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
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.
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.
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/
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.
Ding Chen, Qingchen Yu, Pengyuan Wang +6 authors
xVerify, an efficient answer verifier, demonstrates strong equivalence judgment capabilities in evaluating reasoning models across various objective questions, with F1 scores and accuracy surpassing other methods.
Yu Gao, Lixue Gong, Qiushan Guo +28 authors
Seedream 3.0 improves Chinese-English bilingual image generation by enhancing data training, pre-training techniques, and post-training aesthetics, resulting in higher visual quality and faster image generation.
Jiazhan Feng, Shijue Huang, Xingwei Qu +6 authors
ReTool, a tool-integrated learning framework, enhances reasoning models with real-time code execution and reinforcement learning, significantly improving performance in structured problem-solving tasks like mathematical reasoning.
Yash Savani, Asher Trockman, Zhili Feng +4 authors
Antidistillation sampling modifies a model's next-token probability distribution to disrupt the generation of reasoning traces for distillation without affecting model performance.
Fangzhi Xu, Hang Yan, Chang Ma +6 authors
Genius, an unsupervised self-training framework, enhances LLM reasoning by employing a stepwise foresight re-sampling strategy and advantage-calibrated optimization to optimize responses without external supervision.
Lvmin Zhang, Maneesh Agrawala
FramePack, a neural network for video generation, compresses frames to manage transformer context length and enhances video diffusion models with increased batch sizes and improved frame prediction.
Ning Li, Jingran Zhang, Justin Cui
Evaluation of GPT-4o shows limitations in its ability to integrate world knowledge, apply contextual reasoning, and adhere to complex instructions in image generation and editing.
Yijun Liang, Ming Li, Chenrui Fan +7 authors
ColorBench evaluates vision-language models' color perception, reasoning, and robustness, revealing limitations and emphasizing the need for enhanced color comprehension in multimodal AI.
Haozhe Wang, Chao Qu, Zuming Huang +3 authors
Vision-language models enhanced with reinforcement learning and Forced Rethinking achieve state-of-the-art performance on math and science benchmarks and approach the capabilities of slow-thinking systems.
Tianwei Xiong, Jun Hao Liew, Zilong Huang +2 authors
GigaTok improves image reconstruction, generation, and representation quality by scaling visual tokenizers with semantic regularization and specific architectural practices.
Junliang Guo, Yang Ye, Tianyu He +4 authors
MineWorld, a real-time interactive world model built on Minecraft, uses a visual-action autoregressive Transformer to generate consequent game scenes based on inputs of game scenes and actions, outperforming state-of-the-art diffusion-based models.
Daniel Bolya, Po-Yao Huang, Peize Sun +15 authors
Perception Encoder, trained via contrastive vision-language learning, achieves state-of-the-art performance across various image and video tasks using intermediate embeddings extracted through alignment methods.
Ming Li, Yanhong Li, Ziyue Li +1 authors
A spectral analysis of gradients during the post-training of large language models reveals the impact of data quality using metrics like nuclear norms and effective ranks, providing insights into improving data exploration strategies.
Tsung-Han Wu, Heekyung Lee, Jiaxin Ge +3 authors
REVERSE is a unified framework that reduces hallucinations in Vision-Language Models by integrating hallucination-aware training and on-the-fly self-verification using a novel retrospective resampling technique.
Zheng Liu, Mengjie Liu, Jingzhou Chen +4 authors
FUSION, a multimodal large language model, integrates vision and language through deep, pixel-level and question-level integration, achieving superior performance compared to existing methods.
Zeqi Xiao, Yushi Lan, Yifan Zhou +4 authors
WorldMem uses a memory bank with attention mechanism to enhance scene generation by preserving 3D spatial and temporal consistency across viewpoints and time.
Haozhan Shen, Peng Liu, Jingcheng Li +9 authors
A reinforcement learning framework extends the capabilities of Vision-Language Models (VLMs) in visual reasoning by leveraging rule-based reward formulations, achieving competitive performance and superior generalization compared to supervised fine-tuning.
Peixian Ma, Xialie Zhuang, Chengjin Xu +3 authors
A novel NL2SQL model named SQL-R1 leverages reinforcement learning to improve reasoning in complex scenarios, achieving high accuracy with minimal data.
Nikita Sorokin, Ivan Sedykh, Valentin Malykh
An iterative self-training approach using Proximal Policy Optimization enhances reranker models for code generation, improving accuracy and quality compared to larger models.
Wenlei Shi, Xing Jin
Heimdall, a long CoT verification LLM, significantly boosts solution accuracy through reinforcement learning and Pessimistic Verification, effectively identifying flawed data in synthesized datasets.
Tianyi Zhang, Yang Sui, Shaochen Zhong +3 authors
DFloat11, a lossless compression framework, reduces Large Language Model sizes by 30% through entropy coding, enabling efficient deployment and significantly higher throughput on resource-constrained hardware.
Yang Shi, Jiaheng Liu, Yushuo Guan +12 authors
Mavors, a novel framework for long-context video understanding, uses a multi-granularity video representation to preserve spatial fidelity and temporal continuity through 3D convolutions, Vision Transformers, and transformer-based dependency modeling.
Hardy Chen, Haoqin Tu, Fali Wang +5 authors
Supervised fine-tuning can hinder reinforcement learning in Large Vision-Language Models by inducing imitative reasoning, whereas a novel reinforcement learning approach improves genuine reasoning capabilities and outperforms existing models.
Xing Han Lù, Amirhossein Kazemnejad, Nicholas Meade +7 authors
AgentRewardBench is a benchmark evaluating the effectiveness of LLMs in assessing web agent performance, showing that rule-based methods may underreport success.
Xin Gao, Qizhi Pei, Zinan Tang +5 authors
A collaborative framework using multiple small language models achieves high-quality data synthesis similar to large language models through a peer-review-inspired process.
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