Trends · 研究趋势
数据来自 Hugging Face 论文的 AI 提取关键词,按月统计研究方向的增长与热度。
Yifan Shen, Jian Xu, Boyi Li +6 authors
ChronoVision improves visual temporal reasoning by aligning latent imagery with logic through reconstructive prediction, ROI attention, and reinforcement learning, achieving strong results on video reasoning benchmarks.
Yunfei Xie, Juncheng Wu, Haoqin Tu +6 authors
OpenAI's o1 demonstrates strong performance in medical tasks but shows weaknesses in reasoning, multilinguality, and evaluation metrics compared to GPT-4.
Zican Hu, Xuyang Hu, Yiming Liu +10 authors
BRAID framework enables unified multi-modal reasoning by casting text-image interaction as a Markov decision process, allowing joint optimization through reinforcement learning with vision-language model guidance.
Johan Obando-Ceron, Ghada Sokar, Timon Willi +6 authors
Incorporating Soft MoE modules into value-based networks enhances scalability and performance across different training regimes and model sizes in reinforcement learning.
DiffusionGemma Team, Adrien Ali Taïga, James Assiene +41 authors
DiffusionGemma is a fine-tuned mixture-of-experts language model that uses discrete diffusion to generate text blocks in parallel, achieving high speed while preserving capabilities like multimodal inputs and reasoning.
Zehan Qi, Xiao Liu, Iat Long Iong +10 authors
WebRL enhances open LLMs as web agents through a self-evolving curriculum, robust outcome-supervised reward model, and adaptive reinforcement learning strategies, achieving high success rates compared to proprietary models.
Huizhuo Yuan, Zixiang Chen, Kaixuan Ji +1 authors
Self-play fine-tuning improves diffusion models' performance and alignment with human preferences using competition between model versions, outperforming supervised and RL-based methods with less data.
Akshay Krishnamurthy, Keegan Harris, Dylan J. Foster +2 authors
Experimentation with LLMs in multi-armed bandit environments reveals that robust exploration requires specific prompts, external summarization of history, or algorithmic interventions.
Jonghyun Song, Sangjun Song, Minjae Oh +3 authors
SHAPE analyzes chain-of-thought reasoning via semantic spaces and heuristics to diagnose LLM mathematical reasoning and improve post-training.
Shangqing Tu, Daniel Zhang-Li, Yucheng Wang +21 authors
CogEvol is a family of models that generate structured learning artifacts in a single pass using supervised fine-tuning and reinforcement learning with vision-language rewards, achieving high quality with far fewer parameters and lower cost.
Sheng-Chieh Lin, Luyu Gao, Barlas Oguz +4 authors
Proposes factuality-aware alignment methods for pre-trained language models to enhance factual accuracy during supervised fine-tuning and reinforcement learning, reducing hallucination.
Eloi Alonso, Adam Jelley, Vincent Micheli +4 authors
DIAMOND, a reinforcement learning agent using a diffusion world model, achieves top performance on Atari 100k by capturing visual details effectively.
Yufei Wang, Zhou Xian, Feng Chen +6 authors
RoboGen automates the learning of diverse robotic skills at scale using generative simulation to propose tasks and generate training environments and supervision.
Chunting Zhou, Pengfei Liu, Puxin Xu +12 authors
A 65B parameter LLaMa language model trained with minimal instructional data matches or outperforms models with extensive human preference modeling in most cases, indicating that pretraining is predominantly responsible for knowledge acquisition.
Yecheng Jason Ma, William Liang, Guanzhi Wang +6 authors
Eureka, an LLM-powered reward design algorithm, generates human-level reward functions enhancing reinforcement learning for complex manipulation tasks and RLHF, including pen spinning with a simulated Shadow Hand.
Junha Jung, Minbyul Jeong, Suhyeon Lim +5 authors
A reinforcement learning approach called MRPO is introduced to improve clinical image reasoning by addressing cascading errors through step-wise process rewards, demonstrating superior performance over existing methods.
Gabrielle Kaili-May Liu, Avi Caciularu, Gal Yona +2 authors
Reinforcement learning with metacognitive feedback and metacognitive data selection improve large language model calibration by enabling accurate self-assessment of performance and uncertainty.
Mihir Prabhudesai, Anirudh Goyal, Deepak Pathak +1 authors
AlignProp refines text-to-image diffusion models using backpropagation through the denoising process, leveraging low-rank adapters and gradient checkpointing to optimize for various objectives with higher efficiency.
Juan Rocamonde, Victoriano Montesinos, Elvis Nava +2 authors
Pretrained vision-language models can serve as zero-shot reward models in reinforcement learning, enabling complex task learning from minimal natural language prompts.
Wanxia Cao, Chengzhen Duan, Pei Fu +27 authors
A native multimodal GUI agent trained in real-device environments demonstrates superior performance and stability compared to traditional benchmark-based approaches.
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