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
Archiki Prasad, Weizhe Yuan, Richard Yuanzhe Pang +6 authors
Self-consistency preference optimization (ScPO) enhances model training by iteratively preferring consistent answers, improving performance on reasoning tasks and outperforming larger models on specific benchmarks.
Yuyang Zhao, Chung-Ching Lin, Kevin Lin +6 authors
A framework called GenXD generates 3D and 4D scenes from multiview-temporal data, leveraging camera and object movements, and demonstrates superior performance compared to existing methods.
Qihang Yu, Ju He, Xueqing Deng +2 authors
Randomized AutoRegressive modeling (RAR) improves autoregressive image generation by randomly permuting input sequences during training, outperforming existing methods on the ImageNet-256 benchmark.
Pau Rodriguez, Arno Blaas, Michal Klein +4 authors
Activation Transport (AcT) is a framework that uses optimal transport theory to control model activations, effectively addressing challenges in large language models and text-to-image diffusion models.
Peiqi Liu, Zhanqiu Guo, Mohit Warke +4 authors
DynaMem, a dynamic memory system for robots, enhances open-vocabulary mobile manipulation by representing and updating 3D environments, achieving high success rates in non-stationary tasks.
Chuhan Li, Ziyao Shangguan, Yilun Zhao +3 authors
M3SciQA evaluates foundation models on multi-modal, multi-document scientific question answering, showing significant performance gaps compared to human experts.
Aniket Deroy, Subhankar Maity
This paper develops an approach using GPT-3.5 Turbo to extract relevant information from code-mixed conversations, particularly Roman transliterated Bengali mixed with English on social media.
Hao-Yu Hsu, Zhi-Hao Lin, Albert Zhai +2 authors
AutoVFX generates realistic VFX videos from a single video and natural language instructions using neural scene modeling, LLM-based code generation, and physical simulation.
He-Yen Hsieh, Ziyun Li, Sai Qian Zhang +5 authors
GazeGen uses a lightweight gaze estimation model to enable real-time, gaze-controlled manipulation of visual content, including generating images and videos.
Koichi Namekata, Sherwin Bahmani, Ziyi Wu +3 authors
SG-I2V is a self-guided framework for image-to-video generation that achieves zero-shot control and competitive visual quality and motion fidelity without fine-tuning or external knowledge.
Yang Yue, Yulin Wang, Bingyi Kang +5 authors
A Dynamic Early-Exit Framework for Robotic Vision-Language-Action Models optimizes computation and memory usage for MLLMs in robotics without sacrificing performance.
Chengke Zou, Xingang Guo, Rui Yang +3 authors
DynaMath, a dynamic visual math benchmark, evaluates the robustness of Vision-Language Models in mathematical reasoning across various input conditions, revealing significant accuracy differences between average and worst-case scenarios.
Hongjun An, Yiliang Song, Xuelong Li
The study identifies laws of information capacity in Next-token Prediction and applies Landauer's Principle to relate auto-regressive model training to energy consumption.
Shuhong Zheng, Zhipeng Bao, Ruoyu Zhao +2 authors
Diff-2-in-1 is a unified diffusion-based framework that integrates multi-modal data generation and dense visual perception, improving performance through a self-improving mechanism.
Ziyao Shangguan, Chuhan Li, Yuxuan Ding +4 authors
TOMATO, a new benchmark, evaluates the temporal reasoning capabilities of Multimodal Foundation Models in video understanding, revealing significant limitations beyond performance metrics.
Lucas Georges Gabriel Charpentier, David Samuel
A hybrid model combining masked and causal language modeling within a transformer stack outperforms models using either paradigm alone.
Shivam Duggal, Phillip Isola, Antonio Torralba +1 authors
A new method for learning variable-length token representations for 2D images through recursive processing and token specialization is proposed, validated by reconstruction loss and FID metrics.
Nam V. Nguyen, Thong T. Doan, Luong Tran +2 authors
LibMoE provides a modular framework for accessible research, training, and evaluation of Mixture of Experts algorithms in large language models, enabling standardized pipelines and benchmarking across various tasks.
Zichen Liu, Changyu Chen, Chao Du +2 authors
A sample-efficient algorithm called SEA is introduced for aligning large language models with human preferences using contextual dueling bandits and Thompson sampling.
Yean Cheng, Ziqi Cai, Ming Ding +5 authors
DreamPolish generates high-quality 3D assets with refined geometry and photorealistic textures using multiple neural representations, normal estimation, and a novel score distillation objective.
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