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
Geonung Kim, Janghyeok Han, Sunghyun Cho
VideoFrom3D synthesizes high-quality 3D scene videos using a combination of image and video diffusion models, achieving style consistency without requiring paired datasets.
Liang Wang, Nan Yang, Shaohan Huang +2 authors
Thinking augmented pre-training improves data efficiency and performance of large language models by augmenting text with automatically generated thinking trajectories.
Yunzhen Feng, Julia Kempe, Cheng Zhang +2 authors
Effective chain-of-thoughts in large reasoning models are characterized by fewer failed steps and better structural quality, not necessarily by length or review.
Yanzuo Lu, Xin Xia, Manlin Zhang +4 authors
Hyper-Bagel accelerates multimodal understanding and generation tasks using speculative decoding and multi-stage distillation, achieving significant speedups while maintaining high-quality outputs.
Kaiwen Zheng, Huayu Chen, Haotian Ye +7 authors
Diffusion Negative-aware FineTuning (DiffusionNFT) optimizes diffusion models directly on the forward process via flow matching, improving efficiency and performance compared to existing methods.
Avishai Elmakies, Hagai Aronowitz, Nimrod Shabtay +3 authors
A Group Relative Policy Optimization (GRPO)-based method using BLEU as a reward signal outperforms standard SFT for open-format speech understanding tasks like Spoken Question Answering and Automatic Speech Translation.
Joe Barrow
A web-scale dataset and models for form field detection are introduced, achieving high precision and supporting diverse languages and domains.
Xiang Deng, Jeff Da, Edwin Pan +16 authors
SWE-Bench Pro is a challenging benchmark for coding models, featuring complex, enterprise-level problems that require substantial code modifications, with performance evaluations showing significant limitations in current models.
Yi-Fan Zhang, Haihua Yang, Huanyu Zhang +12 authors
The paper provides a comprehensive guide and introduces BaseReward, a state-of-the-art multimodal reward model, which outperforms existing models across various benchmarks and real-world tasks.
Shaopeng Zhai, Qi Zhang, Tianyi Zhang +7 authors
VLAC, a vision-language-action reward model, enhances real-world robotic reinforcement learning by providing dense rewards and enabling one-shot transfer, significantly improving success rates and sample efficiency.
Zhenpeng Su, Leiyu Pan, Minxuan Lv +5 authors
A novel reinforcement learning algorithm, CE-GPPO, reintroduces gradients from clipped tokens to improve the exploration-exploitation balance in training large language models.
Lars Ankile, Zhenyu Jiang, Rocky Duan +3 authors
A residual learning framework combines behavior cloning and reinforcement learning to improve manipulation policies on high-degree-of-freedom systems using sparse binary rewards.
Minsoo Kim, Arnav Kundu, Han-Byul Kim +2 authors
EpiCache is a KV cache management framework for long conversational question answering that reduces memory usage and improves accuracy through block-wise prefill, episodic KV compression, and adaptive layer-wise budget allocation.
Xuan Ju, Tianyu Wang, Yuqian Zhou +11 authors
EditVerse is a unified framework using self-attention for image and video generation and editing, achieving state-of-the-art performance with a scalable data pipeline and benchmark.
Guizhen Chen, Weiwen Xu, Hao Zhang +4 authors
A two-stage reinforcement learning framework improves geometric reasoning and problem-solving in multimodal language models by first enhancing visual perception.
Shilin Lu, Zhuming Lian, Zihan Zhou +3 authors
SHINE is a training-free framework that uses manifold-steered anchor loss and pretrained customization adapters to seamlessly insert objects into new scenes with high fidelity, addressing challenges like complex lighting and diverse inputs.
Yuze He, Yanning Zhou, Wang Zhao +6 authors
CHARM uses a control-point-based parameterization and autoregressive transformer to generate high-fidelity anime hairstyles efficiently.
Kaiwen He, Zhiwei Wang, Chenyi Zhuang +1 authors
Recon-Act, a self-evolving multi-agent framework, improves adaptability and performance on long-horizon web tasks by generating and utilizing generalized tools through reconnaissance and action teams.
Natasha Butt, Ariel Kwiatkowski, Ismail Labiad +2 authors
A scalable reinforcement learning method for learning continuous chain-of-thought tokens in large language models improves performance and diversity over discrete tokens while preserving out-of-domain predictions.
Chen Wang, Chuhao Chen, Yiming Huang +4 authors
PhysCtrl is a physics-grounded framework for generating realistic, controllable videos from images using a diffusion model with spatiotemporal attention and physics-based constraints.
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