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
Sangmin Bae, Bilge Acun, Haroun Habeeb +5 authors
A comprehensive evaluation of hybrid language models combining self-attention with structured state space models, analyzing inter-layer and intra-layer fusion strategies, and providing design recommendations.
Siyoon Jin, Seongchan Kim, Dahyun Chung +5 authors
MATRIX-11K dataset and MATRIX regularization enhance interaction fidelity and semantic alignment in video DiTs by aligning attention with multi-instance mask tracks.
Cheng-Han Chiang, Xiaofei Wang, Linjie Li +7 authors
SHANKS, a general inference framework, enables spoken language models to generate unspoken reasoning while listening to user input, enhancing real-time interaction and task completion.
Guanghao Li, Kerui Ren, Linning Xu +7 authors
ARTDECO combines feed-forward models and SLAM pipelines for efficient and accurate 3D reconstruction from monocular images.
Ming Zhong, Xiang Zhou, Ting-Yun Chang +9 authors
Vibe Checker evaluates LLMs by combining functional correctness and instruction following to better align with human coding preferences.
Milad Aghajohari, Kamran Chitsaz, Amirhossein Kazemnejad +4 authors
Markovian Thinking, implemented in Delethink, enables efficient and scalable reinforcement learning for long-chain-of-thought reasoning in LLMs by decoupling thinking length from context size, resulting in linear compute and constant memory usage.
Kuofeng Gao, Yiming Li, Chao Du +4 authors
Imperceptible jailbreaks using Unicode variation selectors enable high attack success rates against aligned LLMs without visible prompt modifications.
Yunhao Fang, Weihao Yu, Shu Zhong +3 authors
A memory framework combining short-term and long-term memory in neural networks improves long-sequence modeling efficiency and performance.
Zichen Chen, Jiefeng Chen, Sercan Ö. Arik +3 authors
CoDA, a multi-agent system using specialized LLM agents, enhances visualization automation by managing data complexity and ensuring high-quality visualizations through collaborative workflows.
Zhanfeng Mo, Xingxuan Li, Yuntao Chen +1 authors
MATPO, a reinforcement learning method, optimizes tool-integrated multi-agent roles within a single LLM, improving performance and robustness over single-agent systems.
Oleg Filatov, Jiangtao Wang, Jan Ebert +1 authors
Joint optimal scaling of model and dataset sizes in deep learning is governed by the operator norm of the output layer, a phenomenon termed norm transfer, which provides a necessary condition for optimal learning rate and batch size.
Vage Egiazarian, Roberto L. Castro, Denis Kuznedelev +8 authors
A new quantization method, Micro-Rotated-GPTQ, addresses the challenges of 4-bit floating-point formats MXFP4 and NVFP4, achieving high performance and accuracy in large language model inference.
Tianshi Zheng, Kelvin Kiu-Wai Tam, Newt Hue-Nam K. Nguyen +10 authors
NewtonBench is a benchmark for scientific law discovery that addresses scalability, scientific relevance, and memorization resistance by using metaphysical shifts and interactive model discovery.
Aviv Navon, Aviv Shamsian, Neta Glazer +4 authors
Drax, a discrete flow matching framework for ASR, achieves state-of-the-art recognition accuracy with improved efficiency by constructing an audio-conditioned probability path.
Guijin Son, Donghun Yang, Hitesh Laxmichand Patel +9 authors
A language-mixed chain-of-thought reasoning approach improves performance in Korean-specific tasks by switching between English and Korean, achieving state-of-the-art results across various benchmarks.
Kai-Wei Chang, En-Pei Hu, Chun-Yi Kuan +7 authors
The Game-Time Benchmark evaluates the temporal dynamics and real-time interaction capabilities of conversational spoken language models, highlighting performance gaps in instruction-following and synchronized responses.
Haiquan Qiu, Quanming Yao
Low-precision training of transformer models with flash attention suffers from catastrophic loss explosions due to low-rank representations and biased rounding errors, which are addressed by a minimal modification to the flash attention mechanism.
Syeda Nahida Akter, Shrimai Prabhumoye, Eric Nyberg +4 authors
Introducing reasoning data during pretraining significantly enhances LLM performance compared to post-training, with pretraining benefiting more from diverse data patterns while SFT benefits more from high-quality data.
Sheriff Issaka, Keyi Wang, Yinka Ajibola +17 authors
The African Languages Lab addresses the underserved status of African languages in NLP by creating a large dataset and demonstrating improved model performance through fine-tuning.
Adam Filipek
The Reactive Transformer (RxT) addresses the limitations of stateless Transformers in conversational AI by using an event-driven paradigm with a fixed-size Short-Term Memory (STM) system, achieving linear scaling and low latency.
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