NeurIPS 2025 Papers

Transformers have been widely adopted in natural language processing, computer vision, and other domains due to their exceptional performance across a variety of tasks. However, the computational cost of Transformers is prohibitively high, particularly when handling long input sequences, significantly increasing both training and inference time. Although various token pruning methods have been proposed to reduce the computational burden of Transformers, most approaches overlook critical differences in sequences in terms of length and complexity, leading to suboptimal compression efficiency. In this paper, we propose AD-TP, an Attribution-Driven Adaptive Token Pruning method designed to retain only the most informative tokens. We analyze the performance of using accumulated attention values to measure token importance and find that attention values do not accurately reflect the actual contribution of each token to text understanding. Additionally, we observe significant variations in the length and complexity of different sequences within the dataset. Based on these insights, we adopt Integrated Gradients to evaluate token importance and introduce a lightweight adaptive token retainer module that dynamically generates pruning configurations for each input sequence. In addition, we incorporate both teacher supervision and self-supervised learning objectives to enhance the training efficiency, accuracy, and robustness of the model. Experiments conducted on GLUE, SQuAD, and 20News demonstrate that AD-TP outperforms state-of-the-art token pruning and model compression methods in both accuracy and computational efficiency. On GLUE, AD-TP reduces FLOPs by an average of 7.8× while improving performance by 0.6%.

Generalized Category Discovery (GCD) focuses on classifying known categories while simultaneously discovering novel categories from unlabeled data. However, previous GCD methods face challenges due to inconsistent optimization objectives and category confusion. This leads to feature overlap and ultimately hinders performance on novel categories. To address these issues, we propose the Neural Collapse-inspired Generalized Category Discovery (NC-GCD) framework. By pre-assigning and fixing Equiangular Tight Frame (ETF) prototypes, our method ensures an optimal geometric structure and a consistent optimization objective for both known and novel categories. We introduce a Consistent ETF Alignment Loss that unifies supervised and unsupervised ETF alignment and enhances category separability. Additionally, a Semantic Consistency Matcher (SCM) is designed to maintain stable and consistent label assignments across clustering iterations. Our method significantly enhancing novel category accuracy and demonstrating its effectiveness.

Most expressive variational families---such as normalizing flows---lack practical convergence guarantees, as their theoretical assurances typically hold only at the intractable global optimum. In this work, we present a general recipe for constructing tuning-free, asymptotically exact variational flows on *arbitrary* state spaces from involutive MCMC kernels. The core methodological component is a novel representation of general involutive MCMC kernels as invertible, measure- preserving iterated random function systems, which act as the flow maps of our variational flows. This leads to three new variational families with provable total variation convergence. Our framework resolves key practical limitations of existing variational families with similar guarantees (e.g., MixFlows), while requiring substantially weaker theoretical assumptions. Finally, we demonstrate the competitive performance of our flows across tasks including posterior approximation, Monte Carlo estimates, and normalization constant estimation, outperforming or matching No-U-Turn sampler (NUTS) and black-box normalizing flows.

Reward shaping is effective in addressing the sparse-reward challenge in reinforcement learning (RL) by providing immediate feedback through auxiliary, informative rewards. Based on the reward shaping strategy, we propose a novel multi-task reinforcement learning framework that integrates a centralized reward agent (CRA) and multiple distributed policy agents. The CRA functions as a knowledge pool, aimed at distilling knowledge from various tasks and distributing it to individual policy agents to improve learning efficiency. Specifically, the shaped rewards serve as a straightforward metric for encoding knowledge. This framework not only enhances knowledge sharing across established tasks but also adapts to new tasks by transferring meaningful reward signals. We validate the proposed method on both discrete and continuous domains, including the representative Meta-World benchmark, demonstrating its robustness in multi-task sparse-reward settings and its effective transferability to unseen tasks.

Inspired by the human ability to learn and organize knowledge into hierarchical taxonomies with prototypes, this paper addresses key limitations in current deep hierarchical clustering methods. Existing methods often tie the structure to the number of classes and underutilize the rich prototype information available at intermediate hierarchical levels. We introduce deep taxonomic networks, a novel deep latent variable approach designed to bridge these gaps. Our method optimizes a large latent taxonomic hierarchy, specifically a complete binary tree structured mixture-of-Gaussian prior within a variational inference framework, to automatically discover taxonomic structures and associated prototype clusters directly from unlabeled data without assuming true label sizes. We analytically show that optimizing the ELBO of our method encourages the discovery of hierarchical relationships among prototypes. Empirically, our learned models demonstrate strong hierarchical clustering performance, outperforming baselines across diverse image classification datasets using our novel evaluation mechanism that leverages prototype clusters discovered at all hierarchical levels. Qualitative results further reveal that deep taxonomic networks discover rich and interpretable hierarchical taxonomies, capturing both coarse-grained semantic categories and fine-grained visual distinctions.

While deep neural networks possess the capability to perform semantic segmentation, producing a single deterministic output limits reliability in safety-critical applications, caused by uncertainty and annotation variability. To address this, stochastic segmentation models using Conditional Variational Autoencoders (CVAE), Bayesian networks, and diffusion have been explored. However, existing approaches suffer from limited latent expressiveness and interpretability. Furthermore, our experiments showed that models like Probabilistic U-Net rely excessively on high latent variance, leading to posterior collapse. This work propose a novel framework by integrating Gaussian Mixture Model (GMM) with Normalizing Flow (NF) in CVAE for stochastic segmentation. GMM structures the latent space into meaningful semantic clusters, while NF captures feature deformations with quantified uncertainty. Our method stabilizes latent distributions through constrained variance and mean ranges. Experiments on LIDC, Crack500, and Cityscapes datasets show that our approach outperformed state-of-the-art in curvilinear structure and medical image segmentation.

Point cloud registration is a fundamental task in 3D computer vision. Recent advances have shown that graph-based methods are effective for outlier rejection in this context. However, existing clique-based methods impose overly strict constraints and are NP-hard, making it difficult to achieve both robustness and efficiency. While the k-core reduces computational complexity, which only considers node degree and ignores higher-order topological structures such as triangles, limiting its effectiveness in complex scenarios. To overcome these limitations, we introduce the $k$-truss from graph theory into point cloud registration, leveraging triangle support as a constraint for inlier selection. We further propose a consensus voting-based low-scale sampling strategy to efficiently extract the structural skeleton of the point cloud prior to $k$-truss decomposition. Additionally, we design a spatial distribution score that balances coverage and uniformity of inliers, preventing selections that concentrate on sparse local clusters. Extensive experiments on KITTI, 3DMatch, and 3DLoMatch demonstrate that our method consistently outperforms both traditional and learning-based approaches in various indoor and outdoor scenarios, achieving state-of-the-art results.

Online Continual Learning (OCL) requires models to learn sequentially from data streams with limited memory. Rehearsal-based methods, particularly Experience Replay (ER), are commonly used in OCL scenarios. This paper revisits ER through the lens of $\epsilon$-constraint optimization, revealing that ER implicitly employs a soft constraint on past task performance, with its weighting parameter post-hoc defining a slack variable. While effective, ER's implicit and fixed slack strategy has limitations: it can inadvertently lead to updates that negatively impact generalization, and its fixed trade-off between plasticity and stability may not optimally balance current streaming with memory retention, potentially overfitting to the memory buffer. To address these shortcomings, we propose the \textbf{G}radient-Guided \textbf{E}psilon \textbf{C}onstraint (\textbf{GEC}) method for online continual learning. GEC explicitly formulates the OCL update as an $\epsilon$-constraint optimization problem, which minimize the loss on the current task data and transform the stability objective as constraints and propose a gradient-guided method to dynamically adjusts the update direction based on whether the performance on memory samples violates a predefined slack tolerance $\bar{\varepsilon}$: if forgetting exceeds this tolerance, GEC prioritizes constraint satisfaction; otherwise, it focuses on the current task while controlling the rate of increase in memory loss. Empirical evaluations on standard OCL benchmarks demonstrate GEC's ability to achieve a superior trade-off, leading to improved overall performance.

The widespread deployment of large language models (LLMs) allows users to access their capabilities via black-box APIs, but backdoor attacks pose serious security risks for API users by hijacking the model behavior. This highlights the importance of backdoor detection technologies to help users audit LLMs before use. However, most existing LLM backdoor defenses require white-box access or costly reverse engineering, limiting their practicality for resource-constrained users. Moreover, they mainly target classification tasks, leaving broader generative scenarios underexplored. To solve the problem, this paper introduces ICLScan, a lightweight framework that exploits targeted in-context learning (ICL) as illumination for backdoor detection in black-box LLMs, which effectively supports generative tasks without additional training or model modifications. ICLScan is based on our finding of backdoor susceptibility amplification: LLMs with pre-embedded backdoors are highly susceptible to new trigger implantation via ICL. Including only a small ratio of backdoor examples (containing ICL-triggered input and target output) in the ICL prompt can induce ICL trigger-specific malicious behavior in backdoored LLMs. ICLScan leverages this phenomenon to detect backdoored LLMs by statistically analyzing whether the success rate of new trigger injection via targeted ICL exceeds a threshold. It requires only multiple queries to estimate the backdoor success rate, overcoming black-box access and computational resource limitations. Extensive experiments across diverse LLMs and backdoor attacks demonstrate ICLScan's effectiveness and efficiency, achieving near-perfect detection performance (precision/recall/F1-score/ROC-AUC all approaching 1) with minimal additional overhead across all settings.

Time-resolved single-cell omics data offers high-throughput, genome-wide measurements of cellular states, which are instrumental to reverse-engineer the processes underpinning cell fate. Such technologies are inherently destructive, allowing only cross-sectional measurements of the underlying stochastic dynamical system. Furthermore, cells may divide or die in addition to changing their molecular state. Collectively these present a major challenge to inferring realistic biophysical models. We present a novel approach, unbalanced probability flow inference, that addresses this challenge for biological processes modelled as stochastic dynamics with growth. By leveraging a Lagrangian formulation of the Fokker-Planck equation, our method accurately disentangles drift from intrinsic noise and growth. We showcase the applicability of our approach through evaluation on a range of simulated and real single-cell RNA-seq datasets. Comparing to several existing methods, we find our method achieves higher accuracy while enjoying a simple two-step training scheme.

In this paper, we develop a novel algorithm for constructing time-uniform, asymptotic confidence sequences for quantiles under local differential privacy (LDP). The procedure combines dynamically chained parallel stochastic gradient descent (P-SGD) with a randomized response mechanism, thereby guaranteeing privacy protection while simultaneously estimating the target quantile and its variance. A strong Gaussian approximation for the proposed estimator yields asymptotically anytime-valid confidence sequences whose widths obey the law of the iterated logarithm (LIL). Moreover, the method is fully online, offering high computational efficiency and requiring only $\mathcal{O}(\kappa)$ memory, where $\kappa$ denotes the number of chains and is much smaller than the sample size. Rigorous mathematical proofs and extensive numerical experiments demonstrate the theoretical soundness and practical effectiveness of the algorithm.

Reinforcement learning (RL) post-training is crucial for LLM alignment and reasoning, but existing policy-based methods, such as PPO and DPO, can fall short of fixing shortcuts inherited from pre-training. In this work, we introduce $Q\sharp$, a value-based algorithm for KL-regularized RL that guides the reference policy using the optimal regularized $Q$ function. We propose to learn the optimal $Q$ function using distributional RL on an aggregated online dataset. Unlike prior value-based baselines that guide the model using unregularized $Q$-values, our method is theoretically principled and provably learns the optimal policy for the KL-regularized RL problem. Empirically, $Q\sharp$ outperforms prior baselines in math reasoning benchmarks while maintaining a smaller KL divergence to the reference policy. Theoretically, we establish a reduction from KL-regularized RL to no-regret online learning, providing the first bounds for deterministic MDPs under only realizability. Thanks to distributional RL, our bounds are also variance-dependent and converge faster when the reference policy has small variance. In sum, our results highlight $Q\sharp$ as an effective approach for post-training LLMs, offering both improved performance and theoretical guarantees. The code can be found at \url{https://github.com/jinpz/q_sharp}.

Reasoning models are large language models that use extra "thought tokens" before answering, providing both higher accuracy and explicit reasoning for their response. A major question has been whether language model reasoning generalizes beyond mathematics, programming, and logic, where most previous work has focused. We demonstrate that reasoning models can be post-trained in scientific domains without additional domain pretraining, and require substantially less data compared to contemporary domain-specific models. We report ether0, a 24B parameter LLM (based on Mistral-Small-24B) that can reason in natural language and respond with chemical structures. This reasoning model was trained with reinforcement learning on 577,790 experimentally-grounded chemistry tasks involving synthesized organic molecules. Our model outperforms all previous general-purpose chemistry models, frontier models, and humans, and is more data efficient relative to specialized models. We anticipate that this method can be applied to train highly data-efficient language models specialized for predictive and generative tasks across a wide variety of scientific domains.

As large language models (LLMs) gain popularity in conducting prediction tasks in-context, understanding the sources of uncertainty in in-context learning becomes essential to ensuring reliability. The recent hypothesis of in-context learning performing predictive Bayesian inference opens the avenue for Bayesian uncertainty estimation, particularly for decomposing uncertainty into epistemic uncertainty due to lack of in-context data and aleatoric uncertainty inherent in the in-context prediction task. However, the decomposition idea remains under-explored due to the intractability of the latent parameter posterior from the underlying Bayesian model. In this work, we introduce a variational uncertainty decomposition framework for in-context learning without explicitly sampling from the latent parameter posterior, by optimising auxiliary inputs as probes to obtain an upper bound to the aleatoric uncertainty of an LLM's in-context learning procedure. Through experiments on synthetic and real-world tasks, we show quantitatively and qualitatively that the decomposed uncertainties obtained from our method exhibit desirable properties of epistemic and aleatoric uncertainty.

Multimodal Large Language Models (MLLMs) increasingly excel at perception,understanding, and reasoning. However, current benchmarks inadequately evaluate their ability to perform these tasks continuously in dynamic, real-world environments. To bridge this gap, we introduce RT V-Bench, a fine-grained benchmark for MLLM real-time video analysis. RTV-Bench includes three key principles: (1) Multi-Timestamp Question Answering (MTQA), where answers evolve with scene changes; (2) Hierarchical Question Structure, combining basic and advanced queries; and (3) Multi-dimensional Evaluation, assessing the ability of continuous perception, understanding, and reasoning. RTV-Bench contains 552 diverse videos (167.2 hours) and 4,631 high-quality QA pairs. We evaluated leading MLLMs, including proprietary (GPT-4o, Gemini 2.0), open-source offline (Qwen2.5-VL, VideoLLaMA3), and open-source real-time (VITA-1.5, InternLM-XComposer2.5-OmniLive) models. Experiment results show open-source real-time models largely outperform offline ones but still trail top proprietary models. Our analysis also reveals that larger model size or higher frame sampling rates do not significantly boost RTV-Bench performance, sometimes causing slight decreases.This underscores the need for better model architectures optimized for video stream processing and long sequences to advance real-time video analysis with MLLMs.

Accurate anatomical structure detection is a critical preliminary step for diagnosing diseases characterized by structural abnormalities. In clinical practice, medical experts frequently adjust the zoom level of medical images to obtain comprehensive views for diagnosis. This common interaction results in significant variations in the apparent scale of anatomical structures across different images or fields of view. However, the information embedded in these zoom-induced scale changes is often overlooked by existing detection algorithms. In addition, human organs possess a priori, fixed topological knowledge. To overcome this limitation, we propose ZR-DETR, a zoom-aware probabilistic framework tailored for medical object detection. ZR-DETR uniquely incorporates scale-sensitive zoom embeddings, anatomical relation constraints, and a Gaussian Process-based detection head. This architecture enables the framework to jointly model semantic context, enforce anatomical plausibility, and quantify detection uncertainty. Empirical validation across three diverse medical imaging benchmarks demonstrates that ZR-DETR consistently outperforms strong baselines in both single-domain and unsupervised domain adaptation scenarios.

Low-light image enhancement (LLIE) faces persistent challenges in balancing reconstruction fidelity with cross-scenario generalization. While existing methods predominantly focus on deterministic pixel-level mappings between paired low/normal-light images, they often neglect the continuous physical process of luminance transitions in real-world environments, leading to performance drop when normal-light references are unavailable. Inspired by empirical analysis of natural luminance dynamics revealing power-law distributed intensity transitions, this paper introduces Luminance-Aware Statistical Quantification (LASQ), a novel framework that reformulates LLIE as a statistical sampling process over hierarchical luminance distributions. Our LASQ re-conceptualizes luminance transition as a power-law distribution in intensity coordinate space that can be approximated by stratified power functions, therefore, replacing deterministic mappings with probabilistic sampling over continuous luminance layers. A diffusion forward process is designed to autonomously discover optimal transition paths between luminance layers, achieving unsupervised distribution emulation without normal-light references. In this way, it considerably improves the performance in practical situations, enabling more adaptable and versatile light restoration. This framework is also readily applicable to cases with normal-light references, where it achieves superior performance on domain-specific datasets alongside better generalization-ability across non-reference datasets. The code is available at: https://github.com/XYLGroup/LASQ.

Theoretical efforts to prove advantages of Transformers in comparison with classical architectures such as feedforward and recurrent neural networks have mostly focused on representational power. In this work, we take an alternative perspective and prove that even with infinite compute, feedforward and recurrent networks may suffer from larger sample complexity compared to Transformers, as the latter can adapt to a form of dynamic sparsity. Specifically, we consider a sequence-to-sequence data generating model on sequences of length $N$, where the output at each position only depends on $q \ll N$ relevant tokens, and the positions of these tokens are described in the input prompt. We prove that a single-layer Transformer can learn this model if and only if its number of attention heads is at least $q$, in which case it achieves a sample complexity almost independent of $N$, while recurrent networks require $N^{\Omega(1)}$ samples on the same problem. If we simplify this model, recurrent networks may achieve a complexity almost independent of $N$, while feedforward networks still require $N$ samples. Our proposed sparse retrieval model illustrates a natural hierarchy in sample complexity across these architectures.

Differentially private stochastic gradient descent (DP-SGD) is the most widely used method for training machine learning models with provable privacy guarantees. A key challenge in DP-SGD is setting the per-sample gradient clipping threshold, which significantly affects the trade-off between privacy and utility. While recent adaptive methods improve performance by adjusting this threshold during training, they operate in the standard coordinate system and fail to account for correlations across the coordinates of the gradient. We propose GeoClip, a geometry-aware framework that clips and perturbs gradients in a transformed basis aligned with the geometry of the gradient distribution. GeoClip adaptively estimates this transformation using only previously released noisy gradients, incurring no additional privacy cost. We provide convergence guarantees for GeoClip and derive a closed-form solution for the optimal transformation that minimizes the amount of noise added while keeping the probability of gradient clipping under control. Experiments on both tabular and image datasets demonstrate that GeoClip consistently outperforms existing adaptive clipping methods under the same privacy budget.