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
Ziru Chen, Shijie Chen, Yuting Ning +17 authors
ScienceAgentBench evaluates language agents for scientific discovery tasks, showing limited capability in generating code for data-driven research.
Jiawei Mao, Xiaoke Huang, Yunfei Xie +4 authors
Story-Adapter is a training-free framework that enhances text-to-image models for long story visualization by iteratively refining images and maintaining semantic consistency through global reference cross-attention.
Junyi Zhang, Charles Herrmann, Junhwa Hur +5 authors
A novel geometry-first approach, Motion DUSt3R, directly estimates per-timestep geometry in dynamic scenes, demonstrating strong performance on tasks like video depth and camera pose estimation through fine-tuning on limited dynamic datasets.
Fu-Yun Wang, Ling Yang, Zhaoyang Huang +2 authors
Rectified Diffusion improves generation speed and performance by retraining with matched noise-sample pairs, simplifying the training procedure compared to previous methods.
Xiaoxiao He, Ligong Han, Quan Dao +14 authors
DICE enables precise inversion and flexible editing for discrete diffusion models, enhancing content manipulation in image and text domains.
Boyu Gou, Ruohan Wang, Boyuan Zheng +5 authors
Visual grounding models for GUI agents using pixel-level operations outperform text-based methods, demonstrating the feasibility of human-like visual perception in complex applications.
Han Yang, Kun Su, Yutong Zhang +4 authors
UniMuMo generates outputs across text, music, and motion modalities using a unified encoder-decoder transformer architecture and aligned unpaired data based on rhythmic patterns.
Deqing Fu, Tong Xiao, Rui Wang +5 authors
A token-level detective reward model (TLDR) provides detailed feedback for text tokens, assisting in self-correction and hallucination evaluation, and accelerating human annotation.
Dylan Zhang, Justin Wang, Francois Charton
Data diversification across semantic domains significantly enhances instruction-following capabilities in LLMs, leading to better generalization and performance.
Zachary Novack, Ge Zhu, Jonah Casebeer +3 authors
Presto! accelerates score-based diffusion transformers in text-to-music generation through step and layer distillation, achieving high-quality and fast outputs.
Mor Ventura, Michael Toker, Nitay Calderon +3 authors
NL-Eye is a benchmark that assesses VLMs' visual abductive reasoning skills by evaluating their ability to predict and explain the plausibility of hypothesis images based on premise images.
Jangyeong Kim, Donggoo Kang, Junyoung Choi +5 authors
The proposed method uses state-of-the-art 2D diffusion models and novel techniques to generate consistent, seamless, and well-aligned textures for 3D meshes, improving upon existing methods.
Shubhi Bansal, Sreeharish A, Madhava Prasath J +6 authors
Mamba, a State Space Model variant, offers linear time complexity and efficient long-range dependency handling in medical image analysis, outperforming transformers in merging multimodal data.
M. Jehanzeb Mirza, Mengjie Zhao, Zhuoyuan Mao +12 authors
GLOV method enhances vision tasks by leveraging LLMs as implicit optimizers for VLMs through meta-prompts and offset vector guidance.
Young-Jun Lee, Byungsoo Ko, Han-Gyu Kim +2 authors
The investigation into large language and vision models reveals that while they excel in advanced reasoning, they may lack robust perception, as seen through various benchmarks evaluating different aspects like permutation invariance and cross-modal alignment.
Desai Xie, Zhan Xu, Yicong Hong +5 authors
Progressive noise assignment in autoregressive video diffusion models enables high-quality generation of long videos without quality degradation or abrupt scene changes.
Fabian Paischer, Lukas Hauzenberger, Thomas Schmied +3 authors
Explained Variance Adaptation (EVA) enhances LoRA by data-driven initialization and rank redistribution to achieve faster convergence and superior performance across various fine-tuning tasks.
Qingxiu Dong, Li Dong, Xingxing Zhang +2 authors
SynPO, a self-boosting paradigm using synthetic preference data, enhances LLMs' instruction-following abilities and general performance without extensive human annotation.
Jiachen Li, Qian Long, Jian Zheng +4 authors
T2V-Turbo-v2 enhances a diffusion-based text-to-video model through consistency distillation with tailored dataset supervision, reward model feedback, and conditional guidance, setting a new state-of-the-art on VBench.
Lei Li, Yuanxin Liu, Linli Yao +6 authors
Textual Temporal reasoning Transfer (T3) enhances Video LLMs' temporal understanding by leveraging text-based temporal tasks, leading to superior performance on video benchmarks without additional video data.
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