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。
Zhi Zhou, Yuhao Tan, Zenan Li +4 authors
A theoretical framework for sampling-based test-time scaling in large language models reveals limitations in self-consistency and perplexity, and introduces RPC to improve reasoning performance and reduce sampling costs.
50 篇论文 · 按点赞排序
Yonghao Zhuang, Junda Chen, Bo Pang +6 authors
CAD, a technique for long-context large language model training, improves throughput and balance by decoupling and distributing core attention computations.
Jizhan Fang, Xinle Deng, Haoming Xu +9 authors
LightMem, a memory system inspired by human memory, enhances LLMs by efficiently managing historical interaction information, improving accuracy and reducing computational costs.
Ling Team, Bin Han, Caizhi Tang +25 authors
The Ring-linear model series, including Ring-mini-linear-2.0 and Ring-flash-linear-2.0, uses a hybrid architecture combining linear and softmax attention to reduce inference costs and improve training efficiency.
Shaolei Zhang, Ju Fan, Meihao Fan +2 authors
DeepAnalyze-8B, an agentic LLM, autonomously completes the data science pipeline from raw data to research reports using curriculum-based training and data-grounded trajectory synthesis.
Haoran Wei, Yaofeng Sun, Yukun Li
DeepSeek-OCR uses optical 2D mapping to compress long contexts, achieving high OCR precision with reduced vision tokens and demonstrating practical value in document processing.
Hanrong Ye, Chao-Han Huck Yang, Arushi Goel +29 authors
OmniVinci, an open-source omni-modal LLM, enhances cross-modal understanding and performance across audio, vision, and robotics applications with innovative architecture and efficient data curation.
Zhiheng Xi, Xin Guo, Yang Nan +18 authors
BAlanced Policy Optimization with Adaptive Clipping (BAPO) addresses challenges in off-policy reinforcement learning by dynamically adjusting clipping bounds to improve sample efficiency, stability, and performance in large language models.
Luis Wiedmann, Orr Zohar, Amir Mahla +6 authors
FineVision, a large-scale and curated dataset, enhances vision-language models through rigorous data collection, de-duplication, and human oversight, leading to improved performance.
Jiahan Zhang, Muqing Jiang, Nanru Dai +14 authors
World-in-World evaluates generative world models in closed-loop environments, emphasizing task success over visual quality and revealing insights into controllability, data scaling, and compute allocation.
Ling Team, Anqi Shen, Baihui Li +101 authors
Ring-1T, a trillion-parameter open-source thinking model, addresses training challenges with IcePop, C3PO++, and ASystem, achieving top results across benchmarks and democratizing large-scale reasoning intelligence.
Qianli Ma, Siyu Wang, Yilin Chen +7 authors
AutoPage, a multi-agent system, automates the creation of interactive research webpages through a hierarchical process, ensuring high-quality and efficient results.
Giorgos Nikolaou, Tommaso Mencattini, Donato Crisostomi +3 authors
Transformer language models are proven to be injective, allowing exact input reconstruction from hidden activations, which has implications for transparency and safety.
Jiale Cheng, Yusen Liu, Xinyu Zhang +11 authors
Glyph compresses long textual inputs into images using vision-language models, achieving significant token compression and improved performance in long-context tasks.
Yibin Wang, Zhimin Li, Yuhang Zang +8 authors
UniGenBench++ is a comprehensive benchmark for text-to-image generation that evaluates semantic consistency across diverse scenarios and languages using a hierarchical prompt structure and a robust evaluation pipeline.
Yuandong Pu, Le Zhuo, Songhao Han +10 authors
PICABench and PICAEval evaluate physical realism in image editing by assessing eight sub-dimensions and using VLM-as-a-judge with human annotations, highlighting the need for physics-based solutions.
Junliang Ye, Shenghao Xie, Ruowen Zhao +5 authors
Nano3D is a training-free framework that integrates FlowEdit and TRELLIS for precise 3D object editing, using front-view renderings and region-aware merging strategies to maintain structural fidelity and visual quality.
Siyuan Wang, Gaokai Zhang, Li Lyna Zhang +4 authors
LoongRL, a data-driven reinforcement learning method, enhances long-context reasoning by transforming short multi-hop QA into high-difficulty tasks, improving accuracy and generalization in large language models.
Yuezhou Hu, Jiaxin Guo, Xinyu Feng +1 authors
AdaSPEC enhances speculative decoding by selectively filtering tokens during knowledge distillation, improving token acceptance rates without sacrificing generation quality.
Jiahao Meng, Xiangtai Li, Haochen Wang +8 authors
Open-o3 Video integrates spatio-temporal evidence into video reasoning, achieving state-of-the-art performance on multiple benchmarks and providing valuable reasoning traces for test-time scaling.
GigaBrain Team, Angen Ye, Boyuan Wang +24 authors
GigaBrain-0, a VLA foundation model, uses world model-generated data to enhance cross-task generalization and policy robustness, improving real-world performance on complex manipulation tasks.
Weida Wang, Benteng Chen, Di Zhang +14 authors
Chem-R, a three-phase trained Chemical Reasoning model, achieves superior performance on chemical tasks by integrating core knowledge, expert reasoning, and multi-task optimization.
Qingyan Bai, Qiuyu Wang, Hao Ouyang +10 authors
Ditto framework addresses data scarcity in instruction-based video editing by generating a large dataset and using a curriculum learning strategy to train Editto, achieving superior instruction-following ability.
Jie-Ying Lee, Yi-Ruei Liu, Shr-Ruei Tsai +6 authors
Skyfall-GS creates large-scale, high-quality 3D urban scenes using satellite imagery and diffusion models, offering real-time exploration and improved geometry and texture consistency.
Maximo Eduardo Rulli, Simone Petruzzi, Edoardo Michielon +3 authors
Empirical analysis of Masked Diffusion Language Models (DLMs) reveals distinct attention sinking phenomena and robustness compared to Autoregressive Models (ARMs).
Minglei Shi, Haolin Wang, Wenzhao Zheng +6 authors
SVG, a novel latent diffusion model without VAEs, uses self-supervised representations to enable efficient training, few-step sampling, and high-quality visual generation with semantic and discriminative capabilities.
Junha Song, Sangdoo Yun, Dongyoon Han +2 authors
Reinforcement Learning enhances vision encoders in Multimodal Language Models, leading to better visual representations and performance compared to Supervised Fine-tuning.
Nikita Afonin, Nikita Andriyanov, Nikhil Bageshpura +8 authors
Emergent misalignment occurs in in-context learning across multiple models and datasets, with misaligned responses increasing with the number of examples provided.
Yihao Meng, Hao Ouyang, Yue Yu +9 authors
HoloCine generates coherent multi-shot narratives using a Window Cross-Attention mechanism and Sparse Inter-Shot Self-Attention, enabling end-to-end cinematic creation.
Weinan Jia, Yuning Lu, Mengqi Huang +6 authors
Mixture-of-Groups Attention (MoGA) enables efficient long video generation by addressing the quadratic scaling issue of full attention in Diffusion Transformers.
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