With its significant advantages, RAE is poised to become the new default choice for training diffusion Transformers.

Given the limitations of existing fine-tuning techniques such as GRPO, GVPO has emerged as a reliable and versatile post-training paradigm.

ReCA has generalization capabilities in terms of application scenarios and system scale, and the success rate of tasks has been improved by 4.3%.

DexFlyWheel is a scalable and self-improving data generation paradigm for agile operations.

NovaFlow is able to handle rigid, articulated, and deformable objects in different robot forms.

TreeSynth demonstrates exceptional robustness and scalability in large-scale data synthesis.

GTA significantly outperforms standard SFT baselines and state-of-the-art RL methods in multiple text classification benchmarks.

ACE enables agents to improve themselves by dynamically optimizing the input context.

The rise of Vibe coding has not only changed the form of programming, but also reshaped the software development ecosystem.

Analogous to the concept of thought chains in the field of LLM, CoF is applicable to today's generative video models.

Experiments on three alignment capabilities demonstrate the effectiveness of TAE, particularly its realism, which surpasses the baseline 25.8% at a very low cost.

The emergence of the lottery hypothesis has spurred a series of methods for efficiently training neural networks.

TileLang, with its unified block and thread paradigm and transparent scheduling capabilities, meets the powerful functionality and flexibility required for the development of modern AI systems.

RPN and Fast R-CNN are combined into a single network for object detection by sharing convolutional features.

CSA aims to build systems that are not only secure, but also truly helpful.

CaT can be used at test time to improve inference time, or built into RL (CaT-RL) to improve policies.

MCP is used to connect AI assistants to where data is stored, including content repositories, business tools, and development environments.

MetaFold can handle a variety of clothing and a wide range of language commands, efficiently completing various clothing folding tasks.

ST-Raptor outperforms nine baseline models by up to 20% in answer accuracy.

SubLlME aims to achieve efficient and accurate model performance evaluation through ranking relevance prediction without the need for full-scale evaluation.

BSC-Nav constructs an allocentric cognitive map from egocentric trajectories and contextual clues, and dynamically retrieves spatial knowledge consistent with semantic goals.

Preliminary experiments show that DPCL can separate speech and achieve relatively ideal results.

The goal of dual-mode annealing is to develop a model that can grasp two different response modes: thinking mode and non-thinking mode.

The core principle of BPO is to learn adaptive policies by explicitly comparing the utilities of thinking and non-thinking paths under the same input query.

With its significant advantages, RAE is poised to become the new default choice for training diffusion Transformers.

Given the limitations of existing fine-tuning techniques such as GRPO, GVPO has emerged as a reliable and versatile post-training paradigm.

ReCA has generalization capabilities in terms of application scenarios and system scale, and the success rate of tasks has been improved by 4.3%.

DexFlyWheel is a scalable and self-improving data generation paradigm for agile operations.

NovaFlow is able to handle rigid, articulated, and deformable objects in different robot forms.

TreeSynth demonstrates exceptional robustness and scalability in large-scale data synthesis.

GTA significantly outperforms standard SFT baselines and state-of-the-art RL methods in multiple text classification benchmarks.

ACE enables agents to improve themselves by dynamically optimizing the input context.

The rise of Vibe coding has not only changed the form of programming, but also reshaped the software development ecosystem.

Analogous to the concept of thought chains in the field of LLM, CoF is applicable to today's generative video models.

Experiments on three alignment capabilities demonstrate the effectiveness of TAE, particularly its realism, which surpasses the baseline 25.8% at a very low cost.

The emergence of the lottery hypothesis has spurred a series of methods for efficiently training neural networks.

TileLang, with its unified block and thread paradigm and transparent scheduling capabilities, meets the powerful functionality and flexibility required for the development of modern AI systems.

RPN and Fast R-CNN are combined into a single network for object detection by sharing convolutional features.

CSA aims to build systems that are not only secure, but also truly helpful.

CaT can be used at test time to improve inference time, or built into RL (CaT-RL) to improve policies.

MCP is used to connect AI assistants to where data is stored, including content repositories, business tools, and development environments.

MetaFold can handle a variety of clothing and a wide range of language commands, efficiently completing various clothing folding tasks.

ST-Raptor outperforms nine baseline models by up to 20% in answer accuracy.

SubLlME aims to achieve efficient and accurate model performance evaluation through ranking relevance prediction without the need for full-scale evaluation.

BSC-Nav constructs an allocentric cognitive map from egocentric trajectories and contextual clues, and dynamically retrieves spatial knowledge consistent with semantic goals.

Preliminary experiments show that DPCL can separate speech and achieve relatively ideal results.

The goal of dual-mode annealing is to develop a model that can grasp two different response modes: thinking mode and non-thinking mode.

The core principle of BPO is to learn adaptive policies by explicitly comparing the utilities of thinking and non-thinking paths under the same input query.