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。
50 篇论文 · 按点赞排序
Jintian Zhang, Yuqi Zhu, Mengshu Sun +6 authors
LightThinker improves efficiency in LLMs by dynamically compressing intermediate thoughts, reducing memory usage and inference time while maintaining performance.
Liang Chen, Shuai Bai, Wenhao Chai +5 authors
A unified framework called Dream Engine integrates powerful text-to-image models with versatile multimodal encoders to enable arbitrary text-image interleaved control in image generation.
Chuofan Ma, Yi Jiang, Junfeng Wu +5 authors
UniTok, a discrete visual tokenizer utilizing multi-codebook quantization, achieves competitive performance in image generation and understanding tasks.
Orion Weller, Kathryn Ricci, Eugene Yang +3 authors
Rank1 is a re-ranking model using test-time compute and reasoning language models to enhance performance and explainability compared to smaller models, demonstrating state-of-the-art results on reasoning and instruction-following datasets.
Yancheng He, Shilong Li, Jiaheng Liu +7 authors
DeltaBench evaluates the ability of models and process reward models to detect errors in long Chain-of-Thought reasoning generated by o1-like models across various reasoning tasks.
Alexander Zhang, Marcus Dong, Jiaheng Liu +15 authors
A new benchmark, CodeCriticBench, is introduced to evaluate code critique capacity of LLMs by addressing limitations in existing benchmarks, such as narrow focus and insufficient task diversity.
Guijin Son, Jiwoo Hong, Hyunwoo Ko +1 authors
Multilingual math benchmark MCLM tests various test-time scaling methods on multilingual LLMs, showing that performance gains from these methods are limited outside of English.
Guizhen Chen, Weiwen Xu, Hao Zhang +6 authors
A new logic-puzzle benchmark, FINEREASON, evaluates LLMs' intermediate reasoning steps and demonstrates improved math reasoning capabilities.
Hao Peng, Yunjia Qi, Xiaozhi Wang +4 authors
An agentic reward system that combines human preferences with verifiable correctness signals, demonstrated through RewardAgent, enhances reward reliability and improves large language model performance.
Christoph Schuhmann, Gollam Rabby, Ameya Prabhu +9 authors
Using language models to convert scholarly documents into structured Knowledge Units can legally and effectively preserve and disseminate scientific knowledge.
Chengyin Xu, Kaiyuan Chen, Xiao Li +2 authors
The Clustering-On-Difficulty framework predicts LLM downstream performance by leveraging task clustering and mapping functions, achieving high accuracy with small models.
Chenlong Wang, Zhaoyang Chu, Zhengxiang Cheng +6 authors
CODESYNCBENCH is introduced to evaluate LLMs' adaptation to evolving third-party library APIs, revealing their challenges with dynamic code changes.
Shiven Sinha, Shashwat Goel, Ponnurangam Kumaraguru +3 authors
REFUTE is a benchmark for evaluating language models' ability to falsify incorrect solutions, improving their role in scientific discovery and self-improvement.
Patrick Tser Jern Kon, Jiachen Liu, Qiuyi Ding +7 authors
Curie is an AI agent framework that improves rigor in scientific experimentation through reliability, control, and interpretability, demonstrating a significant improvement over baselines in answering experimental questions.
Tianjin Huang, Haotian Hu, Zhenyu Zhang +8 authors
Stable-SPAM, an enhanced version of SPAM optimizer, stabilizes gradient norms and improves performance in low-bit precision training across various bit levels.
Enzhe Lu, Zhejun Jiang, Jingyuan Liu +22 authors
Mixture of Block Attention (MoBA) improves large language models' efficiency and performance on long-context tasks by dynamically switching between full and sparse attention without predefined biases.
Sotiris Anagnostidis, Gregor Bachmann, Yeongmin Kim +7 authors
A flexible diffusion transformer model reduces computational requirements for image and video generation without sacrificing quality.
Min Zhao, Guande He, Yixiao Chen +3 authors
By analyzing positional embeddings, RIFLEx reduces frequency components to achieve high-quality video extrapolation without additional training or extensive fine-tuning.
Xiuli Bi, Jianfei Yuan, Bo Liu +4 authors
A novel method for generating looping videos from text using pre-trained video latent diffusion models without user annotations, ensuring temporal consistency and enabling dynamic motion.
Qianqi Yan, Yue Fan, Hongquan Li +5 authors
A benchmark assesses state-of-the-art multimodal large language models' ability to detect and reason about semantic inconsistencies in real-world content, revealing gaps in cross-modal reasoning.
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