Explore · 每周精选
发现最受关注的研究论文,追踪研究趋势,订阅感兴趣的期刊与关键词。
606 篇论文 · 按点赞排序
Zeyinzi Jiang, Zhen Han, Chaojie Mao +3 authors
VACE, an all-in-one framework for video creation and editing, integrates multiple tasks within a unified model using a Video Condition Unit and Context Adapter for flexible and consistent video synthesis.
Yuhao Wu, Yushi Bai, Zhiqiang Hu +2 authors
An incentivization-based reinforcement learning approach is used to develop a large language model capable of generating ultra-long, high-quality text without the need for synthetic data or supervised fine-tuning.
Hengguang Zhou, Xirui Li, Ruochen Wang +3 authors
A non-SFT model replicated emergent reasoning characteristics for multimodal tasks using reinforcement learning, achieving higher accuracy than base and SFT models on CVBench.
Siyin Wang, Zhaoye Fei, Qinyuan Cheng +4 authors
Dual Preference Optimization (D$^2$PO) enhances embodied task planning in large vision-language models by jointly optimizing state prediction and action selection using preference learning and tree search for efficient trajectory collection.
Edward Yeo, Yuxuan Tong, Morry Niu +2 authors
Investigation into long chains-of-thought reasoning in large language models reveals the critical role of training compute, reward shaping, and verifiable reward signals in enabling and measuring this capability.
Chengqi Lyu, Songyang Gao, Yuzhe Gu +14 authors
A novel reinforcement learning framework, OREAL, using outcome rewards achieves high accuracy in mathematical reasoning tasks with smaller models compared to larger models trained using distillation.
Guoqing Ma, Haoyang Huang, Kun Yan +112 authors
Step-Video-T2V is a state-of-the-art text-to-video model with 30B parameters, utilizing a deep compression Variational Autoencoder and a DiT with 3D full attention for high-quality video generation.
Kashun Shum, Yuzhen Huang, Hongjian Zou +5 authors
A lightweight data selection method called PreSelect effectively reduces computational requirements while enhancing language model performance by focusing on data predictive of downstream abilities.
Etash Guha, Ryan Marten, Sedrick Keh +47 authors
The OpenThoughts project created open-source datasets leading to reasoning models that match or exceed state-of-the-art benchmarks in math, code, and science.
Rui Xie, Yinhong Liu, Penghao Zhou +7 authors
A novel approach using text-to-video models with local information enhancement and dynamic frequency loss improves spatio-temporal quality in real-world video super-resolution.
Mingjie Pan, Jiyao Zhang, Tianshu Wu +3 authors
A novel object-centric representation bridges the gap between Vision-Language Models' high-level reasoning and the low-level precision needed for robotic manipulation tasks, enabling zero-shot generalization across diverse tasks.
Zorik Gekhman, Eyal Ben David, Hadas Orgad +5 authors
LLMs encode more internal factual knowledge than they express externally, with some knowledge so deeply hidden that it is never generated, despite repeated sampling.
Tong Wu, Shuai Yang, Ryan Po +4 authors
A new framework enhances video world models' long-term consistency by integrating a geometry-grounded long-term spatial memory mechanism.
Alejandro Lozano, Min Woo Sun, James Burgess +13 authors
BIOMEDICA is a scalable framework that creates a large, publicly accessible dataset of image-text pairs from PubMed Central, enabling state-of-the-art performance across various biomedical tasks with efficient CLIP-style models.
Siyuan Huang, Liliang Chen, Pengfei Zhou +7 authors
EnerVerse enhances robotic manipulation by generating future space using convolutional and bidirectional attention, sparse memory, and generated multi-camera data.
Xinyu Yang, Yuwei An, Hongyi Liu +2 authors
Multiverse, a generative model enabling native parallel text generation, achieves performance comparable to autoregressive models through a three-stage MapReduce paradigm and efficient attention mechanisms.
Qi Sun, Edoardo Cetin, Yujin Tang
A self-adaptive framework for large language models uses reinforcement learning to dynamically adjust task-specific components during inference, enhancing adaptability and performance with efficiency.
Jiaqi Li, Junshu Tang, Zhiyong Xu +6 authors
Hunyuan-GameCraft is a novel framework for high-dynamic interactive video generation in game environments that addresses limitations in dynamics, generality, and efficiency through unified input representation, hybrid history-conditioned training, and model distillation.
Chenlong Wang, Yuanning Feng, Dongping Chen +3 authors
NoWait suppresses explicit self-reflection tokens during inference to enhance efficiency in multimodal reasoning without reducing model utility.
Lianghong Guo, Yanlin Wang, Caihua Li +6 authors
A pipeline named SWE-Factory automates the creation and validation of GitHub issue resolution datasets for training and evaluating Large Language Models, using SWE-Builder for environment setup, exit-code-based grading, and automated fail2pass validation.
Yichen He, Guanhua Huang, Peiyuan Feng +4 authors
A quality Paper Search agent, PaSa, uses reinforcement learning on synthetic and real academic queries, outperforming Google and GPT-4-based systems.
Jian Zhang, Zhiyuan Wang, Zhangqi Wang +6 authors
A Multi-Agent framework with feedback mechanisms and Socratic guidance addresses multi-modal scientific problems and enhances reasoning and reflection, outperforming current models.
Qian Chen, Yafeng Chen, Yanni Chen +33 authors
MinMo, a multimodal large language model, integrates speech and text processing to achieve state-of-the-art performance in voice comprehension and generation, while enabling full-duplex conversation and instruction-following capabilities.
Jaehyeong Jo, Sung Ju Hwang
A continuous diffusion model for language modeling that leverages the geometry of discrete distributions outperforms existing discrete models and matches autoregressive models in performance.
Ziming Liu, Yifan Yang, Chengruidong Zhang +4 authors
RAS, a novel sampling strategy for diffusion transformers, dynamically adjusts sampling ratios based on regions of focus, achieving speedups in diffusion models with minimal quality loss.
Dongwei Jiang, Alvin Zhang, Andrew Wang +2 authors
LLMs show resistance to feedback, termed feedback friction, even under ideal conditions, and sampling-based strategies only partially mitigate this issue.
Rongyao Fang, Chengqi Duan, Kun Wang +9 authors
A new reasoning-driven paradigm, GoT, improves text-to-image generation and editing through language understanding and reasoning chains, integrating Qwen2.5-VL and an enhanced diffusion model with a Semantic-Spatial Guidance Module.
Juraj Gottweis, Wei-Hung Weng, Alexander Daryin +31 authors
A multi-agent AI system named AI co-scientist aids in scientific discovery by generating and validating novel hypotheses across biomedical areas, demonstrating potential improvements in drug repurposing, target discovery, and bacterial evolution understanding.
Yuanpeng Tu, Hao Luo, Xi Chen +3 authors
VideoAnydoor is a zero-shot framework for object insertion in videos that preserves detail and controls motion using a text-to-video model, ID extractor, box sequence, pixel warper, and diffusion U-Net.
Zibo Zhao, Zeqiang Lai, Qingxiang Lin +68 authors
Hunyuan3D 2.0, an advanced 3D synthesis system, uses flow-based diffusion transformers for shape generation and diffusion priors for texture synthesis, outperforming existing models in geometry, condition alignment, and texture quality.
北京市昌平区探索星信息技术及软件开发工作室
京ICP备2026059466号