Camera conditioning as relative projective transformation in multi-view transformer

We present JetLM, a new family of LMs, which matches leading full-attention models while significantly improving generation throughput.

We introduce a segmentation-centric backbone architecture built around a spatial grouping layer, capable of generating high-quality segmentation masks through feature extraction.

Theory of Fair Representations, and a computationally tractable fairness performance Pareto front algorithm.

Efficiently generate text avoiding constraint violations, with less distortion of LLM output distribution.

DrivAerStar introduces a validated dataset of 12,000 industrial-grade CFD simulations for automotive aerodynamics, enabling machine learning models to achieve engineering accuracy while drastically reducing computational costs.

EDELINE combines diffusion models with state space models to create a world model for reinforcement learning that overcomes memory limitations in previous approaches.

The paper introduces a pruning and distillation method for hybrid LLMs, compressing Nemotron-H 8B to 4B with better accuracy and ~2× faster inference, advancing the efficiency-accuracy trade-off.

We propose MASP, a meta-learned similarity-based regularization for RL with macro-actions. MASP improves exploration, credit assignment, and transfer across tasks, outperforming Rainbow DQN in challenging benchmarks.

A reliable driving world model that can be accurately steered with various actions with realistic visual simulation.

GradMetaNet is a neural architecture that efficiently processes gradients of other networks by exploiting their symmetries and rank-1 decomposition structure, enabling better learned optimizers, model editing, and loss curvature estimation

We present the first systematic study of lossy latency–quality trade-offs in LLM agents, introducing HFTBench and StreetFighter benchmarks, and proposing an adaptive mixed-precision framework for real-world latency-sensitive tasks.

We propose a novel RL-based MLLM post-training framework named RePIC for the personalized image captioning task. Our method significantly outperforms SFT-based baselines on multi-concept personalized image captioning.

We release an open human-annotated preference dataset with 40 thousand samples spanning General, STEM, Code and Multilingual Samples, which can be used to train SOTA Reward Models on RM-Bench and JudgeBench

We present SonoGym, a scalable simulation platform for challenging robotic ultrasound tasks that allows training reinforcement learning and imitation learning agents.

We demonstrate for the first time, fully quantized training of a 7B LLM using FP4 format.

An architectural optimization method that accelerate inference efficiency by replacing sequential computation in LLM with parallel computation while preserving accuracy.

Image-based Gaussian Splatting

Inference-time hyper-scaling uses key–value cache compression with Delayed Memory Sparsification (DMS) to boost Transformer LLM reasoning accuracy for equivalent compute or memory costs.

The first global synthetic dataset for physics-ML seismic wavefield modeling and full-waveform inversion

We develop flow map methods for state-of-the-art few-step generation, generalizing flow, diffusion, and consistency models.

Feed-forward dynamic 3DGS scene reconstruction from videos.

a sparse attention with $\mathcal O(n \log n)$ complexity for long video generation

We propose a method to enable elastic inference for pretrained ViTs via structured pruning. We do not require labels, generalize to models without a classification head and do not need re-training. We improve over the existing state of the art.

A fully-open large audio-language model with SOTA performance across 20+ benchmarks.

scaling math and code rl

We propose BlurDM, a diffusion-based framework that integrates the physics of motion blur formation into the diffusion processes to improve image deblurring.

Nemotron-CLIMB automates data mixture optimization for pre-training, improving domain adaptation and outperforming Llama-3.2-1B by 2.0% on general reasoning.

We propose GPAS, a simple method that scales activations without scaling gradients to accelerate pretraining convergence of LLMs.

We introduce TrajICL, a novel in-context learning (ICL) framework for pedestrian trajectory forecasting that enables rapid adaptation without costly fine-tuning.

We present a general framework to solve inverse PDE problems through function space diffusion models in a plug-and-play way.

We present an algorithm for test-time scaling of SDE-based diffusion models by searching for noise trajectories which optimize arbitrary rewards, empirically matching/exceeding MCTS performance.

Polar Sparsity scales contextual sparsity to large batches by exploiting stable attention head sparsity and using efficient GPU kernels, achieving up to 2.2× speedups with minimal accuracy loss.

Long-RL enables RL on hour-long videos on a single A100; LongVILA-R1-7B supports 8,192 frames and scores 65.1%/71.1% on VideoMME.

A sequence-level EDM-derived diffusion model using progressively increasing noise levels is designed for probabilistic forecasting of complex dynamics like weather.

using polynomials to build spatio-temporal networks

We introduce two complementary enhancements to Stein Variational Ergodic Search, yielding richer, more efficient exploration trajectories across diverse robotic benchmarks.

We present the first comprehensive benchmark for long-context vision langauge models.

We present the largest pancreatic tumor dataset and will make it public benchmarks.

EMLoC is a memory-efficient fine-tuning framework that uses a lightweight emulator and LoRA correction to enable fine-tuning within an inference-level memory budget.

We propose a method to speedup video diffusion generation through efficient attention.

We generalize attention and neighborhood attention to the two-dimensional sphere in an equivariant manner.

We develop trust region methods for stochastic optimal control to improve sampling from unnormalized densities, transition path sampling, and diffusion model finetuning.

SutureBot: A benchmark and dataset for evaluating goal-conditioned VLAs on precision suturing.