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
Zihan Liu, Zhuolin Yang, Yang Chen +4 authors
Combining supervised fine-tuning and reinforcement learning enhances reasoning models, especially when optimizing sampling temperature and leveraging strong initial fine-tuning, as demonstrated by the improved AceReason-Nemotron-1.1 model.
Saksorn Ruangtanusak, Natthapath Rungseesiripak, Peerawat Rojratchadakorn +2 authors
DoTA-RAG improves retrieval and generation accuracy over massive web datasets using a dynamic routing pipeline and optimized embedding models, achieving high correctness scores while maintaining low latency.
Zihan Zheng, Zerui Cheng, Zeyu Shen +16 authors
LLMs perform well on implementation-heavy competitive programming problems but struggle with nuanced algorithmic reasoning, as highlighted by LiveCodeBench Pro.
Christoph Schuhmann, Robert Kaczmarczyk, Gollam Rabby +6 authors
EmoNet-Voice, a new resource with large pre-training and benchmark datasets, advances speech emotion recognition by offering fine-grained emotion evaluation with synthetic, privacy-preserving audio.
Yining Hong, Rui Sun, Bingxuan Li +7 authors
Embodied Web Agents integrate physical interaction and web-scale reasoning to assess cross-domain intelligence in a novel benchmark environment.
Dingfeng Shi, Jingyi Cao, Qianben Chen +14 authors
TaskCraft automates the generation of scalable, multi-tool, and complex agentic tasks to enhance prompt optimization and fine-tuning of agentic models.
Xiyao Wang, Zhengyuan Yang, Chao Feng +10 authors
ViCrit, an RL task for fine-tuning VLMs, improves visual perception by training models to detect subtle hallucinations in image captions, with gains transferable to various visual domains.
Liyan Xu, Zhenlin Su, Mo Yu +3 authors
A new dataset named CapRetrieval is introduced to evaluate the ability of text encoders to recognize fine-grained entities and events, highlighting challenges in dense retrieval tasks.
Jiacong Wang, Zijiang Kang, Haochen Wang +8 authors
VGR, a novel multimodal large language model, improves visual reasoning by detecting relevant image regions and integrating them into the reasoning process, outperforming existing models on multimodal benchmarks with reduced resource usage.
Meiling Tao, Chenghao Zhu, Dongyi Ding +3 authors
A new benchmark, PersonaFeedback, evaluates Large Language Models' ability to generate personalized responses given explicit user personas, revealing limitations in current systems.
Xiaoran Liu, Siyang He, Qiqi Wang +9 authors
FourierAttention is a training-free framework that enhances memory efficiency in Large Language Models by compressing long-context-insensitive transformer head dimensions using orthogonal Fourier bases, while maintaining accuracy.
Amirmojtaba Sabour, Sanja Fidler, Karsten Kreis
Flow maps, introduced with new continuous-time objectives and training techniques, achieve state-of-the-art performance in few-step image and text-to-image generation.
Wanlong Liu, Junxiao Xu, Fei Yu +7 authors
Question-Free Fine-Tuning (QFFT) improves efficiency and adaptability in cognitive models by leveraging both short and long chain-of-thought patterns, reducing response length while maintaining performance across various scenarios.
Jaehoon Yun, Jiwoong Sohn, Jungwoo Park +9 authors
Med-PRM enhances clinical decision making by verifying reasoning steps against medical knowledge bases, achieving state-of-the-art performance in medical QA benchmarks with improved accuracy.
Mathurin Videau, Badr Youbi Idrissi, Alessandro Leite +3 authors
An autoregressive U-Net learns to embed its own tokens during training, enabling a multi-scale view of text sequences and improved handling of character-level tasks and low-resource languages.
Zheyuan Yang, Zexi Kuang, Xue Xia +1 authors
TestCase-Eval is a benchmark for evaluating LLMs in generating comprehensive and targeted test cases for algorithm problems.
Joanito Agili Lopo, Muhammad Ravi Shulthan Habibi, Tack Hwa Wong +6 authors
Research confirms natural representation alignment in large language models and introduces Inference-Time Language Control to enhance cross-lingual performance.
Alexander Polok, Jiangyu Han, Dominik Klement +3 authors
The combined DiCoW and DiariZen ASR system demonstrates strong performance in multilingual scenarios, with DiCoW preserving its multilingual capabilities and DiariZen improving through fine-tuning.
Yao Zhang, Chenyang Lin, Shijie Tang +4 authors
SwarmAgentic is a framework for automated agentic system generation that optimize agent functionality and collaboration through language-driven exploration, outperforming existing baselines in unconstrained tasks.
Chiwei Zhu, Benfeng Xu, Xiaorui Wang +1 authors
The paper presents a method for generating diverse and complex instruction data for large language models using attributed grounding, achieving top performance on benchmarks with a large synthesized dataset.
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