NeurIPS 2025 Papers

AI agents today are mostly siloed — they either retrieve and reason over vast amount of digital information and knowledge obtained online; or interact with the physical world through embodied perception, planning and action — but rarely both. This separation limits their ability to solve tasks that require integrated physical and digital intelligence, such as cooking from online recipes, navigating with dynamic map data, or interpreting real-world landmarks using web knowledge. We introduce \textsc{Embodied Web Agents}, a novel paradigm for AI agents that fluidly bridge embodiment and web-scale reasoning. To operationalize this concept, we first develop the \textsc{Embodied Web Agents} task environments, a unified simulation platform that integrates realistic 3D indoor and outdoor environments with functional web interfaces. Building upon this platform, we construct and release the \textsc{Embodied Web Agents} Benchmark, which encompasses a diverse suite of tasks including cooking, navigation, shopping, tourism, and geolocation — all requiring coordinated reasoning across physical and digital realms for systematic assessment of cross-domain intelligence. Experimental results reveal significant performance gaps between state-of-the-art AI systems and human capabilities, establishing both challenges and opportunities at the intersection of embodied cognition and web-scale knowledge access.

Hybrid Optical Neural Networks (ONNs, typically consisting of an optical frontend and a digital backend) offer an energy-efficient alternative to fully digital deep networks for real-time, power-constrained systems. However, their adoption is limited by two main challenges: the accuracy gap compared to large-scale networks during training, and discrepancies between simulated and fabricated systems that further degrade accuracy. While previous work has proposed end-to-end optimizations for specific datasets (e.g., MNIST) and optical systems, these approaches typically lack generalization across tasks and hardware designs. To address these limitations, we propose a task-agnostic and hardware-agnostic pipeline that supports image classification and segmentation across diverse optical systems. To assist optical system design before training, we design the metasurface layout based on fabrication constraints. For training, we introduce Neural Tangent Knowledge Distillation (NTKD), which aligns optical models with electronic teacher networks, thereby narrowing the accuracy gap. After fabrication, NTKD also guides fine-tuning of the digital backend to compensate for implementation errors. Experiments on multiple datasets (e.g., MNIST, CIFAR, Carvana Masking) and hardware configurations show that our pipeline consistently improves ONN performance and enables practical deployment in both pre-fabrication simulations and physical implementations.

Progress in offline reinforcement learning (RL) has been impeded by ambiguous problem definitions and entangled algorithmic designs, resulting in inconsistent implementations, insufficient ablations, and unfair evaluations. Although offline RL explicitly avoids environment interaction, prior methods frequently employ extensive, undocumented online evaluation for hyperparameter tuning, complicating method comparisons. Moreover, existing reference implementations differ significantly in boilerplate code, obscuring their core algorithmic contributions. We address these challenges by first introducing a rigorous taxonomy and a transparent evaluation protocol that explicitly quantifies online tuning budgets. To resolve opaque algorithmic design, we provide clean, minimalistic, single-file implementations of various model-free and model-based offline RL methods, significantly enhancing clarity and achieving substantial speed-ups. Leveraging these streamlined implementations, we propose Unifloral, a unified algorithm that encapsulates diverse prior approaches and enables development within a single, comprehensive hyperparameter space. Using Unifloral with our rigorous evaluation protocol, we develop two novel algorithms - TD3-AWR (model-free) and MoBRAC (model-based) - which substantially outperform established baselines. Our implementation is publicly available at https://github.com/EmptyJackson/unifloral.

Large language models (LLMs) have demonstrated remarkable performance across a wide range of natural language processing tasks. However, their exponentially increasing parameters pose significant challenges for deployment on resource-constrained devices. Vector Quantization (VQ) shows great promise for low-bit quantization (e.g., 2 to 4 bits), but existing work faces two key challenges: unconstrained direction error and suboptimal bit allocation. In this paper, we propose RSAVQ, a novel VQ framework to enhance extremely low-bit quantization for LLMs. RSAVQ introduces two geometry-driven innovations that effectively mitigate above limitations: (1) Error Direction Sensitivity Guidance (EDSG), which leverages the Fisher information matrix (FIM)-induced Riemannian metric to project quantization errors onto low-sensitivity directions in the parameter space. Specifically, this projection is performed along the negative natural gradient direction, which effectively suppresses error expansion. (2) Weight Channel Sensitivity Guidance (WCSG) , which constructs a channel-wise sensitivity metric via FIM curvature analysis to dynamically guide bit resource allocation. The approach facilitates a globally optimal quantization solution within prescribed bit constraints. Experiments demonstrate that RSAVQ outperforms existing methods for LLMs. For example, in 2-bit quantization of LLaMA-3 8B, RSAVQ leads baselines like VPTQ and QuIP\# by 0.4 in perplexity (PPL) and 1.5 in zero-shot accuracy. This work offers a practical solution for constrained environments and a theoretical bridge between information geometry and the quantization of neural networks, advancing efficient deep learning.

Prompt caching is critical for reducing latency and cost in LLM inference---OpenAI and Anthropic report up to 50–90\% cost savings through prompt reuse. Despite its widespread success, little is known about what constitutes an optimal prompt caching policy, particularly when optimizing tail latency—a metric of central importance to practitioners. The widely used Least Recently Used (LRU) policy can perform arbitrarily poor on this metric, as it is oblivious to the heterogeneity of conversation lengths. To address this gap, we propose Tail-Optimized LRU, a simple two-line modification that reallocates KV cache capacity to prioritize high-latency conversations by evicting cache entries that are unlikely to affect future turns. Though the implementation is simple, we prove its optimality under a natural stochastic model of conversation dynamics, providing the first theoretical justification for LRU in this setting---a result that may be of independent interest to the caching community. Experimentally, on real conversation data WildChat~\citep{zhao2024wildchat}, Tail-Optimized LRU achieves up to 27.5\% reduction in P90 tail Time to First Token latency and 23.9\% in P95 tail latency compared to LRU, along with up to 38.9\% decrease in SLO violations of 200ms. We believe this provides a practical and theoretically grounded option for practitioners seeking to optimize tail latency in real-world LLM deployments.

With the ever-increasing volume of visual data, the efficient and lossless transmission, along with its subsequent interpretation and understanding, has become a critical bottleneck in modern information systems. The emerged codebook-based solution utilize a globally shared codebook to quantize and dequantize each token, controlling the bpp by adjusting the number of tokens or the codebook size. However, for facial images—which are rich in attributes—such global codebook strategies overlook both the category-specific correlations within images and the semantic differences among tokens, resulting in suboptimal performance, especially at low bpp. Motivated by these observations, we propose a Switchable Token-Specific Codebook Quantization for face image compression, which learns distinct codebook groups for different image categories and assigns an independent codebook to each token. By recording the codebook group to which each token belongs with a small number of bits, our method can reduce the loss incurred when decreasing the size of each codebook group. This enables a larger total number of codebooks under a lower overall bpp, thereby enhancing the expressive capability and improving reconstruction performance. Owing to its generalizable design, our method can be integrated into any existing codebook-based representation learning approach and has demonstrated its effectiveness on face recognition datasets, achieving an average accuracy of 93.51\% for reconstructed images at 0.05 bpp.

We study a two-sided market, wherein, price-sensitive heterogeneous customers and servers arrive and join their respective queues. A compatible customer-server pair can then be matched by the platform, at which point, they leave the system. Our objective is to design pricing and matching algorithms that maximize the platform's profit, while maintaining reasonable queue lengths. As the demand and supply curves governing the price-dependent arrival rates may not be known in practice, we design a novel online-learning-based pricing policy and establish its near-optimality. In particular, we prove a tradeoff among three performance metrics: $\tilde{O}(T^{1-\gamma})$ regret, $\tilde{O}(T^{\gamma/2})$ average queue length, and $\tilde{O}(T^{\gamma})$ maximum queue length for $\gamma \in (0, 1/6]$, significantly improving over existing results (Yang & Ying, 2024). Moreover, barring the permissible range of $\gamma$, we show that this trade-off between regret and average queue length is optimal up to logarithmic factors under a class of policies, matching the optimal one as in (Varma et al., 2023) which assumes the demand and supply curves to be known. Our proposed policy has two noteworthy features: a dynamic component that optimizes the tradeoff between low regret and small queue lengths; and a probabilistic component that resolves the tension between obtaining useful samples for fast learning and maintaining small queue lengths.

Offline reinforcement learning (RL) aims to learn a policy from a fixed dataset without additional environment interaction. However, effective offline policy learning often requires a large and diverse dataset to mitigate epistemic uncertainty. Collecting such data demands substantial online interactions, which are costly or infeasible in many real-world domains. Therefore, improving policy learning from limited offline data—achieving high data efficiency—is critical for practical offline RL. In this paper, we propose a simple yet effective plug-and-play pretraining framework that initializes the feature representation of a $Q$-network to enhance data efficiency in offline RL. Our approach employs a shared $Q$-network architecture trained in two stages: pretraining a backbone feature extractor with a transition prediction head; training a $Q$-network—combining the backbone feature extractor and a $Q$-value head—with *any* offline RL objective. Extensive experiments on the D4RL, Robomimic, V-D4RL, and ExoRL benchmarks show that our method substantially improves both performance and data efficiency across diverse datasets and domains. Remarkably, with only **10\%** of the dataset, our approach outperforms standard offline RL baselines trained on the full data.

The AlphaFold Protein Structure Database (AFDB) offers unparalleled structural coverage at near-experimental accuracy, positioning it as a valuable resource for data-driven protein design. However, its direct use in training deep models that are sensitive to fine-grained atomic geometry—such as inverse folding—exposes a critical limitation. Comparative analysis of structural feature distributions reveals that AFDB structures exhibit distinct statistical regularities, reflecting a systematic geometric bias that deviates from the conformational diversity found in experimentally determined structures from the Protein Data Bank (PDB). While AFDB structures are cleaner and more idealized, PDB structures capture the intrinsic variability and physical realism essential for generalization in downstream tasks. To address this discrepancy, we introduce a Debiasing Structure AutoEncoder (DeSAE) that learns to reconstruct native-like conformations from intentionally corrupted backbone geometries. By training the model to recover plausible structural states, DeSAE implicitly captures a more robust and natural structural manifold. At inference, applying DeSAE to AFDB structures produces debiased structures that significantly improve inverse folding performance across multiple benchmarks. This work highlights the critical impact of subtle systematic biases in predicted structures and presents a principled framework for debiasing, significantly boosting the performance of structure-based learning tasks like inverse folding.

Gene regulatory network (GRN) inference serves as a cornerstone for deciphering cellular decision-making processes. Early approaches rely exclusively on gene expression data, thus their predictive power remain fundamentally constrained by the vast combinatorial space of potential gene-gene interactions. Subsequent methods integrate prior knowledge to mitigate this challenge by restricting the solution space to biologically plausible interactions. However, we argue that the effectiveness of these approaches is contingent upon the precision of prior information and the reduction in the search space will circumscribe the models' potential for novel biological discoveries. To address these limitations, we introduce KINDLE, a three-stage framework that decouples GRN inference from prior knowledge dependencies. KINDLE trains a teacher model that integrates prior knowledge with temporal gene expression dynamics and subsequently distills this encoded knowledge to a student model, enabling accurate GRN inference solely from expression data without access to any prior. KINDLE achieves state-of-the-art performance across four benchmark datasets. Notably, it successfully identifies key transcription factors governing mouse embryonic development and precisely characterizes their functional roles. In mouse hematopoietic stem cell data, KINDLE accurately predicts fate transition outcomes following knockout of two critical regulators (Gata1 and Spi1). These biological validations demonstrate our framework's dual capability in maintaining topological inference precision while preserving discovery potential for novel biological mechanisms.

Linear attention has attracted interest as a computationally efficient approximation to softmax attention, especially for long sequences. Recent studies has explored distilling softmax attention in pre-trained Transformers into linear attention. However, a critical challenge remains: *how to choose the feature dimension that governs the approximation quality*. Existing methods fix this dimension uniformly across all attention layers, overlooking the diverse roles and complexities of them. In this paper, we propose a principled method to automatically determine the feature dimension in linear attention using the concept of statistical *degrees of freedom*, which represent the effective dimensionality of the inputs. We provide a theoretical bound on the approximation error and show that the dimension chosen by our method achieves smaller errors under a fixed computational budget. Furthermore, we introduce an efficient layerwise training strategy to learn nonlinear features tailored to each layer. Experiments on multiple pre-trained transformers demonstrate that our method improves the performance of distilled models compared to baselines without increasing the inference cost. Our findings also provide insight into how the complexity of the attention mechanism evolves across layers.

Mixture-of-Experts (MoE) architectures enable efficient scaling of large language models but face prohibitive memory demands due to massive parameterization. Existing pruning methods rely on heuristic metrics or impractical enumeration of expert subsets, leading to suboptimal performance or scalability. In this paper, we propose Shapley-MoE, an efficient pruning method for MoE models inspired by cooperative game theory. By quantifying each expert’s contribution via Shapley value, our method identifies important experts without exhaustive combination evaluations. To overcome the NP-hard complexity of exact Shapley computation, we introduce a Monte Carlo sampling strategy for efficient approximation that reduces complexity to quadratic time. However, vanilla Monte Carlo sampling still faces issues of insufficient estimation accuracy and low sampling efficiency. To address these issues, we further propose two novel methods to improve sampling accuracy and efficiency: (1) Early Truncation, which early terminates unstable sampling steps caused by overly small expert subsets, and (2) Router-Guided Importance Sampling, which prioritize sampling important expert subsets using gating activation probabilities. Both theoretical and experimental analyses show that both methods can accelerate Shapley value estimation and improve accuracy. Extensive empirical evaluations demonstrate that our pruned MoE models outperform existing expert pruning methods. Notably, when applied to the Qwen2-57B-A14B model, our method reduces the number of experts by 25% with only a 0.92 increase in perplexity and over 96.4% of the average zero-shot accuracy is maintained.

Unsupervised and open-vocabulary 3D object detection has recently gained attention, particularly in autonomous driving, where reducing annotation costs and recognizing unseen objects are critical for both safety and scalability. However, most existing approaches uniformly annotate 3D bounding boxes, ignore objects’ physical states, and require multiple self-training iterations for annotation refinement, resulting in suboptimal quality and substantial computational overhead. To address these challenges, we propose OpenBox, a two-stage automatic annotation pipeline that leverages a 2D vision foundation model. In the first stage, OpenBox associates instance-level cues from 2D images processed by a vision foundation model with the corresponding 3D point clouds via context-aware refinement. In the second stage, it categorizes instances by rigidity and motion state, then generates adaptive bounding boxes with class-specific size statistics. As a result, OpenBox produces high-quality 3D bounding box annotations without requiring self-training. Experiments on the Waymo Open Dataset (WOD), the Lyft Level 5 Perception dataset, and the nuScenes dataset demonstrate improved accuracy and efficiency over baselines.

Modeling the transport dynamics of natural processes from population-level observations is a ubiquitous problem in the natural sciences. Such models rely on key assumptions about the underlying process in order to enable faithful learning of governing dynamics that mimic the actual system behavior. The de facto assumption in current approaches relies on the principle of least action that results in gradient field dynamics and leads to trajectories minimizing an energy functional between two probability measures. However, many real-world systems, such as cell cycles in single-cell RNA, are known to exhibit non-gradient, periodic behavior, which fundamentally cannot be captured by current state-of-the-art methods such as flow and bridge matching. In this paper, we introduce Curly Flow Matching (Curly-FM), a novel approach that is capable of learning non-gradient field dynamics by designing and solving a Schrödinger bridge problem with a non-zero drift reference process---in stark contrast to typical zero-drift reference processes---which is constructed using inferred velocities in addition to population snapshot data. We showcase Curly-FM by solving the trajectory inference problems for single cells, computational fluid dynamics, and ocean currents with approximate velocities. We demonstrate that Curly-FM can learn trajectories that better match both the reference process and population marginals. Curly-FM expands flow matching models beyond the modeling of populations and towards the modeling of known periodic behavior in physical systems. Our code repository is accessible at: https://github.com/kpetrovicc/curly-flow-matching.git

Out-of-tree kernel patches are essential for adapting the Linux kernel to new hardware or enabling specific functionalities. Maintaining and updating these patches across different kernel versions demands significant effort from experienced engineers. Large language models (LLMs) have shown remarkable progress across various domains, suggesting their potential for automating out-of-tree kernel patch migration. However, our findings reveal that LLMs, while promising, struggle with incomplete code context understanding and inaccurate migration point identification. In this work, we propose MigGPT, a framework that employs a novel code fingerprint structure to retain code snippet information and incorporates three meticulously designed modules to improve the migration accuracy and efficiency of out-of-tree kernel patches. Furthermore, we establish a robust benchmark using real-world out-of-tree kernel patch projects to evaluate LLM capabilities. Evaluations show that MigGPT significantly outperforms the direct application of vanilla LLMs, achieving an average completion rate of 72.59\% ($\uparrow 50.74\%$) for migration tasks.

Generalized Category Discovery (GCD) aims to classify test-time samples into either seen categories—available during training—or novel ones, without relying on label supervision. Most existing GCD methods assume simultaneous access to labeled and unlabeled data during training and arising from the same domain, limiting applicability in open-world scenarios involving distribution shifts. Domain Generalization with GCD (DG-GCD) lifts this constraint by requiring models to generalize to unseen domains containing novel categories, without accessing target-domain data during training. The only prior DG-GCD method, DG$^2$CD-Net~\cite{dg2net}, relies on episodic training with multiple synthetic domains and task vector aggregation, incurring high computational cost and error accumulation. We propose \textsc{HiDISC}, a hyperbolic representation learning framework that achieves domain and category-level generalization without episodic simulation. To expose the model to minimal but diverse domain variations, we augment the source domain using GPT-guided diffusion, avoiding overfitting while maintaining efficiency. To structure the representation space, we introduce \emph{Tangent CutMix}, a curvature-aware interpolation that synthesizes pseudo-novel samples in tangent space, preserving manifold consistency. A unified loss—combining penalized Busemann alignment, hybrid hyperbolic contrastive regularization, and adaptive outlier repulsion—facilitates compact, semantically structured embeddings. A learnable curvature parameter further adapts the geometry to dataset complexity. \textsc{HiDISC} achieves state-of-the-art results on PACS~\cite{pacs}, Office-Home~\cite{officehome}, and DomainNet~\cite{domainnet}, consistently outperforming the existing Euclidean and hyperbolic (DG)-GCD baselines.

Over time, a growing wave of large language models from various series has been introduced to the community. Researchers are striving to maximize the performance of language models with constrained parameter sizes. However, from a microscopic perspective, there has been limited research on how to better store knowledge in model parameters, particularly within MLPs, to enable more effective utilization of this knowledge by the model. In this work, we analyze twenty publicly available open-source large language models to investigate the relationship between their strong performance and the way knowledge is stored in their corresponding MLP parameters. Our findings reveal that as language models become more advanced and demonstrate stronger knowledge capabilities, their parameters exhibit increased specialization. Specifically, parameters in the MLPs tend to be more focused on encoding similar types of knowledge. We experimentally validate that this specialized distribution of knowledge contributes to improving the efficiency of knowledge utilization in these models. Furthermore, by conducting causal training experiments, we confirm that this specialized knowledge distribution plays a critical role in improving the model's efficiency in leveraging stored knowledge.

Recent advances have significantly improved our understanding of the generalization performance of gradient descent (GD) methods in deep neural networks. A natural and fundamental question is whether GD can achieve generalization rates comparable to the minimax optimal rates established in the kernel setting. Existing results either yield suboptimal rates of $O(1/\sqrt{n})$, or focus on networks with smooth activation functions, incurring exponential dependence on network depth $L$. In this work, we establish optimal generalization rates for GD with deep ReLU networks by carefully trading off optimization and generalization errors, achieving only polynomial dependence on depth. Specifically, under the assumption that the data are NTK separable from the margin $\gamma$, we prove an excess risk rate of $\widetilde{O}(L^4 (1 + \gamma L^2) / (n \gamma^2))$, which aligns with the optimal SVM-type rate $\widetilde{O}(1 / (n \gamma^2))$ up to depth-dependent factors. A key technical contribution is our novel control of activation patterns near a reference model, enabling a sharper Rademacher complexity bound for deep ReLU networks trained with gradient descent.

The famous theorem of Fritz John states that any convex body has a unique maximal volume inscribed ellipsoid, known as the John Ellipsoid. Computing the John Ellipsoid is a fundamental problem in convex optimization. In this paper, we focus on approximating the John Ellipsoid inscribed in a convex and centrally symmetric polytope defined by $P := \{ x \in \mathbb{R}^d : -\mathbf{1}_n \leq A x \leq \mathbf{1}_n \},$ where $ A \in \mathbb{R}^{n \times d}$ is a rank-$d$ matrix and $ \mathbf{1}_n \in \mathbb{R}^n $ is the all-ones vector. We develop two efficient algorithms for approximating the John Ellipsoid. The first is a sketching-based algorithm that runs in nearly input-sparsity time $ \widetilde{O}(\mathrm{nnz}(A) + d^\omega) $, where $ \mathrm{nnz}(A) $ denotes the number of nonzero entries in the matrix $A$ and $ \omega \approx 2.37$ is the current matrix multiplication exponent. The second is a treewidth-based algorithm that runs in time $ \widetilde{O}(n \tau^2)$, where $\tau$ is the treewidth of the dual graph of the matrix $A$. Our algorithms significantly improve upon the state-of-the-art running time of $ \widetilde{O}(n d^2) $ achieved by [Cohen, Cousins, Lee, and Yang, COLT 2019].