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。
Guilherme Penedo, Hynek Kydlíček, Loubna Ben allal +5 authors
FineWeb, a 15-trillion token dataset from Common Crawl snapshots, outperforms other open datasets in LLM training, and its educational subset, FineWeb-Edu, significantly improves performance on knowledge and reasoning benchmarks.
50 篇论文 · 按点赞排序
Lihe Yang, Bingyi Kang, Zilong Huang +4 authors
Depth Anything V2 improves monocular depth estimation through synthetic images, larger teacher models, and pseudo-labeled real images, achieving better efficiency and accuracy than Stable Diffusion models.
Kuzma Khrabrov, Anton Ber, Artem Tsypin +10 authors
A new dataset and benchmark, $\nabla^2$DFT, enhances neural network potentials for molecular property prediction, Hamiltonian prediction, and conformational optimization with extensive, diverse molecular data.
Daixuan Cheng, Yuxian Gu, Shaohan Huang +3 authors
Instruction Pre-Training enhances language models by generating and incorporating instruction-response pairs into unsupervised multitask pre-training.
Alexander Nikulin, Ilya Zisman, Alexey Zemtsov +3 authors
XLand-100B is a large-scale dataset for in-context reinforcement learning, containing extensive learning histories in the XLand-MiniGrid environment.
Lital Binyamin, Yoad Tewel, Hilit Segev +3 authors
CountGen enhances text-to-image diffusion models by accurately controlling the number of depicted objects using features that carry object identity information, surpassing existing methods on benchmark datasets.
Lin Chen, Xilin Wei, Jinsong Li +12 authors
The ShareGPT4Video series enhances video understanding and generation using dense captions and an efficient captioning strategy by addressing temporal and spatial challenges in video annotation.
Zhangchen Xu, Fengqing Jiang, Luyao Niu +4 authors
Magpie, a self-synthesis method, generates high-quality instruction data from aligned LLMs, enabling competitive performance in alignment benchmarks compared to models fine-tuned with larger, supervised datasets.
DeepSeek-AI, Qihao Zhu, Daya Guo +37 authors
DeepSeek-Coder-V2, a Mixture-of-Experts language model, excels in code-specific tasks by enhancing coding and mathematical reasoning capabilities while expanding language support and context length.
Peize Sun, Yi Jiang, Shoufa Chen +4 authors
LlamaGen applies the next-token prediction paradigm from large language models to image generation, achieving state-of-the-art performance with various model sizes and conditions.
Denis Bobkov, Vadim Titov, Aibek Alanov +1 authors
StyleFeatureEditor combines W-space and F-space for effective image attribute manipulation in StyleGAN, balancing reconstruction quality and detailed editing.
Yushun Zhang, Congliang Chen, Ziniu Li +5 authors
Adam-mini reduces memory usage by optimizing learning rate allocation in parameter blocks, achieving comparable or better performance than AdamW with increased throughput.
Tri Dao, Albert Gu
A new framework connects state-space models and transformers, leading to a faster architecture for language modeling.
Sander Schulhoff, Michael Ilie, Nishant Balepur +28 authors
A taxonomy and meta-analysis are provided for prompting techniques and vocabulary in GenAI systems, with a focus on natural language prefix-prompting.
Ziyan Jiang, Xueguang Ma, Wenhu Chen
The LongRAG framework improves information retrieval and answer extraction by using longer retrieval units and leveraging long-context LLMs, leading to better performance on benchmark datasets.
Shengbang Tong, Ellis Brown, Penghao Wu +11 authors
Cambrian-1, a vision-centric multimodal LLM family, introduces the Spatial Vision Aggregator (SVA) and a new benchmark CV-Bench to enhance visual grounding and representation learning in multimodal systems.
Ziyu Liu, Tao Chu, Yuhang Zang +8 authors
A new benchmark (MMDU) and instruction tuning dataset (MMDU-45k) are introduced to evaluate and enhance large vision-language models' performance in multi-turn and multi-image conversations, highlighting the need for improved conversational instruction tuning data.
Qihang Yu, Mark Weber, Xueqing Deng +3 authors
Transformer-based 1D Tokenizer (TiTok) reduces computational demands and enhances image synthesis efficiency by significantly decreasing token count and outperforming state-of-the-art methods in both speed and quality.
Junlin Wang, Jue Wang, Ben Athiwaratkun +2 authors
A Mixture-of-Agents (MoA) methodology combines multiple large language models (LLMs) to achieve state-of-the-art performance on various evaluations.
Yuang Peng, Yuxin Cui, Haomiao Tang +7 authors
dreambench++ uses advanced multimodal GPT models to create human-aligned evaluations for generative models in image generation, improving assessment accuracy and efficiency.
Yubo Wang, Xueguang Ma, Ge Zhang +14 authors
MMLU-Pro extends the MMLU benchmark with more challenging reasoning questions, eliminates trivial ones, and demonstrates better stability and discriminative power for language models.
Jeffrey Li, Alex Fang, Georgios Smyrnis +56 authors
DataComp for Language Models (DCLM) provides a benchmark with a standardized corpus and recipes to improve language models through controlled dataset experiments, showing the importance of data curation for performance and compute efficiency.
Chufan Shi, Cheng Yang, Yaxin Liu +11 authors
ChartMimic is a new benchmark assessing large multimodal models' visually-grounded code generation for chart rendering, evaluating multi-level cognitive capabilities with diverse chart types and metrics.
Weiyun Wang, Shuibo Zhang, Yiming Ren +13 authors
NN-NIAH is a benchmark that evaluates MLLMs' comprehension of long multimodal documents across retrieval, counting, and reasoning tasks.
Tao Zhang, Xiangtai Li, Hao Fei +5 authors
OMG-LLaVA integrates pixel-level vision understanding and reasoning capabilities through a universal segmentation method combined with LLM, achieving high performance across multiple benchmarks.
Ting-Hsuan Chen, Jiewen Chan, Hau-Shiang Shiu +3 authors
A video editing framework combining deformation fields and diffusion priors produces high-quality canonical images for diverse video editing tasks with accelerated training using low-rank adaptation.
Yuri Kuratov, Aydar Bulatov, Petr Anokhin +4 authors
BABILong benchmark evaluates large language models' reasoning across long documents, revealing that models use only a small portion of context and that recurrent memory transformers handle longer contexts better.
Duy-Kien Nguyen, Mahmoud Assran, Unnat Jain +3 authors
Vanilla Transformers achieve high performance in computer vision tasks by treating individual pixels as tokens, challenging the conventional use of local neighborhoods.
Terry Yue Zhuo, Minh Chien Vu, Jenny Chim +30 authors
Benchmark Bench assesses LLMs' capability to solve practical programming tasks by invoking function calls from various libraries, revealing significant performance gaps compared to human developers.
Ning-Hsu Wang, Yu-Lun Liu
A new depth estimation framework for 360-degree imagery uses pseudo-labels generated by perspective depth models to improve accuracy, especially in zero-shot scenarios.
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