Felix X. -F. Ye

An Yu, Ting Yu Tsai, Zhenfei Zhang et al.

Recent multimodal large language models are computationally expensive because Transformers must process a large number of visual tokens. We present \textbf{ReDiPrune}, a training-free token pruning method applied before the vision-language projector, where visual features remain rich and discriminative. Unlike post-projection pruning methods that operate on compressed representations, ReDiPrune selects informative tokens directly from vision encoder outputs, preserving fine-grained spatial and semantic cues. Each token is scored by a lightweight rule that jointly consider text-conditioned relevance and max-min diversity, ensuring the selected tokens are both query-relevant and non-redundant. ReDiPrune is fully plug-and-play, requiring no retraining or architectural modifications, and can be seamlessly inserted between the encoder and projector. Across four video and five image benchmarks, it consistently improves the accuracy-efficiency trade-off. For example, on EgoSchema with LLaVA-NeXT-Video-7B, retaining only 15\% of visual tokens yields a +2.0\% absolute accuracy gain while reducing computation by more than $6\times$ in TFLOPs. Code is available at https://github.com/UA-CVML/ReDiPrune.

Felix X. -F. Ye, Xingjie Li, An Yu et al.

Entropic optimal transport (EOT) via Sinkhorn iterations is widely used in modern machine learning, yet GPU solvers remain inefficient at scale. Tensorized implementations suffer quadratic HBM traffic from dense $n\times m$ interactions, while existing online backends avoid storing dense matrices but still rely on generic tiled map-reduce reduction kernels with limited fusion. We present \textbf{FlashSinkhorn}, an IO-aware EOT solver for squared Euclidean cost that rewrites stabilized log-domain Sinkhorn updates as row-wise LogSumExp reductions of biased dot-product scores, the same normalization as transformer attention. This enables FlashAttention-style fusion and tiling: fused Triton kernels stream tiles through on-chip SRAM and update dual potentials in a single pass, substantially reducing HBM IO per iteration while retaining linear-memory operations. We further provide streaming kernels for transport application, enabling scalable first- and second-order optimization. On A100 GPUs, FlashSinkhorn achieves up to $32\times$ forward-pass and $161\times$ end-to-end speedups over state-of-the-art online baselines on point-cloud OT, improves scalability on OT-based downstream tasks. For reproducibility, we release an open-source implementation at https://github.com/ot-triton-lab/ot_triton.

Sean Hill, Felix X. -F. Ye

Stochastic dynamical systems with slow or metastable behavior evolve, on long time scales, on an unknown low-dimensional manifold in high-dimensional ambient space. Building a reduced simulator from short-burst ambient ensembles is a long-standing problem: local-chart methods like ATLAS suffer from exponential landmark scaling and per-step reprojection, while autoencoder alternatives leave tangent-bundle geometry poorly constrained, and the errors propagate into the learned drift and diffusion. We observe that the ambient covariance~$Λ$ already encodes coordinate-invariant tangent-space information, its range spanning the tangent bundle. Using this, we construct a tangent-bundle penalty and an inverse-consistency penalty for a three-stage pipeline (chart learning, latent drift, latent diffusion) that learns a single nonlinear chart and the latent SDE. The penalties induce a function-space metric, the $ρ$-metric, strictly weaker than the Sobolev $H^1$ norm yet achieving the same chart-quality generalization rate up to logarithmic factors. For the drift, we derive an encoder-pullback target via Itô's formula on the learned encoder and prove a bias decomposition showing the standard decoder-side formula carries systematic error for any imperfect chart. Under a $W^{2,\infty}$ chart-convergence assumption, chart-level error propagates controllably to weak convergence of the ambient dynamics and to convergence of radial mean first-passage times. Experiments on four surfaces embedded in up to $201$ ambient dimensions reduce radial MFPT error by $50$--$70\%$ under rotation dynamics and achieve the lowest inter-well MFPT error on most surface--transition pairs under metastable Müller--Brown Langevin dynamics, while reducing end-to-end ambient coefficient errors by up to an order of magnitude relative to an unregularized autoencoder.

Ting Yu Tsai, An Yu, Lucy Lee et al.

Animal models, particularly rats, play a critical role in seizure research for studying epileptogenesis and treatment response. However, progress is limited by the lack of datasets with precise temporal annotations and standardized evaluation protocols. Existing animal behavior datasets often have limited accessibility, coarse labeling, and insufficient temporal localization of clinically meaningful events. To address these limitations, we introduce RatSeizure, the first publicly benchmark for fine-grained seizure behavior analysis. The dataset consists of recorded clips annotated with seizure-related action units and temporal boundaries, enabling both behavior classification and temporal localization. We further propose RaSeformer, a saliency-context Transformer for temporal action localization that highlights behavior-relevant context while suppressing redundant cues. Experiments on RatSeizure show that RaSeformer achieves strong performance and provides a competitive reference model for this challenging task. We also establish standardized dataset splits and evaluation protocols to support reproducible benchmarking.

Kevin Rojas, Yuchen Zhu, Sichen Zhu et al.

Diffusion models have demonstrated remarkable performance in generating unimodal data across various tasks, including image, video, and text generation. On the contrary, the joint generation of multimodal data through diffusion models is still in the early stages of exploration. Existing approaches heavily rely on external preprocessing protocols, such as tokenizers and variational autoencoders, to harmonize varied data representations into a unified, unimodal format. This process heavily demands the high accuracy of encoders and decoders, which can be problematic for applications with limited data. To lift this restriction, we propose a novel framework for building multimodal diffusion models on arbitrary state spaces, enabling native generation of coupled data across different modalities. By introducing an innovative decoupled noise schedule for each modality, we enable both unconditional and modality-conditioned generation within a single model simultaneously. We empirically validate our approach for text-image generation and mixed-type tabular data synthesis, demonstrating that it achieves competitive performance.

An Yu, Weiheng Lu, Jian Li et al.

Video Moment Retrieval is a task in video understanding that aims to localize a specific temporal segment in an untrimmed video based on a natural language query. Despite recent progress in moment retrieval from videos using both traditional techniques and Multimodal Large Language Models (MLLM), most existing methods still rely on coarse temporal understanding and a single visual modality, limiting performance on complex videos. To address this, we introduce \textit{S}hot-aware \textit{M}ultimodal \textit{A}udio-enhanced \textit{R}etrieval of \textit{T}emporal \textit{S}egments (SMART), an MLLM-based framework that integrates audio cues and leverages shot-level temporal structure. SMART enriches multimodal representations by combining audio and visual features while applying \textbf{Shot-aware Token Compression}, which selectively retains high-information tokens within each shot to reduce redundancy and preserve fine-grained temporal details. We also refine prompt design to better utilize audio-visual cues. Evaluations on Charades-STA and QVHighlights show that SMART achieves significant improvements over state-of-the-art methods, including a 1.61\% increase in R1@0.5 and 2.59\% gain in R1@0.7 on Charades-STA.

Felix X. -F. Ye, Sichen Yang, Mauro Maggioni

We introduce a nonlinear stochastic model reduction technique for high-dimensional stochastic dynamical systems that have a low-dimensional invariant effective manifold with slow dynamics, and high-dimensional, large fast modes. Given only access to a black box simulator from which short bursts of simulation can be obtained, we design an algorithm that outputs an estimate of the invariant manifold, a process of the effective stochastic dynamics on it, which has averaged out the fast modes, and a simulator thereof. This simulator is efficient in that it exploits of the low dimension of the invariant manifold, and takes time steps of size dependent on the regularity of the effective process, and therefore typically much larger than that of the original simulator, which had to resolve the fast modes. The algorithm and the estimation can be performed on-the-fly, leading to efficient exploration of the effective state space, without losing consistency with the underlying dynamics. This construction enables fast and efficient simulation of paths of the effective dynamics, together with estimation of crucial features and observables of such dynamics, including the stationary distribution, identification of metastable states, and residence times and transition rates between them.

Xingjie Li, Fei Lu, Felix X. -F. Ye

Efficient simulation of SDEs is essential in many applications, particularly for ergodic systems that demand efficient simulation of both short-time dynamics and large-time statistics. However, locally Lipschitz SDEs often require special treatments such as implicit schemes with small time-steps to accurately simulate the ergodic measure. We introduce a framework to construct inference-based schemes adaptive to large time-steps (ISALT) from data, achieving a reduction in time by several orders of magnitudes. The key is the statistical learning of an approximation to the infinite-dimensional discrete-time flow map. We explore the use of numerical schemes (such as the Euler-Maruyama, a hybrid RK4, and an implicit scheme) to derive informed basis functions, leading to a parameter inference problem. We introduce a scalable algorithm to estimate the parameters by least squares, and we prove the convergence of the estimators as data size increases. We test the ISALT on three non-globally Lipschitz SDEs: the 1D double-well potential, a 2D multi-scale gradient system, and the 3D stochastic Lorenz equation with degenerate noise. Numerical results show that ISALT can tolerate time-step magnitudes larger than plain numerical schemes. It reaches optimal accuracy in reproducing the invariant measure when the time-step is medium-large.

Felix X. -F. Ye, Yi-an Ma, Hong Qian

Inference in hidden Markov model has been challenging in terms of scalability due to dependencies in the observation data. In this paper, we utilize the inherent memory decay in hidden Markov models, such that the forward and backward probabilities can be carried out with subsequences, enabling efficient inference over long sequences of observations. We formulate this forward filtering process in the setting of the random dynamical system and there exist Lyapunov exponents in the i.i.d random matrices production. And the rate of the memory decay is known as $λ_2-λ_1$, the gap of the top two Lyapunov exponents almost surely. An efficient and accurate algorithm is proposed to numerically estimate the gap after the soft-max parametrization. The length of subsequences $B$ given the controlled error $ε$ is $B=\log(ε)/(λ_2-λ_1)$. We theoretically prove the validity of the algorithm and demonstrate the effectiveness with numerical examples. The method developed here can be applied to widely used algorithms, such as mini-batch stochastic gradient method. Moreover, the continuity of Lyapunov spectrum ensures the estimated $B$ could be reused for the nearby parameter during the inference.