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
Weixun Wang, XiaoXiao Xu, Wanhe An +85 authors
The Agentic Learning Ecosystem (ALE) introduces a principled infrastructure for agent development, combining post-training optimization, sandbox environments, and policy alignment to enhance long-horizon training stability and performance in real-world tasks.
Yu Gao, Haoyuan Guo, Tuyen Hoang +41 authors
Seedance 1.0 offers high-performance video generation by integrating advanced data curation, efficient architecture, post-training optimization, and model acceleration, resulting in superior quality and speed.
Jingfeng Yao, Yuda Song, Yucong Zhou +1 authors
A unified visual tokenizer pre-training framework (VTP) improves generative performance by optimizing image-text contrastive, self-supervised, and reconstruction losses, leading to better scaling properties and higher zero-shot accuracy and faster convergence.
Zixiao Wang, Yuxin Wang, Xiaorui Wang +8 authors
MetaStone-S1, a reflective generative model using a self-supervised process reward model, achieves high performance with reduced parameters and supports test time scaling.
Siyu Yuan, Zehui Chen, Zhiheng Xi +3 authors
Agent-R, an iterative self-training framework using Monte Carlo Tree Search, enhances large language models' ability to recover from errors and correct actions in real-time within interactive environments.
Chenghao Xiao, Hou Pong Chan, Hao Zhang +3 authors
Recent multimodal embedding approaches leveraging multimodal large language models (MLLMs) fine-tuned with contrastive learning (CL) have shown promising results, yet the underlying reasons behind their superiority remain underexplored. This work argues that a crucial advantage of MLLM-based approaches stems from implicit cross-modal alignment achieved during generative pretraining, where the language decoder learns to exploit multimodal signals within a shared representation space for generating unimodal outputs. Through analysis of anisotropy and kernel similarity structure, we empirically confirm that latent alignment emerges within MLLM representations, allowing CL to serve as a lightweight refinement stage. Leveraging this insight, we propose a Language-Centric Omnimodal Embedding framework, termed LCO-Emb. Extensive experiments across diverse backbones and benchmarks demonstrate its effectiveness, achieving state-of-the-art performance across modalities. Furthermore, we identify a Generation-Representation Scaling Law (GRSL), showing that the representational capabilities gained through contrastive refinement scales positively with the MLLM's generative capabilities. This suggests that improving generative abilities evolves as an effective paradigm for enhancing representation quality. We provide a theoretical explanation of GRSL, which formally links the MLLM's generative quality to the upper bound on its representation performance, and validate it on a challenging, low-resource visual-document retrieval task, showing that continual generative pretraining before CL can further enhance the potential of a model's embedding capabilities. Codes, models, and resources are available at https://github.com/LCO-Embedding/LCO-Embedding.
Aarash Feizi, Shravan Nayak, Xiangru Jian +14 authors
GroundCUA, a large-scale desktop grounding dataset, enables the development of GroundNext models that achieve state-of-the-art performance in mapping instructions to UI elements with less training data.
Nikita Kachaev, Mikhail Kolosov, Daniil Zelezetsky +2 authors
Systematic study reveals that naive action fine-tuning degrades visual representations in Vision-Language-Action models, but targeted strategies can mitigate this and improve generalization.
Chujie Zheng, Kai Dang, Bowen Yu +7 authors
The paper provides a theoretical foundation for optimizing sequence-level rewards in reinforcement learning using token-level objectives, highlighting the importance of techniques like importance sampling correction, clipping, and Routing Replay for stabilizing training, especially with large language models.
Zijian Li, Xin Guan, Bo Zhang +9 authors
WebWeaver, a dual-agent framework, addresses open-ended deep research challenges by integrating adaptive planning and focused synthesis to produce high-quality, reliable reports.
Yushen Zuo, Qi Zheng, Mingyang Wu +10 authors
4KAgent, a unified agentic super-resolution system, enhances low-resolution images to 4K using a profiling module, perception agent, and restoration agent, achieving state-of-the-art performance across various imaging domains.
Yang Zhou, Yifan Wang, Jianjun Zhou +16 authors
OmniWorld, a large-scale, multi-domain, multi-modal dataset, addresses the limitations of existing 4D world modeling datasets and benchmarks, enabling significant performance improvements in 4D reconstruction and video generation.
Yunxin Li, Xinyu Chen, Shenyuan Jiang +9 authors
We present Uni-MoE 2.0 from the Lychee family. As a fully open-source omnimodal large model (OLM), it substantially advances Lychee's Uni-MoE series in language-centric multimodal understanding, reasoning, and generating. Based on the Qwen2.5-7B dense architecture, we build Uni-MoE-2.0-Omni from scratch through three core contributions: dynamic-capacity Mixture-of-Experts (MoE) design, a progressive training strategy enhanced with an iterative reinforcement strategy, and a carefully curated multimodal data matching technique. It is capable of omnimodal understanding, as well as generating images, text, and speech. Architecturally, our new MoE framework balances computational efficiency and capability for 10 cross-modal inputs using shared, routed, and null experts, while our Omni-Modality 3D RoPE ensures spatio-temporal cross-modality alignment in the self-attention layer. For training, following cross-modal pretraining, we use a progressive supervised fine-tuning strategy that activates modality-specific experts and is enhanced by balanced data composition and an iterative GSPO-DPO method to stabilise RL training and improve reasoning. Data-wise, the base model, trained on approximately 75B tokens of open-source multimodal data, is equipped with special speech and image generation tokens, allowing it to learn these generative tasks by conditioning its outputs on linguistic cues. Extensive evaluation across 85 benchmarks demonstrates that our model achieves SOTA or highly competitive performance against leading OLMs, surpassing Qwen2.5-Omni (trained with 1.2T tokens) on over 50 of 76 benchmarks. Key strengths include video understanding (+7% avg. of 8), omnimodallity understanding (+7% avg. of 4), and audiovisual reasoning (+4%). It also advances long-form speech processing (reducing WER by 4.2%) and leads in low-level image processing and controllable generation across 5 metrics.
Shoufa Chen, Chongjian Ge, Yuqi Zhang +19 authors
Goku, a state-of-the-art family of joint image-and-video generation models using rectified flow Transformers, sets new benchmarks in text-to-image and text-to-video tasks.
Hyungjoo Chae, Sunghwan Kim, Junhee Cho +18 authors
The paper introduces Web-Shepherd, a process reward model for web navigation, which improves accuracy and cost-effectiveness in step-level trajectory assessment compared to existing multimodal large language models.
Xiaoxi Li, Guanting Dong, Jiajie Jin +5 authors
Search-o1 enhances large reasoning models with an agentic retrieval-augmented generation mechanism and a Reason-in-Documents module to improve performance on complex reasoning tasks.
Tong Zheng, Hongming Zhang, Wenhao Yu +8 authors
Parallel-R1, a reinforcement learning framework, enhances large language models' reasoning capabilities by enabling parallel thinking through a progressive curriculum, leading to significant performance improvements on math benchmarks.
Kevin Qinghong Lin, Yuhao Zheng, Hangyu Ran +5 authors
VCode introduces a benchmark for generating SVG code from images to preserve symbolic meaning, highlighting gaps in visual-centric coding and proposing VCoder to improve performance.
Sicong Leng, Jing Wang, Jiaxi Li +12 authors
Variance-Aware Sampling and large-scale CoT data improve multimodal reasoning models by stabilizing RL fine-tuning and enhancing performance on benchmarks.
Tongyi DeepResearch Team, Baixuan Li, Bo Zhang +53 authors
Tongyi DeepResearch, a large language model with agentic capabilities, achieves top performance in various deep research tasks through an end-to-end training framework and automatic data synthesis.
Yang Xiao, Mohan Jiang, Jie Sun +18 authors
LIMI demonstrates that sophisticated agentic intelligence can emerge from minimal, strategically curated demonstrations, outperforming data-intensive models on agency benchmarks.
Samuel Cahyawijaya, Holy Lovenia, Joel Ruben Antony Moniz +89 authors
SEA-VL is an open-source project aimed at creating a large, culturally relevant dataset for Southeast Asian languages to improve AI inclusivity.
Yu Sun, Xingyu Qian, Weiwen Xu +7 authors
ReasonMed, a large medical reasoning dataset, enhances the accuracy of medical question answering models by combining detailed reasoning paths with concise summaries, setting new benchmarks for model performance.
Xiaoxi Li, Wenxiang Jiao, Jiarui Jin +8 authors
DeepAgent, an end-to-end deep reasoning agent, autonomously performs thinking, tool discovery, and action execution using memory folding and reinforcement learning, outperforming baselines in various tool-use and application tasks.
Jian Hu
REINFORCE++ is an enhanced reinforcement learning algorithm that combines PPO techniques without a critic network, offering simplicity, stability, and efficiency.
Haotong Lin, Sili Chen, Junhao Liew +5 authors
Depth Anything 3 (DA3) uses a plain transformer for geometry prediction from visual inputs, achieving state-of-the-art results in camera pose estimation, any-view geometry, visual rendering, and monocular depth estimation.
Yilei Jiang, Yaozhi Zheng, Yuxuan Wan +4 authors
A modular multi-agent framework named ScreenCoder decomposes UI design-to-code translation into grounding, planning, and generation stages, achieving superior layout accuracy and code correctness through specialized agents and fine-tuned multimodal models.
Yizhou Wang, Chen Tang, Han Deng +29 authors
A scientific reasoning foundation model pre-trained on diverse scientific data supports multiple tasks and enhances cross-domain generalization and fidelity through specialized training techniques.
Thaddäus Wiedemer, Yuxuan Li, Paul Vicol +6 authors
Veo 3, a generative video model, exhibits zero-shot capabilities across various visual tasks, suggesting a trajectory towards becoming a unified, generalist vision foundation model.
Jiuhai Chen, Zhiyang Xu, Xichen Pan +10 authors
A diffusion transformer is used in a unified multimodal model framework to improve image generation while maintaining image understanding capabilities.
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