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。
Loubna Ben Allal, Anton Lozhkov, Elie Bakouch +19 authors
SmolLM2, a small language model with 1.7 billion parameters, achieves strong performance through overtraining on diverse datasets, outperforming other recent small models.
50 篇论文 · 按点赞排序
Gaojie Lin, Jianwen Jiang, Jiaqi Yang +2 authors
OmniHuman, a Diffusion Transformer-based framework, achieves realistic human video generation by integrating motion-related conditions into its training process, supporting various inputs and driving modalities.
Niklas Muennighoff, Zitong Yang, Weijia Shi +7 authors
The use of budget forcing during test time improves reasoning performance in language models, as demonstrated by the s1 model which outperforms the o1-preview on competition math questions.
Alexey Gorbatovski, Boris Shaposhnikov, Viacheslav Sinii +2 authors
Direct Alignment Algorithms improve language model alignment by introducing a supervised fine-tuning phase and adjusting preference optimization strength, showing that ranking objectives are crucial for performance.
Hila Chefer, Uriel Singer, Amit Zohar +5 authors
VideoJAM enhances generative video models by integrating motion coherence with appearance fidelity, achieving top performance in motion coherence and visual quality.
Yixin Ye, Zhen Huang, Yang Xiao +3 authors
LIMO, a new model, achieves high mathematical reasoning performance using minimal training data, challenging the notion that extensive datasets are necessary for complex reasoning.
Ganqu Cui, Lifan Yuan, Zefan Wang +20 authors
PRIME leverages implicit process rewards to improve the reinforcement learning of large language models, achieving better performance with less data compared to traditional methods.
Daniil Laptev, Nikita Balagansky, Yaroslav Aksenov +1 authors
A new data-free cosine similarity method maps and steers features across layers of large language models, enhancing interpretability and targeted control in text generation.
Edward Yeo, Yuxuan Tong, Morry Niu +2 authors
Investigation into long chains-of-thought reasoning in large language models reveals the critical role of training compute, reward shaping, and verifiable reward signals in enabling and measuring this capability.
Yuri Chervonyi, Trieu H. Trinh, Miroslav Olšák +7 authors
AlphaGeometry2, an enhanced version of AlphaGeometry, uses Gemini architecture and a novel knowledge-sharing mechanism to solve a broader range of Olympiad geometry problems with higher accuracy.
Yuzhe Yang, Yifei Zhang, Minghao Wu +5 authors
TwinMarket, a multi-agent framework using large language models, simulates socio-economic dynamics and emergent phenomena like financial bubbles and recessions in a simulated stock market.
Ahmed Masry, Juan A. Rodriguez, Tianyu Zhang +19 authors
A new vision-text alignment method, AlignVLM, effectively maps visual features to LLM embeddings, improving performance in document understanding and robustness to noise.
Dawei Li, Renliang Sun, Yue Huang +6 authors
Preference leakage, a bias in LLM-as-a-judge systems caused by the relatedness between synthetic data generators and evaluators, has been empirically confirmed and found to be pervasive and challenging.
Baohao Liao, Yuhui Xu, Hanze Dong +5 authors
Reward-Guided Speculative Decoding improves inference efficiency in large language models by combining a draft and target model with a reward-based strategy, achieving better performance and cost savings.
Alec Helbling, Tuna Han Salih Meral, Ben Hoover +2 authors
ConceptAttention, a method using DiT attention layers, generates precise saliency maps for textual concepts in images, achieving state-of-the-art performance in zero-shot image segmentation without additional training.
Peiqing Yang, Shangchen Zhou, Jixin Zhao +2 authors
MatAnyone, a memory-based video matting framework, uses region-adaptive memory fusion for robust and accurate matting in complex scenarios.
Shashwat Goel, Joschka Struber, Ilze Amanda Auzina +6 authors
Evaluating and supervising advanced language models is increasingly challenging, and using other language models for oversight can introduce risks due to correlated mistakes.
Huaye Zeng, Dongfu Jiang, Haozhe Wang +3 authors
Automated test-case synthesis improves reward models and reinforces coding models, demonstrating significant performance gains.
Danah Yatim, Rafail Fridman, Omer Bar-Tal +1 authors
A zero-shot, training-free method using text-to-video diffusion transformer and Vision Language Model synthesizes and integrates dynamic content into real-world videos based on text instructions.
Zuyan Liu, Yuhao Dong, Jiahui Wang +4 authors
Ola is an omni-modal language model that achieves competitive performance across image, video, and audio understanding through progressive modality alignment and efficient cross-modal training.
Zhen Ye, Xinfa Zhu, Chi-Min Chan +17 authors
A single-layer vector quantizer codec and Transformer-based framework for text-to-speech synthesis improves speech naturalness and emotional expressiveness with scalable compute resources.
Nikita Gushchin, David Li, Daniil Selikhanovych +3 authors
A novel distillation method accelerates diffusion bridge models for image-to-image translation tasks, improving inference speed and quality.
Xun Liang, Simin Niu, Zhiyu Li +8 authors
The SafeRAG benchmark evaluates the security of RAG systems by simulating various attack scenarios, revealing significant vulnerabilities in RAG components.
Rohit Gandikota, Zongze Wu, Richard Zhang +3 authors
SliderSpace is a framework that automatically finds multiple interpretable directions in diffusion models from a single text prompt, enabling better control and discovery of model capabilities across various applications.
Bo Pang, Hanze Dong, Jiacheng Xu +3 authors
A method is introduced to enable long chain-of-thought reasoning in large language models using in-context learning and supervised fine-tuning without distillation from pre-existing models or extensive human annotations.
Qika Lin, Tianzhe Zhao, Kai He +5 authors
A two-stage framework using self-supervised quantized representation (SSQR) integrates Knowledge Graphs (KGs) with Large Language Models (LLMs), producing more distinguishable codes and enhancing performance on KG tasks.
Lichen Bai, Masashi Sugiyama, Zeke Xie
The Weak-to-Strong Diffusion (W2SD) framework enhances the alignment of learned distributions with real data distributions by leveraging differences between weak and strong models, leading to significant improvements in human preference, aesthetic quality, and prompt adherence across multiple modalities and architectures.
Xinyan Guan, Jiali Zeng, Fandong Meng +6 authors
DeepRAG improves answer accuracy by 21.99% through strategic and adaptive retrieval, modeled as a Markov Decision Process for efficient reasoning-augmented generation.
Da Yu, Edith Cohen, Badih Ghazi +5 authors
SCONE extends language model performance by using contextualized n-gram embeddings, precomputed and stored off-accelerator, to minimize inference costs while maintaining fixed FLOPS.
Ken M. Nakanishi
Scalable-Softmax (SSMax) enhances Transformer-based models by improving attention distribution and long-context performance, resolving issues posed by the Softmax function.
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