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
Zhe Ren, Yimeng Chen, Dandan Guo +9 authors
Modern self-improving agents are surveyed as adaptive systems that convert experience into capability gains through updates to foundation models and operational scaffolds.
Xiaohan Zhang, Yuqing Wen, Junlin Chen +9 authors
MultiRef-Compass introduces a benchmark and evaluation protocol for generating synchronized audio-video content from multiple references and text instructions, assessing quality, consistency, and instruction adherence.
Minh-Quan Le, Armand Comas, Alexandros Lattas +7 authors
A coupled Markov jump process with cross-modal attention and remasking enables a training-free single-pass sampler for joint multimodal generation that improves with more denoising steps.
Lingkai Kong, Zijian Wu, Yuzhe Gu +10 authors
AdvancedMathBench evaluates advanced mathematical reasoning through proof generation and verification tasks, revealing substantial performance gaps in frontier models.
Donghyun Lee, Yuhang Li, Ruokai Yin +1 authors
KronQ improves post-training quantization by incorporating gradient covariance via Kronecker-factored Hessian approximations, enabling bidirectional incoherence processing and gradient-aware mixed-precision allocation.
Zhongwei Ren, Yunchao Wei, Yao Zhao +5 authors
UniVR learns complex visual reasoning, physical dynamics, and long-term planning from pure visual demonstrations using a reinforcement learning approach with global and step-level rewards, evaluated on a new large-scale benchmark.
Haotian Liang, Mingkang Chen, Yufei Huang +27 authors
Hy-Embodied-RxBrain is a multimodal foundation model that unifies language reasoning and visual imagination into sequential embodied planning, enabling joint subgoal planning and world-state prediction with reduced reliance on large-scale robot action data.
Gabrielle Kaili-May Liu, Areeb Gani, Jacqueline Lu +3 authors
This review surveys metacognitive capabilities in large language models, covering evaluation benchmarks, enhancement techniques, and future research directions.
GigaWorld Team, Angen Ye, Angyuan Ma +26 authors
GigaWorld-Policy-0.5 improves robot control efficiency by using action-centered world modeling with specialized transformer experts and automated hyperparameter search.
Lianghua Huang, Zhi-Fan Wu, Yupeng Shi +24 authors
Wan-Streamer v0.3 treats video as a persistent world plus a temporal event stream to enable real-time pretraining and full-duplex audio-visual interaction via vision-language-action-style multimodal mapping.
Haotian Lin, Silin Chen, Xiaodong Gu +8 authors
ACQUIRE improves automated software repair by using targeted question-answering to build explicit repository knowledge before generating patches.
Nikita Starodubcev, Ilia Sudakov, Ilya Drobyshevskiy +2 authors
Register tokens improve diffusion transformers by cleaning high-noise feature maps, and a proposed register guidance technique enhances visual coherence.
Yunfan Jiang, Yevgen Chebotar, Ruijie Zheng +8 authors
RoboTTT extends robot policies to 8K-step visuomotor contexts via test-time training with fast recurrent weights, enabling long-horizon manipulation and in-context imitation without added inference latency.
Simon Rouard, Michael Krause, Axel Roebel +2 authors
A multi-instrument transcription model combines synthetic pre-training, real-audio fine-tuning, reinforcement learning, and instrument conditioning to transcribe diverse real-world music.
Rui Wang, Hongru Wang, Yi Chen +4 authors
On-policy distillation acts as an exploration catalyst guided by token-level signals, but suffers from student-teacher mismatch and length exploitation that can be mitigated by lightweight signal regulations.
Hongbo Wang, Huaibo Huang, Jie Cao +3 authors
Hallo4D reduces spatial and temporal inconsistencies in 3D and 4D generation by using multimodal language models to detect errors and guide consensus-based image optimization without retraining.
Shengyuan Liu, Jia-Xuan Jiang, Boyun Zheng +8 authors
AMID is an autonomous multi-agent framework that automates medical imaging model development through data-conditioned planning and verification-guided optimization, yielding high-performing and auditable results across diverse tasks.
Zanyi Wang, Xin Lin, Haodong Li +2 authors
Pretrained diffusion transformers can be adapted for dense prediction tasks by mapping tokens to task-native outputs instead of generating RGB images, achieving state-of-the-art results with minimal additional parameters.
Ye Yuan, Weien Li, Rui Song +19 authors
This work proposes a unified framework for discrete diffusion models based on discrete state space construction, unifying existing methods and highlighting design trade-offs.
Samuel Yeh, Yiwen Zhu, Shaleen Deep +1 authors
OAT uses one-class learning with neural controlled differential equations to identify failure steps in agent trajectories by modeling successful dynamics, requiring no step-level failure annotations.
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