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。
Jinguo Zhu, Weiyun Wang, Zhe Chen +44 authors
InternVL3 is a multimodal pre-trained language model that jointly learns from both multimodal data and text, improving performance and scalability through advanced techniques and setting a new state-of-the-art in multimodal tasks.
50 篇论文 · 按点赞排序
Bang Liu, Xinfeng Li, Jiayi Zhang +44 authors
This survey covers the design, evaluation, and improvement of intelligent agents based on modular, brain-inspired architectures, focusing on self-enhancement, multi-agent collaboration, and safety in AI systems.
Andrés Marafioti, Orr Zohar, Miquel Farré +14 authors
SmolVLM, a series of compact multimodal models, achieves high performance with minimal GPU memory usage, making efficient deployment on mobile and edge devices possible.
Yiying Yang, Wei Cheng, Sijin Chen +6 authors
OmniSVG, using pre-trained Vision-Language Models and discrete tokenization, generates high-quality SVGs efficiently with a large, multimodal dataset.
Kimi Team, Angang Du, Bohong Yin +89 authors
Kimi-VL, an efficient Mixture-of-Experts vision-language model, excels in multimodal reasoning, long-context understanding, and diverse vision-language tasks, achieving competitive performance with reduced computational cost.
Zonghang Li, Tao Li, Wenjiao Feng +2 authors
Prima.cpp, a distributed inference system, enables running large language models on home devices using CPU/GPU and low memory, overcoming hardware limitations.
Cong Wei, Bo Sun, Haoyu Ma +10 authors
MoCha generates realistic talking character animations from speech and text using a speech-video attention mechanism and joint training on speech-labeled and text-labeled data, enabling multi-character conversations and superior realism.
Yang Yue, Zhiqi Chen, Rui Lu +5 authors
Despite initial claims, reinforcement learning with verifiable rewards does not introduce fundamentally new reasoning abilities to LLMs, instead it enhances performance by biasing output distribution toward rewarded paths without expanding the reasoning boundary.
Team Seawead, Ceyuan Yang, Zhijie Lin +51 authors
A mid-sized video generation model with 7 billion parameters, trained efficiently with substantial GPU hours, achieves competitive performance and strong generalization across various applications.
Minju Seo, Jinheon Baek, Seongyun Lee +1 authors
PaperCoder is a multi-agent LLM framework that converts machine learning papers into functional code repositories through planning, analysis, and generation stages.
Yuxin Zuo, Kaiyan Zhang, Shang Qu +7 authors
Test-Time Reinforcement Learning (TTRL) enhances Large Language Models (LLMs) using unlabeled data through reinforcement learning, improving performance across tasks.
Khalil Hennara, Sara Chrouf, Mohamed Motaism Hamed +3 authors
A method for integrating a new language into a large language model enhances performance in the target language while preserving existing knowledge.
Gleb Rodionov, Roman Garipov, Alina Shutova +5 authors
A parallel inference engine for LLMs allows concurrent task collaboration through a shared attention cache, improving efficiency without additional fine-tuning.
Karan Dalal, Daniel Koceja, Gashon Hussein +12 authors
Test-Time Training (TTT) layers enable pre-trained Transformers to generate coherent one-minute videos from text storyboards, outperforming alternatives like Mamba~2 and Gated DeltaNet.
Yiping Wang, Qing Yang, Zhiyuan Zeng +11 authors
Reinforcement learning with verifiable reward using one training example significantly enhances math reasoning capabilities of large language models.
Siyuan Li, Luyuan Zhang, Zedong Wang +8 authors
MergeVQ and MergeAR enhance MIM with VQ by integrating token merging and KV cache compression, improving both image generation and visual representation learning.
Shizhe Diao, Yu Yang, Yonggan Fu +12 authors
Pre-training datasets are typically collected from web content and lack inherent domain divisions. For instance, widely used datasets like Common Crawl do not include explicit domain labels, while manually curating labeled datasets such as The Pile is labor-intensive. Consequently, identifying an optimal pre-training data mixture remains a challenging problem, despite its significant benefits for pre-training performance. To address these challenges, we propose CLustering-based Iterative Data Mixture Bootstrapping (CLIMB), an automated framework that discovers, evaluates, and refines data mixtures in a pre-training setting. Specifically, CLIMB embeds and clusters large-scale datasets in a semantic space and then iteratively searches for optimal mixtures using a smaller proxy model and a predictor. When continuously trained on 400B tokens with this mixture, our 1B model exceeds the state-of-the-art Llama-3.2-1B by 2.0%. Moreover, we observe that optimizing for a specific domain (e.g., Social Sciences) yields a 5% improvement over random sampling. Finally, we introduce ClimbLab, a filtered 1.2-trillion-token corpus with 20 clusters as a research playground, and ClimbMix, a compact yet powerful 400-billion-token dataset designed for efficient pre-training that delivers superior performance under an equal token budget. We analyze the final data mixture, elucidating the characteristics of an optimal data mixture. Our data is available at: https://research.nvidia.com/labs/lpr/climb/
Nikai Du, Zhennan Chen, Zhizhou Chen +5 authors
TextCrafter is a progressive method that enhances the generation of intricate visual text within images, addressing issues like distortion, blurriness, and omissions through token focus enhancement.
Shiyu Liu, Yucheng Han, Peng Xing +21 authors
Step1X-Edit, an image editing model that uses multimodal LLM and diffusion image decoder, outperforms open-source models and approaches proprietary models in quality.
Abhay Kumar, Louis Owen, Nilabhra Roy Chowdhury +1 authors
ZClip is an adaptive gradient clipping algorithm that uses z-score-based anomaly detection to dynamically adjust clipping thresholds and prevent large gradient spikes during LLM training.
Jianhao Yan, Yafu Li, Zican Hu +5 authors
LUFFY enhances zero-RL models with off-policy guidance, improving reasoning and generalization through balanced imitation and exploration.
Shuming Ma, Hongyu Wang, Shaohan Huang +5 authors
BitNet b1.58 2B4T, a 1-bit Large Language Model with 2 billion parameters, matches the performance of full-precision models while improving computational efficiency.
Yi Peng, Chris, Xiaokun Wang +12 authors
Skywork R1V extends large language models to multimodal reasoning with efficient transfer, enhanced visual-text alignment, and dynamic reasoning chain optimization, achieving competitive performance in various benchmarks.
Sara Vera Marjanović, Arkil Patel, Vaibhav Adlakha +14 authors
DeepSeek-R1 generates detailed reasoning chains, offering insights into model behavior and cognitive phenomena, with findings highlighting performance nuances and safety concerns.
Jiacheng Liu, Taylor Blanton, Yanai Elazar +28 authors
OLMoTrace traces language model outputs to their training data in real-time, providing insights into fact checking, hallucination, and creativity through verbatim matches and infini-gram technology.
Zhiqiu Lin, Siyuan Cen, Daniel Jiang +12 authors
CameraBench assesses and enhances camera motion understanding by providing a large-scale dataset, taxonomy, and benchmark for evaluating SfM and VLMs, emphasizing the need for both semantic and geometric information.
Weiye Xu, Jiahao Wang, Weiyun Wang +10 authors
VisuLogic is a benchmark to evaluate genuine vision-centric reasoning in multimodal large language models, revealing significant performance gaps compared to human accuracy.
Shuai Wang, Zhi Tian, Weilin Huang +1 authors
A decoupled diffusion transformer improves performance and training speed in image generation by separating semantic extraction and high-frequency decoding.
Prateek Chhikara, Dev Khant, Saket Aryan +2 authors
Mem0, a memory-centric architecture with graph-based memory, enhances long-term conversational coherence in LLMs by efficiently extracting, consolidating, and retrieving information, outperforming existing memory systems in terms of accuracy and computational efficiency.
Shivalika Singh, Yiyang Nan, Alex Wang +10 authors
Measuring progress is fundamental to the advancement of any scientific field. As benchmarks play an increasingly central role, they also grow more susceptible to distortion. Chatbot Arena has emerged as the go-to leaderboard for ranking the most capable AI systems. Yet, in this work we identify systematic issues that have resulted in a distorted playing field. We find that undisclosed private testing practices benefit a handful of providers who are able to test multiple variants before public release and retract scores if desired. We establish that the ability of these providers to choose the best score leads to biased Arena scores due to selective disclosure of performance results. At an extreme, we identify 27 private LLM variants tested by Meta in the lead-up to the Llama-4 release. We also establish that proprietary closed models are sampled at higher rates (number of battles) and have fewer models removed from the arena than open-weight and open-source alternatives. Both these policies lead to large data access asymmetries over time. Providers like Google and OpenAI have received an estimated 19.2% and 20.4% of all data on the arena, respectively. In contrast, a combined 83 open-weight models have only received an estimated 29.7% of the total data. We show that access to Chatbot Arena data yields substantial benefits; even limited additional data can result in relative performance gains of up to 112% on the arena distribution, based on our conservative estimates. Together, these dynamics result in overfitting to Arena-specific dynamics rather than general model quality. The Arena builds on the substantial efforts of both the organizers and an open community that maintains this valuable evaluation platform. We offer actionable recommendations to reform the Chatbot Arena's evaluation framework and promote fairer, more transparent benchmarking for the field
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