NeurIPS 2025 论文列表

Low-Rank Adaptation (LoRA) offers a parameter-efficient paradigm for tuning large models. While recent spectral initialization methods improve convergence and performance over the naive “Noise \& Zeros” scheme, their extra computational and storage overhead undermines efficiency. In this paper, we establish update magnitude as the fundamental driver of LoRA performance and propose LoRAM, a magnitude-driven “Basis \& Basis” initialization scheme that matches spectral methods without their inefficiencies. Our key contributions are threefold: (i) Magnitude of weight updates determines convergence. We prove low-rank structures intrinsically bound update magnitudes, unifying hyperparameter tuning in learning rate, scaling factor, and initialization as mechanisms to optimize magnitude regulation. (ii) Spectral initialization succeeds via magnitude amplification. We demystify that the presumed knowledge-driven benefit of spectral component essentially arises from the boost in the weight update magnitude. (iii) A novel and compact initialization strategy, LoRAM, scales deterministic orthogonal bases using pretrained weight magnitudes to simulate spectral gains. Extensive experiments show that LoRAM serves as a strong baseline, retaining the full efficiency of LoRA while matching or outperforming spectral initialization across benchmarks.

We present a universal theoretical framework for understanding *long-context language modeling* based on a *bipartite* mutual information scaling law that we rigorously verify in natural language. We demonstrate that bipartite mutual information captures multi-token interactions distinct from and scaling independently of conventional two-point mutual information, and show that this provides a more complete characterization of the dependencies needed for accurately modeling long sequences. Leveraging this scaling law, we formulate the **L**ong-context **L**anguage **M**odeling (**L**$^2$**M**) condition, which lower bounds the necessary scaling of a model's history state—the latent variables responsible for storing past information—for effective long-context modeling. We validate the framework and its predictions on transformer and state-space models. Our work provides a principled foundation to understand long-context modeling and to design more efficient architectures with stronger long-context capabilities, with potential applications beyond natural language.

We study whether and how the choice of optimization algorithm can impact group fairness in deep neural networks. Through stochastic differential equation analysis of optimization dynamics in an analytically tractable setup, we demonstrate that the choice of optimization algorithm indeed influences fairness outcomes, particularly under severe imbalance. Furthermore, we show that when comparing two categories of optimizers, adaptive methods and stochastic methods, RMSProp (from the adaptive category) has a higher likelihood of converging to fairer minima than SGD (from the stochastic category). Building on this insight, we derive two new theoretical guarantees showing that, under appropriate conditions, RMSProp exhibits fairer parameter updates and improved fairness in a single optimization step compared to SGD. We then validate these findings through extensive experiments on three publicly available datasets, namely CelebA, FairFace, and MS-COCO, across different tasks as facial expression recognition, gender classification, and multi-label classification, using various backbones. Considering multiple fairness definitions including equalized odds, equal opportunity, and demographic parity, adaptive optimizers like RMSProp and Adam consistently outperform SGD in terms of group fairness, while maintaining comparable predictive accuracy. Our results highlight the role of adaptive updates as a crucial yet overlooked mechanism for promoting fair outcomes. We release the source code at: https://github.com/Mkolahdoozi/Some-Optimizers-Are-More-Equal.

When simulating partial differential equations, hybrid solvers combine coarse numerical solvers with learned correctors. They promise accelerated simulations while adhering to physical constraints. However, as shown in our theoretical framework, directly applying learned corrections to solver outputs leads to significant autoregressive errors, which originate from amplified perturbations that accumulate during long-term rollouts, especially in chaotic regimes. To overcome this, we propose the Indirect Neural Corrector ($\mathrm{INC}$), which integrates learned corrections into the governing equations rather than applying direct state updates. Our key insight is that $\mathrm{INC}$ reduces the error amplification on the order of $\Delta t^{-1} + L$, where $\Delta t$ is the timestep and $L$ the Lipschitz constant. At the same time, our framework poses no architectural requirements and integrates seamlessly with arbitrary neural networks and solvers. We test $\mathrm{INC}$ in extensive benchmarks, covering numerous differentiable solvers, neural backbones, and test cases ranging from a 1D chaotic system to 3D turbulence. INC improves the long-term trajectory performance ($R^2$) by up to 158.7\%, stabilizes blowups under aggressive coarsening, and for complex 3D turbulence cases yields speed-ups of several orders of magnitude. INC thus enables stable, efficient PDE emulation with formal error reduction, paving the way for faster scientific and engineering simulations with reliable physics guarantees.

Concept bottleneck models (CBMs) ensure interpretability by decomposing predictions into human interpretable concepts. Yet the annotations used for training CBMs that enable this transparency are often noisy, and the impact of such corruption is not well understood. In this study, we present the first systematic study of noise in CBMs and show that even moderate corruption simultaneously impairs prediction performance, interpretability, and the intervention effectiveness. Our analysis identifies a susceptible subset of concepts whose accuracy declines far more than the average gap between noisy and clean supervision and whose corruption accounts for most performance loss. To mitigate this vulnerability we propose a two-stage framework. During training, sharpness-aware minimization stabilizes the learning of noise-sensitive concepts. During inference, where clean labels are unavailable, we rank concepts by predictive entropy and correct only the most uncertain ones, using uncertainty as a proxy for susceptibility. Theoretical analysis and extensive ablations elucidate why sharpness-aware training confers robustness and why uncertainty reliably identifies susceptible concepts, providing a principled basis that preserves both interpretability and resilience in the presence of noise.

Pre-trained stable diffusion models (SD) have shown great advances in visual correspondence. In this paper, we investigate the capabilities of Diffusion Transformers (DiTs) for accurate dense correspondence. Distinct from SD, DiTs exhibit a critical phenomenon in which very few feature activations exhibit significantly larger values than others, known as massive activations, leading to uninformative representations and significant performance degradation for DiTs. The massive activations consistently concentrate at very few fixed dimensions across all image patch tokens, holding little local information. We analyze these dimension-concentrated massive activations and uncover that their concentration is inherently linked to the Adaptive Layer Normalization (AdaLN) in DiTs. Building on these findings, we propose the Diffusion Transformer Feature (DiTF), a training-free AdaLN-based framework that extracts semantically discriminative features from DiTs. Specifically, DiTF leverages AdaLN to adaptively localize and normalize massive activations through channel-wise modulation. Furthermore, a channel discard strategy is introduced to mitigate the adverse effects of massive activations. Experimental results demonstrate that our DiTF outperforms both DINO and SD-based models and establishes a new state-of-the-art performance for DiTs in different visual correspondence tasks (e.g., with +9.4\% on Spair-71k and +4.4\% on AP-10K-C.S.).

Humans excel at constructing panoramic mental models of their surroundings, maintaining object permanence and inferring scene structure beyond visible regions. In contrast, current artificial vision systems struggle with persistent, panoramic understanding, often processing scenes egocentrically on a frame-by-frame basis. This limitation is pronounced in the Track Any Point (TAP) task, where existing methods fail to track 2D points outside the field of view. To address this, we introduce TAP-Vid 360, a novel task that requires predicting the 3D direction to queried scene points across a video sequence, even when far outside the narrow field of view of the observed video. This task fosters learning allocentric scene representations without needing dynamic 4D ground truth scene models for training. Instead, we exploit 360 videos as a source of supervision, resampling them into narrow field-of-view perspectives while computing ground truth directions by tracking points across the full panorama using a 2D pipeline. We introduce a new dataset and benchmark, TAP360-10k comprising 10k perspective videos with ground truth directional point tracking. Our baseline adapts CoTracker v3 to predict per-point rotations for direction updates, outperforming existing TAP and TAP-Vid 3D methods.

Offline reinforcement learning (RL) aims to learn optimal policies from static datasets while enhancing generalization to out-of-distribution (OOD) data. To mitigate overfitting to suboptimal behaviors in offline datasets, existing methods often relax constraints on policy and data or extract informative patterns through data-driven techniques. However, there has been limited exploration into structurally guiding the optimization process toward flatter regions of the solution space that offer better generalization. Motivated by this observation, we present \textit{FANS}, a generalization-oriented structured network framework that promotes flatter and robust policy learning by guiding the optimization trajectory through modular architectural design. FANS comprises four key components: (1) Residual Blocks, which facilitate compact and expressive representations; (2) Gaussian Activation, which promotes smoother gradients; (3) Layer Normalization, which mitigates overfitting; and (4) Ensemble Modeling, which reduces estimation variance. By integrating FANS into a standard actor-critic framework, we highlight that this remarkably simple architecture achieves superior performance across various tasks compared to many existing advanced methods. Moreover, we validate the effectiveness of FANS in mitigating overestimation and promoting generalization, demonstrating the promising potential of architectural design in advancing offline RL.

Reducing the key-value (KV) cache burden in Large Language Models (LLMs) significantly accelerates inference. Dynamically selecting critical KV caches during decoding helps maintain performance. Existing methods use random linear hashing to identify important tokens, but this approach is inefficient due to the orthogonal distribution of queries and keys within two narrow cones in LLMs. We introduce Spotlight Attention, a novel method that employs non-linear hashing functions to optimize the embedding distribution of queries and keys, enhancing coding efficiency and robustness. We also developed a lightweight, stable training framework using a Bradley-Terry ranking-based loss, enabling optimization of the non-linear hashing module on GPUs with 16GB memory in 8 hours. Experimental results show that Spotlight Attention drastically improves retrieval precision while shortening the length of the hash code at least 5$\times$ compared to traditional linear hashing. Finally, we exploit the computational advantages of bitwise operations by implementing specialized CUDA kernels, achieving hashing retrieval for 512K tokens in under 100$\mu$s on a single A100 GPU, with end-to-end throughput up to 3$\times$ higher than vanilla decoding.

Person retrieval has attracted rising attention. Existing methods are mainly divided into two retrieval modes, namely image-only and text-only. However, they are unable to make full use of the available information and are difficult to meet diverse application requirements. To address the above limitations, we propose a new Composed Person Retrieval (CPR) task, which combines visual and textual queries to identify individuals of interest from large-scale person image databases. Nevertheless, the foremost difficulty of the CPR task is the lack of available annotated datasets. Therefore, we first introduce a scalable automatic data synthesis pipeline, which decomposes complex multimodal data generation into the creation of textual quadruples followed by identity-consistent image synthesis using fine-tuned generative models. Meanwhile, a multimodal filtering method is designed to ensure the resulting SynCPR dataset retains 1.15 million high-quality and fully synthetic triplets. Additionally, to improve the representation of composed person queries, we propose a novel Fine-grained Adaptive Feature Alignment (FAFA) framework through fine-grained dynamic alignment and masked feature reasoning. Moreover, for objective evaluation, we manually annotate the Image-Text Composed Person Retrieval (ITCPR) test set. The extensive experiments demonstrate the effectiveness of the SynCPR dataset and the superiority of the proposed FAFA framework when compared with the state-of-the-art methods. All code and data will be provided at https://github.com/Delong-liu-bupt/Composed_Person_Retrieval.

Training-free guided generation is a widely used and powerful technique that allows the end user to exert further control over the generative process of flow/diffusion models. Generally speaking, two families of techniques have emerged for solving this problem for *gradient-based guidance*: namely, *posterior guidance* (*i.e.*, guidance via projecting the current sample to the target distribution via the target prediction model) and *end-to-end guidance* (*i.e.*, guidance by performing backpropagation throughout the entire ODE solve). In this work, we show that these two seemingly separate families can actually be *unified* by looking at posterior guidance as a *greedy strategy* of *end-to-end guidance*. We explore the theoretical connections between these two families and provide an in-depth theoretical of these two techniques relative to the *continuous ideal gradients*. Motivated by this analysis we then show a method for *interpolating* between these two families enabling a trade-off between compute and accuracy of the guidance gradients. We then validate this work on several inverse image problems and property-guided molecular generation.

Deep reinforcement learning (DRL) has achieved remarkable success across multiple domains, including competitive games, natural language processing, and robotics. Despite these advancements, policies trained via DRL often struggle to generalize to evaluation environments with different parameters. This challenge is typically addressed by training with multiple contexts and/or by leveraging additional structure in the problem. However, obtaining sufficient training data across diverse contexts can be impractical in real-world applications. In this work, we consider contextual Markov decision processes (CMDPs) with transition and reward functions that exhibit regularity in context parameters. We introduce the context-enhanced Bellman equation (CEBE) to improve generalization when training on a single context. We prove both analytically and empirically that the CEBE yields a first-order approximation to the Q function trained across multiple contexts. We then derive context sample enhancement (CSE) as an efficient data augmentation method for approximating the CEBE in deterministic control environments. We numerically validate the performance of CSE in simulation environments, showcasing its potential to improve generalization in DRL.

Scaling up self-supervised learning has driven breakthroughs in language and vision, yet comparable progress has remained elusive in reinforcement learning (RL). In this paper, we study building blocks for self-supervised RL that unlock substantial improvements in scalability, with network depth serving as a critical factor. Whereas most RL papers in recent years have relied on shallow architectures (around 2 -- 5 layers), we demonstrate that increasing the depth up to 1024 layers can significantly boost performance. Our experiments are conducted in an unsupervised goal-conditioned setting, where no demonstrations or rewards are provided, so an agent must explore (from scratch) and learn how to maximize the likelihood of reaching commanded goals. Evaluated on simulated locomotion and manipulation tasks, our approach increases performance on the self-supervised contrastive RL algorithm by $2\times$ -- $50\times$, outperforming other goal-conditioned baselines. Increasing the model depth not only increases success rates but also qualitatively changes the behaviors learned.

Semi-Unbalanced Optimal Transport (SemiUOT) shows great promise in matching two probability measures by relaxing one of the marginal constraints. Previous solvers often incorporate an entropy regularization term, which can result in inaccurate matching solutions. To address this issue, we focus on determining the marginal probability distribution of SemiUOT with KL divergence using the proposed Equivalent Transformation Mechanism (ETM) approach. Furthermore, we extend the ETM-based method into exploiting the marginal probability distribution of Unbalanced Optimal Transport (UOT) with KL divergence for validating its generalization. Once the marginal probabilities of UOT/SemiUOT are determined, they can be transformed into a classical Optimal Transport (OT) problem. Moreover, we propose a KKT-Multiplier regularization term combined with Multiplier Regularized Optimal Transport (MROT) to achieve more accurate matching results. We conduct several numerical experiments to demonstrate the effectiveness of our proposed methods in addressing UOT/SemiUOT problems.

Serial dependence reflects how recent sensory history shapes current perception, producing two opposing biases: repulsion, where perception is repelled from recent stimuli, and attraction, where perception is drawn toward them. Repulsion typically occurs at the sensory perception stage, while attraction arises at the post-perception stage. To uncover the neural basis of these effects, we developed a two-layer continuous attractor neural network model incorporating synaptic short-term plasticity (STP). The lower layer, dominated by synaptic depression, models sensory processing and drives repulsion due to sustained neurotransmitter depletion. The higher layer, dominated by synaptic facilitation, models post-perception processing and drives attraction by sustained high neurotransmitter release probability. Our model successfully explains the serial dependence phenomena observed in the visual orientation judgment experiments, highlighting STP as the critical mechanism, with its time constants defining the temporal windows of repulsion and attraction. Furthermore, the model provides a neural foundation for the Bayesian interpretation of serial dependence. This study advances our understanding of how the neural system leverages STP to balance sensitivity in sensory perception with stability in post-perceptual cognition.

Tables and table-based use cases play a crucial role in many important real-world applications, such as spreadsheets, databases, and computational notebooks, which traditionally require expert-level users like data engineers, data analysts, and database administrators to operate. Although LLMs have shown remarkable progress in working with tables (e.g., in spreadsheet and database copilot scenarios), comprehensive benchmarking of such capabilities remains limited. In contrast to an extensive and growing list of NLP benchmarks, evaluations of table-related tasks are scarce, and narrowly focus on tasks like NL-to-SQL and Table-QA, overlooking the broader spectrum of real-world tasks that professional users face. This gap limits our understanding and model progress in this important area. In this work, we introduce MMTU, a large-scale benchmark with over 28K questions across 25 real-world table tasks, designed to comprehensively evaluate models ability to understand, reason, and manipulate real tables at the expert-level. These tasks are drawn from decades’ worth of computer science research on tabular data, with a focus on complex table tasks faced by professional users. We show that MMTU require a combination of skills -- including table understanding, reasoning, and coding -- that remain challenging for today's frontier models, where even frontier reasoning models like OpenAI GPT-5 and DeepSeek R1 score only around 69% and 57% respectively, suggesting significant room for improvement. We highlight key findings in our evaluation using MMTU and hope this benchmark drives further advances in understanding and developing foundation models for structured data processing and analysis. Our code and data are available at https://github.com/MMTU-Benchmark/MMTU and https://huggingface.co/datasets/MMTU-benchmark/MMTU.

We study the problem of preconditioning in the setting of sequential prediction. From the theoretical lens of linear dynamical systems, we show that applying a convolution to the input sequence translates to applying a polynomial to the unknown transition matrix in the hidden space. With this insight, we develop a novel preconditioning method that convolves the input sequence with the coefficients of the Chebyshev or Legendre polynomials. We formally prove that this improves the regret of a wide family of prediction methods. We proceed to apply this preconditioning technique to the method of spectral filtering. This gives the first sublinear regret bound that is also hidden-dimension free (up to logarithmic factors) even when the hidden transition matrix is asymmetric. From rigorous experiments on synthetic data we show that our simple preconditioning method generalizes to both 1) settings where the data is \emph{not} from a linear dynamical system and 2) a broad range of learning algorithms, including recurrent neural networks.

Model merging provides a scalable alternative to multi-task training by combining specialized finetuned models through parameter arithmetic, enabling efficient deployment without the need for joint training or access to all task data. While recent methods have shown promise, existing evaluations are limited in both model scale and task diversity, leaving open questions about their applicability to large, domain-specialized LLMs. To tackle the challenges, we introduce MergeBench, a comprehensive evaluation suite designed to assess model merging at scale. MergeBench builds on state-of-the-art open-source language models, including Llama and Gemma families at 2B to 9B scales, and covers five key domains: instruction following, mathematics, multilingual understanding, coding and safety. We standardize finetuning and evaluation protocols, and assess eight representative merging methods across multi-task performance, forgetting and runtime efficiency. Based on extensive experiments, we provide practical guidelines for algorithm selection and share insights showing that model merging tends to perform better on stronger base models, with techniques such as merging coefficient tuning and sparsification improving knowledge retention. However, several challenges remain, including the computational cost on large models, the gap for in-domain performance compared to multi-task models, and the underexplored role of model merging in standard LLM training pipelines. We hope MergeBench provides a foundation for future research to advance the understanding and practical application of model merging.

This paper establishes a continuous time approximation, a piece-wise continuous differential equation, for the discrete Heavy-Ball (HB) momentum method with explicit discretization error. Investigating continuous differential equations has been a promising approach for studying the discrete optimization methods. Despite the crucial role of momentum in gradient-based optimization methods, the gap between the original dynamics and the continuous time approximations due to the discretization error has not been comprehensively bridged yet. In this work, we study the HB momentum method in continuous time while putting more focus on the discretization error to provide additional theoretical tools to this area. In particular, we design a first-order piece-wise continuous differential equation, where we add a number of counter terms to account for the discretization error explicitly. As a result, we provide a continuous time model for the HB momentum method that allows the control of discretization error to arbitrary order of the learning rate. As an application, we leverage it to find a new implicit regularization of the directional smoothness and investigate the implicit bias of HB for diagonal linear networks, indicating how our results can be used in deep learning. Our theoretical findings are further supported by numerical experiments.