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
Rui Yang, Hanyang Chen, Junyu Zhang +10 authors
EmbodiedBench is a benchmark comprising diverse tasks to evaluate MLLMs in embodied agents, revealing strengths and weaknesses across high-level and low-level capabilities.
Xintao Wang, Heng Wang, Yifei Zhang +9 authors
CoSER, a dataset and evaluation protocol for role-playing language agents, enhances LLMs by simulating established characters with high-quality data and advanced models.
Kunat Pipatanakul, Pittawat Taveekitworachai, Potsawee Manakul +1 authors
Research demonstrates that language-specific LLMs can achieve advanced reasoning capabilities comparable to DeepSeek R1 using limited computational resources while preserving target language performance.
Pranav Nair, Puranjay Datta, Jeff Dean +2 authors
Matryoshka Quantization enables a single model to be served at multiple precision levels, improving accuracy of int2 quantization and efficiency.
Jikun Kang, Wenqi Wu, Filippos Christianos +5 authors
The Large Memory Model (LM2), a decoder-only Transformer with an auxiliary memory module, improves multi-step reasoning, relational argumentation, and large-context processing while preserving Transformer capabilities.
Jihoon Tack, Jack Lanchantin, Jane Yu +7 authors
A novel pretraining framework, Continuous Concept Mixing (CoCoMix), integrates continuous concepts from a sparse autoencoder into language models, improving sample efficiency and interpretability.
Xiao Wang, Ibrahim Alabdulmohsin, Daniel Salz +3 authors
Pre-training vision-language models on a 100-billion example dataset benefits cultural diversity and multilingual performance more than traditional benchmarks.
Dongzhi Jiang, Renrui Zhang, Ziyu Guo +11 authors
MME-CoT provides a comprehensive evaluation of Chain-of-Thought reasoning in Large Multimodal Models across multiple domains, revealing insights into model performance, efficiency, and potential pitfalls.
Henry Hengyuan Zhao, Difei Gao, Mike Zheng Shou
The paper presents WorldGUI, a novel benchmark for GUI tasks with varying initial states, and GUI-Thinker, a critique mechanism-based framework, which together improve GUI automation performance significantly.
Yiwen Tang, Zoey Guo, Zhuhao Wang +8 authors
An encoder-free architecture, ENEL, demonstrated competitive performance in 3D understanding tasks through LLM-embedded Semantic Encoding and Hierarchical Geometry Aggregation.
Sean McLeish, John Kirchenbauer, David Yu Miller +5 authors
The study analyzes scaling laws with diverse hyperparameters and architectures, releasing an extensive dataset named Gemstones to facilitate complex scaling law predictions in transformer models.
Xiao-Wen Yang, Xuan-Yi Zhu, Wen-Da Wei +5 authors
Introducing a self-backtracking mechanism to LLMs enhances reasoning efficiency and capability, outperforming supervised fine-tuning methods.
Zhihui Xie, Jie chen, Liyu Chen +3 authors
A reinforcement learning framework trains a critic model to generate feedback for code generation, enhancing performance and mitigating errors.
Shilong Zhang, Wenbo Li, Shoufa Chen +7 authors
A two-stage framework, FlashVideo, for text-to-video generation improves computational efficiency while maintaining high resolution and detail fidelity.
Hanmeng Liu, Zhizhang Fu, Mengru Ding +4 authors
This survey examines advancements in logical reasoning within large language models, analyzing capabilities across various paradigms and strategies for improvement.
Weigao Sun, Disen Lan, Yiran Zhong +2 authors
LASP-2 is a new sequence parallelism method that improves communication and computation parallelism for training linear attention transformer models with very-long input sequences, demonstrating significant speed improvements over previous methods.
David Abel, André Barreto, Michael Bowling +13 authors
The puzzle of determining agency in systems is explored through the lens of reinforcement learning, establishing that agency is inherently frame-dependent.
Md. Ashraful Islam, Mohammed Eunus Ali, Md Rizwan Parvez
CodeSim, a multi-agent framework, improves code generation and debugging through human-like perception and step-by-step simulation, achieving state-of-the-art results across several benchmarks.
Ling Yang, Zhaochen Yu, Bin Cui +1 authors
Hierarchical reasoning with LLMs using scaled thought templates improves mathematical reasoning and outperforms existing models on benchmarks like MATH and AIME.
Yihe Deng, Yu Yang, Junkai Zhang +2 authors
A two-player adversarial reinforcement learning framework generates high-quality synthetic data for multilingual guardrail models, improving safety and efficiency across languages.
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