This paper introduces EUGens: Efficient, Unified and General Dense Layers, capable of accurately approximating general feedforward fully connected layers, but of linear rather than quadratic time complexity.

We propose OCTDiff, a bridged diffusion model with Adaptive Noise Aggregation and Multi-Scale Cross-Attention, trained with quality-guided loss. It achieves SOTA enhancement on portable OCT images and shows potential for broader clinical use.

A training-free diffusion-based sequential monte carlo method for sampling from unnormalized distributions

A generative model imputes missing outcomes to drive Thompson Sampling decisions, yielding a flexible algorithm with regret tied to offline prediction quality.

This paper proposes a novel deep reinforcement learning algorithm robust to unobserved confounding bias in observed data over complex and high-dimensional domains.

This paper is a local misspecificiation analysis of three data-driven stochastic optimization methods.

We study the graphical characterization of structural causal bandits framework under Markov equivalence class.

We present POSSM, a novel architecture that combines input cross-attention with a recurrent state-space model, achieving competitive accuracy, fast inference, and efficient generalization for real-time neural decoding applications.

Undocumented versions of Meta-World have clouded algorithmic performance. This work strives to disambiguate Meta-World results from the literature, while also providing insights into benchmark design.

We develop an optimizer ASGO that can provably exploit the low-rank gradients and block-wise diagonal Hessians in training.

We imrove sample complexity of single time scale actor critic to $O(\epsilon^{-3})$ from $O(\epsilon^{-4})$ for obtaining $\epsilon$-close global optimal policy.

A multi-scale, multi-fidelity Bayesian Optimization (BO) approach where {data mixtures, model scale, training steps} are adaptively selected, achieving >2.6x speedups compared to multi-fidelity BO and random search baselines.

We refine SVG generation using online reinforcement learning with image reconstruction, semantic, and code-level rewards, boosting accuracy, efficiency, and interpretability.

This work proposes a Bayesian framework with variational inference that adapts the prior and posterior to covariate shifts, improving uncertainty estimates by capturing predictive reliability rather than just input dissimilarity.

We present CausalDynamics, a benchmark for advancing the structural discovery of dynamical causal models consisting of true causal graphs derived from thousands of increasingly complex coupled differential equations.

We develop a score-based variational inference method that learns a product-of-t-experts model via a Feynman-identity latent-variable formulation, reducing inference to a sequence of convex quadratic programs with provable convergence.

We bridge the gap between causal image editing and large-scale text-to-image generation.

We isolate “reasoning embeddings” from LLMs using a residual method and show they uniquely predict brain activity, revealing distinct neural correlates of reasoning beyond shallow linguistic features.

We fit scaling laws for large language models with varying width-to-depth ratios and parameter counts.

We apply context-aware deep neural networks to predict functional connectivity from gene expression in humans.

If and when integrating studies with different outcome measures improves efficiency, showing that gains hinge on strong linking assumptions but carry bias risks, while weaker ones may offer finite-sample benefits.

We jointly evaluate VLMs and diffusion models by seeing whether VLMs can evaluate diffusion model failure modes.

Framework for generating data-driven realistic security game instances and relevant benchmarks

We investiage the use of autoregressive models for exchangeable sequences in decision-making, showing multi-step inference improves decision making and standard causal architectures outperform existing custom ones.

Our paper shows that ensemble learning via majority voting achieves exponentially faster risk decay, improving base learners with slow rates.

We train reward models to encourage fairer step-by-step reasoning in LLMs, reducing bias on high-stakes decision-making tasks

We meta-learn a transformer based in-context learning fMRI visual cortex encoder which can adapt to new human subjects without any fine-tuning

We built a transformer-based model to predict brain activity across the whole brain from visual input, then use this model to label the categoriacal selectivity of areas beyond the visual cortex to better understand higher-order visual processing.

Multimodal satellite imagery dataset for natural disaster monitoring

We present NuCLR, a self-supervised framework that learns high-quality, population-aware neuron-level embeddings directly from spike train data using a spatio-temporal transformer and tailored contrastive loss.

A novel semi-supervised learning paradigm that unifies view-wise co-training, meta-learned supervision, and adversarial perturbation through a structured triadic game.

We develop the Less Greedy Equivalence Search algorithm for learning causal structure from observational and interventional data with prior knowledge.

A two-system RNN achieves continual learning by using a 'what' system to infer compositional structures in tasks and a 'how' system to compose low-rank components, demonstrating transfer and compositional generalization.

This paper establishes the mathematical foundation of value decomposition in MARL.

NeurIPT, advancing scalable and generalizable EEG decoding across diverse BCI tasks.

HoliTom introduces a training-free holistic outer-inner token merging framework for video LLMs, significantly accelerating inference with negligible performance degradation.

We show how to efficiently verify approximate optimaility of smooth policies and strategies in bandits and games

We develop a framework for building causally interpretable models where counterfactual queries can be consistently evaluated from a model and observational data.

We provide tight asymptotic bounds for the expected top-ℓ order statistics under minimal moment assumptions, revealing a sharp separation between the maximum and second-largest values.

We designed and tested an autoencoder model for stitching together multi-area multi-animal neuronal recording datasets and inpainting neural dynamics from unobserved brain areas in each experiment.

We present a transformer brain encoder that achieves state of the art performance, by leveraging brain-region to image-feature cross-attention mechanism, efficiently mapping high-dimensional retinotopic features to brain areas.

Statistical prediction may be sufficient to drive the emergence of internal causal models and causal inference capacities in deep neural networks.

We study estimation correlation from bivariate data separated vertically, i.e., component-wise into two servers with privacy constraints.

A multi-metric Common Task Framework that standardizes characterization and comparison of modeling methods for dynamic systems.