MVP achieves single-step action generation with both high expressive power and extremely fast computation by modeling the average velocity field.

WorldGen is capable of creating geometrically unified, visually rich, and highly efficient real-time rendering worlds.

Model Souping can generate a better model by averaging the weights of multiple fine-tunings.

By leveraging GPU parallelism to efficiently expand the decoding tree, fast and scalable optimization of the inference path is achieved.

Skills are reusable capability modules that encapsulate knowledge and processes, enabling AI to transform from general-purpose models into specialized intelligent agents.

SoCE is a model optimization paradigm based on an automatic category-aware expert selection mechanism and combined with multiple benchmark tasks.

DePass is used to interpret the Transformer model by decomposing the forward pass.

A file format for storing medical imaging data

iSeal achieves a 100% fingerprint success rate (FSR) against more than 10 attacks on 12 LLMs.

It effectively solves the key challenges in LVLM secure alignment.

VLM can achieve cross-modal understanding, reasoning, and generation tasks by aligning and fusing image and text information.

VLA can generate robot movements directly based on visual images and verbal commands.

The NSG statistic quantifies the ratio of spatial probability gradient to temporal density change.

Mem-I has achieved significant improvements over existing memory-enhanced agent baselines in multiple benchmark tests.

SSP demonstrates the potential of self-game theory as a scalable and data-efficient training paradigm for agent LLM.

CudaForge is a simple, effective, and low-cost multi-agent workflow for CUDA kernel generation and optimization.

FractalForensics exhibits good robustness and vulnerability to common image processing operations and Deepfake operations.

ScaleNet is a novel approach that extends pre-trained Visual Transformer (ViT) through weight sharing.

FlashMoBA makes the theoretically optimal block size practical, achieving up to 14.7x speedup on GPUs.

CoT Hijacking is a novel jailbreak attack method in which benign reasoning systematically weakens the rejection behavior.

InstanceAssemble enables high-quality and controllable image generation under multimodal conditions.

Layout-to-Image provides a flexible control mechanism for image generation.

HiPO is used for adaptive LLM inference, mainly including hybrid data construction and hybrid reinforcement learning.

As a novel semantic-aware framework, it is used to reconstruct 3D models from sparse views.

MVP achieves single-step action generation with both high expressive power and extremely fast computation by modeling the average velocity field.

WorldGen is capable of creating geometrically unified, visually rich, and highly efficient real-time rendering worlds.

Model Souping can generate a better model by averaging the weights of multiple fine-tunings.

By leveraging GPU parallelism to efficiently expand the decoding tree, fast and scalable optimization of the inference path is achieved.

Skills are reusable capability modules that encapsulate knowledge and processes, enabling AI to transform from general-purpose models into specialized intelligent agents.

SoCE is a model optimization paradigm based on an automatic category-aware expert selection mechanism and combined with multiple benchmark tasks.

DePass is used to interpret the Transformer model by decomposing the forward pass.

A file format for storing medical imaging data

iSeal achieves a 100% fingerprint success rate (FSR) against more than 10 attacks on 12 LLMs.

It effectively solves the key challenges in LVLM secure alignment.

VLM can achieve cross-modal understanding, reasoning, and generation tasks by aligning and fusing image and text information.

VLA can generate robot movements directly based on visual images and verbal commands.

The NSG statistic quantifies the ratio of spatial probability gradient to temporal density change.

Mem-I has achieved significant improvements over existing memory-enhanced agent baselines in multiple benchmark tests.

SSP demonstrates the potential of self-game theory as a scalable and data-efficient training paradigm for agent LLM.

CudaForge is a simple, effective, and low-cost multi-agent workflow for CUDA kernel generation and optimization.

FractalForensics exhibits good robustness and vulnerability to common image processing operations and Deepfake operations.

ScaleNet is a novel approach that extends pre-trained Visual Transformer (ViT) through weight sharing.

FlashMoBA makes the theoretically optimal block size practical, achieving up to 14.7x speedup on GPUs.

CoT Hijacking is a novel jailbreak attack method in which benign reasoning systematically weakens the rejection behavior.

InstanceAssemble enables high-quality and controllable image generation under multimodal conditions.

Layout-to-Image provides a flexible control mechanism for image generation.

HiPO is used for adaptive LLM inference, mainly including hybrid data construction and hybrid reinforcement learning.

As a novel semantic-aware framework, it is used to reconstruct 3D models from sparse views.