Researchers from an undisclosed institution introduced Hybrid Domain Knowledge Fusion (HDKF), a relighting-specific training framework that fuses physics, reflectance, and realism priors from syntheti…
Researchers propose HorizonRelight, a diffusion transformer-based framework that reframes long-horizon video relighting as temporally conditioned latent domain translation, using masked target-domain …
A new thesis introduces scalable tools for nested optimization in deep learning, addressing bilevel problems such as hyperparameter optimization and generative adversarial networks, where classical me…
Researchers propose JacNet, a neural network architecture that directly learns the Jacobian of an input-output function, enabling easy enforcement of structural priors such as invertibility and k-Lips…
NVIDIA introduced OmniDreams, a real-time generative world model for closed-loop autonomous vehicle simulation, built from the Cosmos diffusion model and trained on 21k hours of driving scenarios. In …
A new study from researchers demonstrates that large language models (LLMs) can match or outperform traditional hyperparameter optimization (HPO) methods such as Bayesian optimization within constrain…
Researchers propose AXPO (Agent eXplorative Policy Optimization), a reinforcement learning method that addresses the Thinking-Acting Gap in multimodal agentic reasoning by resampling tool calls for al…
SpatialClaw, a training-free framework that uses code as the action interface for spatial reasoning, achieves 59.9% average accuracy across 20 benchmarks, outperforming the recent spatial agent by +11…
Researchers introduced a novel think-answer distillation framework that masks salient reasoning prefixes to force compact vision-language models to anchor their thinking on visual evidence, reducing v…
Researchers introduced Zone of Proximal Policy Optimization (ZPPO), a method that keeps a teacher model inside prompts rather than policy gradients, improving knowledge distillation for small student …
Researchers introduced GenRecal, a general-purpose distillation framework that transfers knowledge from large vision-language models (VLMs) to smaller ones by aligning feature representations across h…
A new study identifies Relevant Visual Information Shift (RVIS) during decoding as the primary cause of performance degradation in visual token pruning methods for Multimodal Large Language Models (ML…
Researchers propose Masters (Masking Teacher and Reinforcing Student), a mask-progressive reinforcement learning distillation framework that enables compact vision-language models to learn from large …
Researchers propose Recursive Think-Answer Process (R-TAP), a method enabling large language models (LLMs) and vision-language models (VLMs) to iteratively refine their answers, outperforming single-p…
Researchers introduced RefineBench, a benchmark of 1,000 challenging problems across 11 domains, to evaluate language models' ability to refine their own responses. In self-refinement, even frontier m…
Researchers introduced Unified Reinforcement and Imitation Learning (RIL), a training algorithm that combines reinforcement learning with adversarial imitation learning to create lightweight vision-la…
Researchers propose VLsI (Verbalized Layers-to-Interactions), a new family of vision-language models in 2B and 7B sizes that uses layer-wise distillation with intermediate verbalizers to align smaller…
Researchers introduced CARV, a compute-aware variance-accounting framework that improves Monte Carlo estimation in diffusion-model pipelines, achieving 2-3x effective compute multipliers in text-to-3D…
Jonathan Lorraine, a research scientist at NVIDIA's Learning and Perception Research group, focuses on AI for science, developing agents that propose, run, and verify their own experiments, and on mea…
ROSA, a robotics foundation model serving system for robot factories, improves factory productivity by up to 12.06x over conventional dedicated serving systems, according to a paper proposing the syst…