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
Yunjie Ji, Xiaoyu Tian, Sitong Zhao +5 authors
AM-Thinking-v1, a 32B dense language model, achieves state-of-the-art performance in mathematical and coding tasks by leveraging supervised fine-tuning and reinforcement learning, demonstrating the capabilities of mid-scale open-source models.
Xingjin Wang, Howe Tissue, Lu Wang +2 authors
The study provides a scaling law for Continual Pre-Training (CPT) of large language models, characterizing the transition in performance curves and accounting for distribution shift and learning rate changes.
Rushi Qiang, Yuchen Zhuang, Yinghao Li +8 authors
MLE-Dojo provides an interactive framework for reinforcement learning and iterative improvement of large language model agents in real-world machine learning engineering tasks.
Kaixin Yao, Longwen Zhang, Xinhao Yan +6 authors
CAST is a method for reconstructing high-quality 3D scenes from single RGB images using GPT models, occlusion-aware 3D generation, and physics-aware correction with Signed Distance Fields.
Stefan Vasilev, Christian Herold, Baohao Liao +3 authors
Unilogit dynamically adjusts logits to enable selective forgetting in Large Language Models while maintaining overall utility and outperforming existing methods.
Zimu Lu, Yunqiao Yang, Houxing Ren +7 authors
A benchmark named WebGen-Bench measures LLM-based agents' ability to generate multi-file website codebases, achieving higher accuracy with open-source models than proprietary ones.
Peiying Zhang, Nanxuan Zhao, Jing Liao
A two-stage pipeline using a T2V diffusion model with path-level representation and distillation of customized T2I models achieves high-quality and diverse SVG generation with style customization.
Xiaobao Wu
A survey of the Learning from Rewards paradigm in Large Language Models, encompassing reinforcement learning techniques and reward-guided decoding strategies, enabling dynamic feedback and aligned preferences.
Jaehyun Jeon, Jang Han Yoon, Min Soo Kim +4 authors
G-FOCUS, a novel inference-time reasoning strategy, enhances Vision-Language Models for assessing UI persuasiveness, complementing A/B testing.
Hanjung Kim, Jaehyun Kang, Hyolim Kang +3 authors
UniSkill learns embodiment-agnostic skill representations from unlabeled cross-embodiment video data, enabling human skill transfer to robots in both simulation and real-world settings.
Saurabh Dash, Yiyang Nan, John Dang +22 authors
Multi-modal language models are enhanced through synthetic data creation and cross-modal merging techniques, achieving superior performance in multilingual settings.
Prime Intellect Team, Sami Jaghouar, Justus Mattern +11 authors
INTELLECT-2, a globally distributed reinforcement learning training of a 32 billion parameter language model, uses fully asynchronous RL across a dynamic, heterogeneous swarm of contributors, improving upon QwQ-32B with custom infrastructure and training modifications.
Zehan Wang, Siyu Chen, Lihe Yang +4 authors
Prior Depth Anything framework integrates metric depth measurement and geometric depth prediction into a coarse-to-fine pipeline for accurate, dense, and detailed depth maps, demonstrating superior zero-shot performance across multiple tasks.
Long Cheng, Jiafei Duan, Yi Ru Wang +12 authors
PointArena evaluates multimodal pointing models across various scenarios, demonstrating the importance of precise pointing in enabling abstract reasoning in real-world applications.
Niladri Shekhar Dutt, Duygu Ceylan, Niloy J. Mitra
A multimodal large language model is trained to suggest and apply procedural edits to photographs, preserving object identity and providing explainable results.
Chenze Shao, Fandong Meng, Jie Zhou
A Continuous VAR framework facilitates direct autoregressive generation of continuous visual data without quantization, using strictly proper scoring rules for optimization.
Hao Yang, Yan Yang, Ruikun Zhang +1 authors
GPT-4o, while generating visually appealing images, often lacks structural fidelity and serves as a visual prior to enhance dehazing, derainning, and low-light enhancement tasks in image restoration.
Shiqi Chen, Jinghan Zhang, Tongyao Zhu +5 authors
Merging models across modalities effectively transfers reasoning abilities from LLMs to VLMs without additional training, revealing layer-specific contributions to perception and reasoning.
Junjie Ye, Caishuang Huang, Zhuohan Chen +12 authors
An automated pipeline generates diverse, code-verifiable instruction-following test samples, revealing performance variation across constraints in large language models and enhancing instruction following via reinforcement learning.
Ziyang Huang, Xiaowei Yuan, Yiming Ju +2 authors
The IKEA agent optimizes reinforcement learning in LLMs by synergistically using internal and external knowledge, reducing hallucinations and unnecessary retrievals.
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