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
Guojun Xiong, Zhiyang Deng, Keyi Wang +10 authors
A unified architecture combining LLMs and RL improves trading performance and other financial tasks through parameter-efficient fine-tuning and policy gradient optimization.
Zhenxing Mi, Kuan-Chieh Wang, Guocheng Qian +5 authors
ThinkDiff improves multimodal reasoning in text-to-image diffusion models using vision-language models aligned with the decoder of an encoder-decoder large language model.
Yi-Fan Zhang, Tao Yu, Haochen Tian +17 authors
MM-RLHF improves Multimodal Large Language Models' alignment with human preferences through a large preference dataset and innovative reward model techniques.
Zekun Qi, Wenyao Zhang, Yufei Ding +15 authors
Incorporating semantic orientation described in natural language enhances robotic manipulation capabilities, improving accuracy in tasks involving object orientations.
Shi Yu, Zhiyuan Liu, Chenyan Xiong
Crawl4LLM is an efficient web crawling method that prioritizes webpages based on their influence in LLM pretraining, significantly reducing the amount of low-quality data and improving downstream performance.
Haoming Xu, Ningyuan Zhao, Liming Yang +7 authors
ReLearn mitigates issues in current unlearning methods for large language models by proposing a data augmentation and fine-tuning pipeline that effectively targets forgetting while preserving quality and coherence, using novel metrics for evaluation.
Haowei Liu, Xi Zhang, Haiyang Xu +8 authors
A hierarchical agent framework called PC-Agent is proposed to improve the performance of MLLM-based GUI agents in complex PC environments through better perception and decision-making processes.
Seanie Lee, Dong Bok Lee, Dominik Wagner +5 authors
SafeRoute improves safety and efficiency in deploying large language models by selectively routing hard examples to a large safety guard model.
Ruotian Ma, Peisong Wang, Cheng Liu +6 authors
S$^2$R is a framework that enhances LLM reasoning through iterative self-verification and self-correction using supervised fine-tuning and reinforcement learning with minimal resources.
Guanzheng Chen, Xin Li, Michael Qizhe Shieh +1 authors
LongPO enables short-context LLMs to adapt to long-context tasks by self-learning from preference data and maintaining short-context performance.
Yein Park, Chanwoong Yoon, Jungwoo Park +2 authors
Temporal Heads in language models process and edit temporal knowledge through specific attention mechanisms, affecting recall without impacting general performance.
William Jurayj, Jeffrey Cheng, Benjamin Van Durme
Evaluating large language models includes confidence thresholding to manage response risk at test time, showing that increased compute improves both accuracy and confidence.
Shangqing Tu, Yucheng Wang, Daniel Zhang-Li +8 authors
A novel dataset and optimization method improve long-output generation in vision-language models while maintaining high-fidelity to input images.
Yixin Ou, Yunzhi Yao, Ningyu Zhang +5 authors
Analysis of knowledge circuit evolution in Large Language Models reveals key patterns in knowledge acquisition and suggests improvements for continual pre-training.
Rotem Shalev-Arkushin, Rinon Gal, Amit H. Bermano +1 authors
ImageRAG improves the generation of rare concepts by integrating dynamic image retrieval into existing image generation models without requiring specialized training.
Debangshu Banerjee, Tarun Suresh, Shubham Ugare +2 authors
A reasoning-augmented constrained decoding algorithm balances syntactic and semantic correctness in LLM outputs without reducing their reasoning capabilities.
Quentin Garrido, Nicolas Ballas, Mahmoud Assran +5 authors
Deep neural network models trained to predict masked regions in natural videos learn intuitive physics properties in a learned representation space, contrasting with models operating in pixel space or using language.
Michael Luo, Xiaoxiang Shi, Colin Cai +8 authors
Autellix improves LLM serving system efficiency by prioritizing program-level dependencies, significantly boosting throughput while maintaining latency.
Pan Lu, Bowen Chen, Sheng Liu +3 authors
OctoTools, an open-source agentic framework, enhances complex reasoning across diverse tasks by utilizing tool cards, planning, and execution without additional training.
Longxu Dou, Qian Liu, Fan Zhou +38 authors
Sailor2, a multilingual language model family tailored for Southeast Asian languages, achieves competitive performance against GPT-4 and includes a detailed cookbook for efficient development.
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