NeurIPS 2025 Papers

Machine learning has promised to change the landscape of laboratory chemistry, with impressive results in molecular property prediction and reaction retro-synthesis. However, chemical datasets are often inaccessible to the machine learning community as they tend to require cleaning, thorough understanding of the chemistry, or are simply not available. In this paper, we introduce a novel dataset for yield prediction, providing the first-ever transient flow dataset for machine learning benchmarking, covering over 1200 process conditions. While previous datasets focus on discrete parameters, our experimental set-up allow us to sample a large number of continuous process conditions, generating new challenges for machine learning models. We focus on solvent selection, a task that is particularly difficult to model theoretically and therefore ripe for machine learning applications. We showcase benchmarking for regression algorithms, transfer-learning approaches, feature engineering, and active learning, with important applications towards solvent replacement and sustainable manufacturing.

Visual anomaly detection is significant in safety-critical and reliability-sensitive scenarios. Prior studies mainly emphasize the design and training of scoring functions, while little effort has been devoted to constructing decision rules based on these score functions. A recent work Ma et al. (2025b) highlights this issue and proposes the SAC-BL algorithm to address it. This method consists of a strong anomaly constraint (SAC) network and a betting-like (BL) algorithm serving as the decision rule. The SAC-BL algorithm can control the false discovery rate (FDR). However the performance of SAC-BL algorithm on anomalous examples, or its false positive rate (FPR), has not been thoroughly investigated. This paper provides a deeper analysis of this problem and explores how to theoretically reduce its FPR. First, we show that as the number of testing examples tends to infinity, the SAC-BL algorithm performs well on abnormal data if the scores follow the generalized Gaussian-like distribution family. But such conditions about the number of testing examples and the distribution of scores are overly restrictive for the real-world applications. So, we attempt to decrease the FPR of the SAC-BL algorithm under the condition of finite samples for practical anomaly detection. To this end, we redesign the BL algorithm by incorporating a randomization strategy and propose a novel stochastic BL (SBL) algorithm. The combination of the SAC network and the SBL algorithm yields our method, SAC-SBL. Theoretical results show that the SAC-SBL algorithm can achieve smaller FPR than SAC-BL algorithm while controlling its FDR. Finally, extensive experimental results demonstrate the superiority of our method over SAC-BL algorithm on multiple visual anomaly detection benchmarks.

Video Anomaly Detection (VAD) aims to locate unusual activities or behaviors within videos. Recently, offline VAD has garnered substantial research attention, which has been invigorated by the progress in large language models (LLMs) and vision-language models (VLMs), offering the potential for a more nuanced understanding of anomalies. However, online VAD has seldom received attention due to real-time constraints and computational intensity. In this paper, we introduce a novel **M**emory-based online scoring queue scheme for **T**raining-free VAD (MoniTor), to address the inherent complexities in online VAD. Specifically, MoniTor applies a streaming input to VLMs, leveraging the capabilities of pre-trained large-scale models. To capture temporal dependencies more effectively, we incorporate a novel prediction mechanism inspired by Long Short-Term Memory (LSTM) networks. This ensures the model can effectively model past states and leverage previous predictions to identify anomalous behaviors. Thereby, it better understands the current frame. Moreover, we design a scoring queue and an anomaly prior to dynamically store recent scores and cover all anomalies in the monitoring scenario, providing guidance for LLMs to distinguish between normal and abnormal behaviors over time. We evaluate MoniTor on two large datasets (i.e., UCF-Crime and XD-Violence) containing various surveillance and real-world scenarios. The results demonstrate that MoniTor outperforms state-of-the-art methods and is competitive with weakly supervised methods without training. Code is available at https://github.com/YsTvT/MoniTor.

Accurate weather forecasting across time scales is critical for anticipating and mitigating the impacts of climate change. Recent data-driven methods based on deep learning have achieved significant success in the medium range, but struggle at longer subseasonal-to-seasonal (S2S) horizons due to error accumulation in their autoregressive approach. In this work, we propose OmniCast, a scalable and skillful probabilistic model that unifies weather forecasting across timescales. OmniCast consists of two components: a VAE model that encodes raw weather data into a continuous, lower-dimensional latent space, and a diffusion-based transformer model that generates a sequence of future latent tokens given the initial conditioning tokens. During training, we mask random future tokens and train the transformer to estimate their distribution given conditioning and visible tokens using a per-token diffusion head. During inference, the transformer generates the full sequence of future tokens by iteratively unmasking random subsets of tokens. This joint sampling across space and time mitigates compounding errors from autoregressive approaches. The low-dimensional latent space enables modeling long sequences of future latent states, allowing the transformer to learn weather dynamics beyond initial conditions. OmniCast performs competitively with leading probabilistic methods at the medium-range timescale while being 10× to 20× faster, and achieves state-of-the-art performance at the subseasonal-to-seasonal scale across accuracy, physics-based, and probabilistic metrics. Furthermore, we demonstrate that OmniCast can generate stable rollouts up to 100 years ahead. Code and model checkpoints are available at https://github.com/tung-nd/omnicast.

Point cloud completion is essential for robust 3D perception in safety-critical applications such as robotics and augmented reality. However, existing models perform static inference and rely heavily on inductive biases learned during training, limiting their ability to adapt to novel structural patterns and sensor-induced distortions at test time. To address this limitation, we propose PointMAC, a meta-learned framework for robust test-time adaptation in point cloud completion. It enables sample-specific refinement without requiring additional supervision. Our method optimizes the completion model under two self-supervised auxiliary objectives that simulate structural and sensor-level incompleteness. A meta-auxiliary learning strategy based on Model-Agnostic Meta-Learning (MAML) ensures that adaptation driven by auxiliary objectives is consistently aligned with the primary completion task. During inference, we adapt the shared encoder on-the-fly by optimizing auxiliary losses, with the decoder kept fixed. To further stabilize adaptation, we introduce Adaptive $\lambda$-Calibration, a meta-learned mechanism for balancing gradients between primary and auxiliary objectives. Extensive experiments on synthetic, simulated, and real-world datasets demonstrate that PointMAC achieves state-of-the-art results by refining each sample individually to produce high-quality completions. To the best of our knowledge, this is the first work to apply meta-auxiliary test-time adaptation to point cloud completion.

State-Space Models (SSMs) have proven to be powerful tools for online function approximation and for modeling long-range dependencies in sequential data. While recent methods such as HiPPO have demonstrated strong performance using a few polynomial bases, they remain limited by their reliance on closed-form solutions for specific, well-behaved bases. The SaFARi framework generalizes this approach, enabling the construction of SSMs from arbitrary frames, including non-orthogonal and redundant ones, thus allowing an infinite diversity of possible "species'' within the SSM family. In this paper, we introduce WaLRUS (Wavelets for Long-range Representation Using SSMs), a new species of SaFARi built from Daubechies wavelet frames. We instantiate two variants, scaled-Walrus and translated-Walrus, and show that their multiresolution and localized nature offers significant advantages in representing non-smooth and transient signals. We compare Walrus to HiPPO-based models and demonstrate improved accuracy, better numerical properties, and more efficient implementations for online function approximation tasks.

The sequential nature of modern LLMs makes them expensive and slow, and speculative sam- pling has proven to be an effective solution to this problem. Methods like EAGLE perform autoregression at the feature level, reusing top- layer features from the target model to achieve better results than vanilla speculative sampling. A growing trend in the LLM community is scaling up training data to improve model intelligence without increasing inference costs. However, we observe that scaling up data provides limited improvements for EAGLE. We identify that this limitation arises from EAGLE’s feature prediction constraints. In this paper, we introduce EAGLE-3, which abandons feature prediction in favor of direct token prediction and replaces reliance on top-layer features with multi-layer feature fusion via a technique named training-time test. These improvements significantly enhance performance and enable the draft model to fully benefit from scaling up training data. Our experiments include both chat models and reasoning models, evaluated on five tasks. The results show that EAGLE-3 achieves a speedup ratio up to 6.5x, with about 1.4x improvement over EAGLE-2. In the SGLang framework, EAGLE- 3 achieves a 1.38x throughput improvement at a batch size of 64.

Emotion recognition based on physiological signals plays a vital role in psychological health and human–computer interaction, particularly with the substantial advances in multimodal emotion recognition techniques. However, two key challenges remain unresolved: 1) how to effectively model the intra-modal long-range dependencies and inter-modal correlations in multimodal physiological emotion signals, and 2) how to address the performance limitations resulting from missing multimodal data. In this paper, we propose a multimodal bidirectional Mamba (BiMamba) network with test-time adaptation (TTA) for emotion recognition named BiM-TTA. Specifically, BiM-TTA consists of a multimodal BiMamba network and a multimodal TTA. The former includes intra-modal and inter-modal BiMamba modules, which model long-range dependencies along the time dimension and capture cross-modal correlations along the channel dimension, respectively. The latter (TTA) mitigates the amplified distribution shifts caused by missing multimodal data through two-level entropy-based sample filtering and mutual information sharing across modalities. By addressing these challenges, BiM-TTA achieves state-of-the-art results on two multimodal emotion datasets.

Diffusion models generate samples by estimating the score function of the target distribution at various noise levels. The model is trained using samples drawn from the target distribution, progressively adding noise. Previous sample complexity bounds have a polynomial dependence on the dimension $d$, apart from $\log({|\mathcal{H}|})$, where $\mathcal{H}$ is the hypothesis class. In this work, we establish the first (nearly) dimension-free sample complexity bounds, modulo any dependence due to $\log( |\mathcal{H}|)$, for learning these score functions, achieving a double exponential improvement in dimension over prior results. A key aspect of our analysis is to use a single function approximator to jointly estimate scores across noise levels, a critical feature in practice which enables generalization across timesteps. We introduce a novel martingale-based error decomposition and sharp variance bounds, enabling efficient learning from dependent data generated by Markov processes, which may be of independent interest. Building on these insights, we propose Bootstrapped Score Matching (BSM), a variance reduction technique that utilizes previously learned scores to improve accuracy at higher noise levels. These results provide crucial insights into the efficiency and effectiveness of diffusion models for generative modeling.

Can we scale 4D pretraining to learn general space-time representations that reconstruct an object from a few views at some times to any view at any time? We provide an affirmative answer with 4D-LRM, the first large-scale 4D reconstruction model that takes input from unconstrained views and timestamps and renders arbitrary novel view-time combinations. Unlike prior 4D approaches, e.g., optimization-based, geometry-based, or generative, that struggle with efficiency, generalization, or faithfulness, 4D-LRM learns a unified space-time representation and directly predicts per-pixel 4D Gaussian primitives from posed image tokens across time, enabling fast, high-quality rendering at, in principle, infinite frame rate. Our results demonstrate that scaling spatiotemporal pretraining enables accurate and efficient 4D reconstruction. We show that 4D-LRM generalizes to novel objects, interpolates across time, and handles diverse camera setups. It reconstructs 24-frame sequences in one forward pass with less than 1.5 seconds on a single A100 GPU.

Time series data is ubiquitous, with forecasting applications spanning from finance to healthcare. Beyond popular deterministic methods, generative models are gaining attention due to advancements in areas like image synthesis and video generation, as well as their inherent ability to provide probabilistic predictions. However, existing generative approaches mostly involve recurrent generative operations or repeated denoising steps, making the prediction laborious, particularly for long-term forecasting. Most of them only conduct experiments for relatively short-term forecasting, with limited comparison to deterministic methods in long-term forecasting, leaving their practical advantages unclear. This paper presents TARFVAE, a novel generative framework that combines the Transformer-based autoregressive flow (TARFLOW) and variational autoencoder (VAE) for efficient one-step generative time series forecasting. Inspired by the rethinking that complex architectures for extracting time series representations might not be necessary, we add a flow module, TARFLOW, to VAE to promote spontaneous learning of latent variables that benefit predictions. TARFLOW enhances VAE's posterior estimation by breaking the Gaussian assumption, thereby enabling a more informative latent space. TARFVAE uses only the forward process of TARFLOW, avoiding autoregressive inverse operations and thus ensuring fast generation. During generation, it samples from the prior latent space and directly generates full-horizon forecasts via the VAE decoder. With simple MLP modules, TARFVAE achieves superior performance over state-of-the-art deterministic and generative models across different forecast horizons on benchmark datasets while maintaining efficient prediction speed, demonstrating its effectiveness as an efficient and powerful solution for generative time series forecasting. Our code is available at https://github.com/Gavine77/TARFVAE.

Graph Neural Networks (GNNs) are vulnerable to backdoor attacks. Existing defenses primarily rely on detecting structural anomalies, distributional outliers, or perturbation-induced prediction instability, which struggle to handle the more subtle, feature-based attacks that do not introduce obvious topological changes. Our empirical analysis reveals that both structure-based and feature-based attacks not only cause early loss convergence of target nodes but also induce a class-coherent loss drift, where this early convergence gradually spreads to nearby clean nodes, leading to significant distribution overlap. To address this issue, we propose LoSplit, the first training-time defense framework in graph that leverages this early-stage loss drift to accurately split target nodes. Our method dynamically selects epochs with maximal loss divergence, clusters target nodes via Gaussian Mixture Models (GMM), and applies a Decoupling-Forgetting strategy to break the association between target nodes and malicious label. Extensive experiments on multiple real-world datasets demonstrate the effectiveness of our approach, significantly reducing attack success rates while maintaining high clean accuracy across diverse backdoor attack strategies.

Offline Reinforcement Learning (RL) provides a promising avenue for training policies from pre-collected datasets when gathering additional interaction data is infeasible. However, existing offline RL methods often assume stationarity or only consider synthetic perturbations at test time—assumptions that often fail in real-world scenarios characterized by abrupt, time-varying offsets. These offsets can lead to partial observability, causing agents to misperceive their true state and degrade performance. To overcome this challenge, we introduce Forecasting in Non-stationary Offline RL (FORL), a framework that unifies (i) conditional diffusion-based candidate state generation, trained without presupposing any specific form of future non-stationarity, and (ii) zero-shot time-series foundation models. FORL targets environments prone to unexpected, potentially non-Markovian offsets, requiring robust agent performance from the onset of each episode. Empirical evaluations on offline RL benchmarks, augmented with real-world time-series data to simulate realistic non-stationarity, demonstrate that FORL consistently improves performance compared to competitive baselines. By integrating zero-shot forecasting with the agent’s experience we aim to bridge the gap between offline RL and the complexity of real-world, non-stationary environments.

We introduce Toto, a time series forecasting foundation model with 151 million parameters. Toto uses a modern decoder-only architecture coupled with architectural innovations designed to account for specific challenges found in multivariate observability time series data. Toto's pre-training corpus is a mixture of observability data, open datasets, and synthetic data, and is 4-10$\times$ larger than those of leading time series foundation models. Additionally, we introduce BOOM, a large-scale benchmark consisting of 350 million observations across 2,807 real-world time series. For both Toto and BOOM, we source observability data exclusively from our own telemetry and internal observability metrics. Extensive evaluations demonstrate that Toto achieves state-of-the-art performance on both BOOM and on established general purpose time series forecasting benchmarks. Toto's model weights, inference code, and evaluation scripts, as well as BOOM's data and evaluation code, are all available as open source under the Apache 2.0 License.

The signature kernel is a recent state-of-the-art tool for analyzing high-dimensional sequential data, valued for its theoretical guarantees and strong empirical performance. In this paper, we present a novel method for efficiently computing the signature kernel of long, high-dimensional time series via adaptively truncated recursive local power series expansions. Building on the characterization of the signature kernel as the solution of a Goursat PDE, our approach employs tilewise Neumann‐series expansions to derive rapidly converging power series approximations of the signature kernel that are locally defined on subdomains and propagated iteratively across the entire domain of the Goursat solution by exploiting the geometry of the time series. Algorithmically, this involves solving a system of interdependent Goursat PDEs via adaptively truncated local power series expansions and recursive propagation of boundary conditions along a directed graph in a topological ordering. This method strikes an effective balance between computational cost and accuracy, achieving substantial performance improvements over state-of-the-art approaches for computing the signature kernel. It offers (a) adjustable and superior accuracy, even for time series with very high roughness; (b) drastically reduced memory requirements; and (c) scalability to efficiently handle very long time series (e.g., with up to one million data points or more) on a single GPU. As demonstrated in our benchmarks, these advantages make our method particularly well-suited for rough-path-assisted machine learning, financial modeling, and signal processing applications involving very long and highly volatile sequential data.

Consider this prompt "Draw a unicorn with two horns". Should large language models (LLMs) recognize that a unicorn has only one horn by definition and ask users for clarifications, or proceed to generate something anyway? We introduce *concept incongruence* to capture such phenomena where concept boundaries clash with each other, either in user prompts or in model representations, often leading to under-specified or mis-specified behaviors. In this work, we take the first step towards defining and analyzing model behavior under concept incongruence. Focusing on temporal boundaries in the Role-Play setting, we propose three behavioral metrics---abstention rate, conditional accuracy, and answer rate---to quantify model behavior under incongruence due to the role's death. We show that models fail to abstain after death and suffer from an accuracy drop compared to the Non-Role-Play setting. Through probing experiments, we identify two main causes: (i) unreliable encoding of the "death" state across different years, leading to unsatisfactory abstention behavior, and (ii) role playing causes shifts in the model’s temporal representations, resulting in accuracy drops. We leverage these insights to improve consistency in the model's abstention and answer behaviors. Our findings suggest that concept incongruence leads to unexpected model behaviors and point to future directions on improving model behavior under concept incongruence.

From clinical healthcare to daily living, continuous sensor monitoring across multiple modalities has shown great promise for real-world intelligent decision-making but also faces various challenges. In this work, we argue for modeling such heterogeneous data sources under the multimodal paradigm and introduce a new framework, MAESTRO. We introduce MAESTRO, a novel framework that overcomes key limitations of existing multimodal learning approaches: (1) reliance on a single primary modality for alignment, (2) pairwise modeling of modalities, and (3) assumption of complete modality observations. These limitations hinder the applicability of these approaches in real-world multimodal time-series settings, where primary modality priors are often unclear, the number of modalities can be large (making pairwise modeling impractical), and sensor failures often result in arbitrary missing observations. At its core, MAESTRO facilitates dynamic intra- and cross-modal interactions based on task relevance, and leverages symbolic tokenization and adaptive attention budgeting to construct long multimodal sequences, which are processed via sparse cross-modal attention. The resulting cross-modal tokens are routed through a sparse Mixture-of-Experts (MoE) mechanism, enabling black-box specialization under varying modality combinations. We evaluate MAESTRO against 10 baselines on four diverse datasets spanning three applications, and observe average relative improvements of 4% and 8% over the best existing multimodal and multivariate approaches, respectively, under complete observations. Under partial observations—with up to 40% of missing modalities—MAESTRO achieves an average 9% improvement. Further analysis also demonstrates the robustness and efficiency of MAESTRO's sparse, modality-aware design for learning from dynamic time series.

Probabilistic time series forecasting, essential in domains like healthcare and neuroscience, requires models capable of capturing uncertainty and intricate temporal dependencies. While deep learning has advanced forecasting, generating calibrated probability distributions over continuous future values remains challenging. We introduce Temporal Generative Flow Networks (Temporal GFNs), adapting Generative Flow Networks (GFNs) – a powerful framework for generating compositional objects – to this sequential prediction task. GFNs learn policies to construct objects (eg. forecast trajectories) step-by-step, sampling final objects proportionally to a reward signal. However, applying GFNs directly to continuous time series necessitates addressing their inherently discrete action spaces and ensuring differentiability. Our framework tackles this by representing time series segments as states and sequentially generating future values via quantized actions chosen by a forward policy. We introduce two key innovations: (1) An adaptive, curriculum-based quantization strategy that dynamically adjusts the number of discretization bins based on reward improvement and policy entropy, balancing precision and exploration throughout training. (2) A straight-through estimator mechanism enabling the forward policy to output both discrete (hard) samples for trajectory construction and continuous (soft) samples for stable gradient propagation. Training utilizes a trajectory balance loss objective, ensuring flow consistency, augmented by an entropy regularizer. We provide rigorous theoretical bounds on the quantization error's impact and the adaptive factor's range. We demonstrate how Temporal GFNs offer a principled way to leverage the structured generation capabilities of GFNs for probabilistic forecasting in continuous domains.

Imputation methods play a critical role in enhancing the quality of practical time-series data, which often suffer from pervasive missing values. Recently, diffusion-based generative imputation methods have demonstrated remarkable success compared to autoregressive and conventional statistical approaches. Despite their empirical success, the theoretical understanding of how well diffusion-based models capture complex spatial and temporal dependencies between the missing values and observed ones remains limited. Our work addresses this gap by investigating the statistical efficiency of conditional diffusion transformers for imputation and quantifying the uncertainty in missing values. Specifically, we derive statistical sample complexity bounds based on a novel approximation theory for conditional score functions using transformers, and, through this, construct tight confidence regions for missing values. Our findings also reveal that the efficiency and accuracy of imputation are significantly influenced by the missing patterns. Furthermore, we validate these theoretical insights through simulation and propose a mixed-masking training strategy to enhance the imputation performance.