in reinforcement learning (RL), particularly with offline datasets. However, their performance is fundamentally limited by the quality and diversity of training data. This often restricts their generalization to tasks outside their training distribution or longer planning horizons. To overcome this challenge, we propose *State-Covering Trajectory Stitching* (SCoTS), a novel reward-free trajectory augmentation method that incrementally stitches together short trajectory segments, systematically generating diverse and extended trajectories. SCoTS first learns a temporal distance-preserving latent representation that captures the underlying temporal structure of the environment, then iteratively stitches trajectory segments guided by directional exploration and novelty to effectively cover and expand this latent space. We demonstrate that SCoTS significantly improves the performance and generalization capabilities of diffusion planners on offline goal-conditioned benchmarks requiring stitching and long-horizon reasoning. Furthermore, augmented trajectories generated by SCoTS significantly improve the performance of widely used offline goal-conditioned RL algorithms across diverse environments.

Diffusion models operating in discrete state spaces have emerged as powerful approaches, demonstrating remarkable efficacy across diverse domains, including reasoning tasks and molecular design. Despite their promising applications, the theoretical foundations of these models remain substantially underdeveloped, with the existing literature predominantly focusing on continuous-state diffusion models. A critical gap persists in the theoretical understanding of discrete diffusion modeling: the absence of a rigorous framework for quantifying estimation error with finite data. Consequently, the fundamental question of how precisely one can reconstruct the true underlying distribution from a limited training set remains unresolved. In this work, we analyze the estimation error induced by a score estimation of the discrete diffusion models. One of the main difficulties in the analysis stems from the fact that the cardinality of the state space can be exponentially large with respect to its dimension, which results in an intractable error bound by a naive approach. To overcome this difficulty, we make use of a property that the state space can be smoothly embedded in a continuous Euclidean space that enables us to derive a cardinality independent bound, which is more practical in real applications. In particular, we consider a setting where the state space is structured as a hypercube graph, and another where the induced graph Laplacian can be asymptotically well approximated by the ordinary Laplacian defined on the continuous space, and then derive state space size independent bounds.