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。
50 篇论文 · 按点赞排序
Shuai Yuan, Yantai Yang, Xiaotian Yang +4 authors
InfiniteVGGT enables continuous 3D visual geometry understanding through a causal transformer with adaptive memory management, outperforming existing streaming methods in long-term stability while introducing a new benchmark for extended evaluation.
Shikun Sun, Liao Qu, Huichao Zhang +8 authors
Visual autoregressive models face training instability due to asynchronous policy conflicts, which are addressed through a novel framework enhancing group relative policy optimization with intermediate rewards, dynamic time-step reweighting, and mask propagation algorithms.
Jiewen Chan, Zhenjun Zhao, Yu-Lun Liu
Reconstructing dynamic 3D scenes from monocular videos requires simultaneously capturing high-frequency appearance details and temporally continuous motion. Existing methods using single Gaussian primitives are limited by their low-pass filtering nature, while standard Gabor functions introduce energy instability. Moreover, lack of temporal continuity constraints often leads to motion artifacts during interpolation. We propose AdaGaR, a unified framework addressing both frequency adaptivity and temporal continuity in explicit dynamic scene modeling. We introduce Adaptive Gabor Representation, extending Gaussians through learnable frequency weights and adaptive energy compensation to balance detail capture and stability. For temporal continuity, we employ Cubic Hermite Splines with Temporal Curvature Regularization to ensure smooth motion evolution. An Adaptive Initialization mechanism combining depth estimation, point tracking, and foreground masks establishes stable point cloud distributions in early training. Experiments on Tap-Vid DAVIS demonstrate state-of-the-art performance (PSNR 35.49, SSIM 0.9433, LPIPS 0.0723) and strong generalization across frame interpolation, depth consistency, video editing, and stereo view synthesis. Project page: https://jiewenchan.github.io/AdaGaR/
Chengsong Huang, Tong Zheng, Langlin Huang +3 authors
RelayLLM enables efficient collaborative reasoning between small and large language models through token-level dynamic invocation, achieving high accuracy with minimal computational overhead.
Junyi Chen, Tong He, Zhoujie Fu +3 authors
VINO is a unified visual generator that uses a shared diffusion backbone with multimodal inputs to perform image and video generation and editing tasks.
Haoran He, Yuxiao Ye, Jie Liu +7 authors
Online reinforcement learning for diffusion model fine-tuning suffers from reward hacking due to proxy reward mismatches, which GARDO addresses through selective regularization, adaptive reference updates, and diversity-aware reward amplification.
Mingjie Pan, Siyuan Feng, Qinglin Zhang +9 authors
A scalable online post-training system enables real-world robot policy adaptation through distributed, multi-task learning that maintains generality while improving task proficiency.
Zefang Zong, Dingwei Chen, Yang Li +6 authors
AT²PO is a unified framework for multi-turn agentic reinforcement learning that improves exploration diversity, credit assignment, and policy optimization through tree search and turn-level learning objectives.
Falcon LLM Team, Iheb Chaabane, Puneesh Khanna +8 authors
Falcon-H1R is a 7B-parameter language model that achieves competitive reasoning performance through efficient training strategies and architectural design, enabling scalable reasoning capabilities in compact models.
Chaofan Tao, Jierun Chen, Yuxin Jiang +11 authors
SWE-Lego achieves state-of-the-art performance in software engineering task resolution through a lightweight supervised fine-tuning approach combined with a curated dataset and refined training procedures.
Boyang Wang, Haoran Zhang, Shujie Zhang +8 authors
Visual identity prompting enhances manipulation data augmentation for robot policies by providing explicit visual guidance to diffusion models, improving policy performance in both simulation and real-world settings.
Mengtian Li, Jinshu Chen, Songtao Zhao +3 authors
DreamStyle is a unified video stylization framework that supports multiple style conditions while addressing style inconsistency and temporal flicker through a specialized data curation pipeline and LoRA training approach.
Mengqi He, Xinyu Tian, Xin Shen +4 authors
Selective adversarial attacks targeting high-entropy tokens in vision-language models achieve significant semantic degradation with reduced budgets and demonstrate transferable vulnerabilities across different architectures.
Runyang You, Hongru Cai, Caiqi Zhang +5 authors
Large language models face limitations in evaluating complex, multi-step tasks, prompting the development of agent-based evaluation systems that utilize planning, tool-augmented verification, and multi-agent collaboration for more robust assessments.
Shuhang Chen, Yunqiu Xu, Junjie Xie +7 authors
Visual mathematical problem solving remains challenging for multimodal large language models, prompting the development of CogFlow, a cognitive-inspired three-stage framework that enhances perception, internalization, and reasoning through synergistic rewards and visual-gated policy optimization.
Max Ruiz Luyten, Mihaela van der Schaar
Large language model training methods that optimize for correctness can cause reasoning path diversity collapse, but a new variational framework provides principled solutions to maintain both accuracy and creativity.
Sixiao Zheng, Minghao Yin, Wenbo Hu +3 authors
VerseCrafter is a 4D-aware video world model that enables unified control over camera and object dynamics through 4D geometric control representation and video diffusion models.
Ming Zhang, Kexin Tan, Yueyuan Huang +20 authors
An LLM-powered agentic system for transparent, evidence-based novelty assessment in peer review that retrieves and analyzes prior work through semantic search and hierarchical taxonomy construction.
Hao Bai, Alexey Taymanov, Tong Zhang +2 authors
WebGym presents a large-scale open-source environment for training visual web agents using reinforcement learning with high-throughput asynchronous sampling, achieving superior performance on unseen websites compared to proprietary models.
Yan Wang, Yitao Xu, Nanhan Shen +3 authors
Research challenges the assumption of domain specialization in Mixture of Experts models by identifying a persistent central committee of experts that dominates routing behavior across different domains and architectures.
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