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
Jiawei Guo, Ziming Li, Xueling Liu +13 authors
CodeEditorBench evaluates Large Language Models in real-world code editing tasks, demonstrating that closed-source models perform better across various scenarios and problems.
Gowthami Somepalli, Anubhav Gupta, Kamal Gupta +5 authors
A new framework extracts style descriptors from images to attribute style in generated images to training data, with applications demonstrated in style retrieval and analysis for the Stable Diffusion model.
Atsuyuki Miyai, Jingkang Yang, Jingyang Zhang +7 authors
Unsolvable Problem Detection evaluates Vision Language Models' ability to identify unsolvable questions in Visual Question Answering through various settings, revealing areas for improvement and offering both training-free and training-based solutions.
Zixuan Huang, Justin Johnson, Shoubhik Debnath +2 authors
PointInfinity, a transformer-based point cloud diffusion model, efficiently generates high-resolution point clouds with state-of-the-art quality by using a fixed-size, resolution-invariant latent representation.
Niklas Stoehr, Mitchell Gordon, Chiyuan Zhang +1 authors
Memorized paragraphs in language models exhibit distinguishable gradient patterns and can be localized to specific low-layer attention heads and distinctive tokens.
Risto Luukkonen, Jonathan Burdge, Elaine Zosa +5 authors
A multilingual training approach on a large language model improves capabilities for small languages, translation, and generation in multiple languages.
Wentian Zhang, Haozhe Liu, Jinheng Xie +3 authors
Cross-attention in text-conditional diffusion models divides inference into semantics-planning and fidelity-improving stages, leading to an efficient generation method called TGATE.
Xingyi Zhou, Anurag Arnab, Shyamal Buch +5 authors
A streaming dense video captioning model with a novel memory module and decoding algorithm improves state-of-the-art performance on multiple benchmarks while handling long videos efficiently.
Han Cai, Muyang Li, Zhuoyang Zhang +3 authors
A Condition-Aware Neural Network (CAN) method for image generation dynamically adjusts neural network weights based on input conditions, demonstrating significant improvements for diffusion transformer models and surpassing existing models in terms of FID score and computational efficiency.
Shujie Hu, Long Zhou, Shujie Liu +8 authors
WavLLM, a speech LLM with dual encoders and prompt-aware LoRA weight adapter, achieves state-of-the-art performance across various speech tasks using a two-stage curriculum learning approach.
Ruohong Zhang, Liangke Gui, Zhiqing Sun +8 authors
A novel framework using video captions as a proxy enhances preference modeling and performance of video LMMs by aligning with GPT-4V's reward mechanism through direct preference optimization.
Ali Behrouz, Michele Santacatterina, Ramin Zabih
MambaMixer, a new architecture with selective token and channel mixing, demonstrates competitive performance in vision and outstanding performance in time series forecasting compared to state-of-the-art methods.
Hanting Chen, Zhicheng Liu, Xutao Wang +2 authors
DiJiang, a Frequency Domain Kernelization approach, reduces the computational load of pre-trained Transformers with little additional training, achieving comparable performance to vanilla Transformers with significantly lower costs and faster inference.
Yisheng He, Weihao Yuan, Siyu Zhu +3 authors
NeRF editing achieves high-fidelity results by decomposing images into low-frequency and high-frequency components and editing in the low-frequency space to maintain consistency and detail.
Shuo Chen, Zhen Han, Bailan He +5 authors
A comprehensive jailbreak evaluation dataset is created to assess the robustness of various LLMs and MLLMs against jailbreak attacks, revealing differences in robustness between open-source and proprietary models, particularly in visual versus textual attacks.
Ritesh Sarkhel, Xiaoqi Ren, Lauro Beltrao Costa +5 authors
Noise-Aware Training (NAT) method improves efficiency and robustness in training named entity extractors using weakly labeled documents, outperforming transfer learning in terms of macro-F1 score and reducing human labeling effort.
Mauro Comi, Alessio Tonioni, Max Yang +5 authors
A novel approach combining touch and vision data improves surface reconstruction and novel view synthesis, especially for non-Lambertian objects, using optimized 3D Gaussian primitives.
Yunzhi Zhang, Zizhang Li, Amit Raj +5 authors
A 3D Congealing framework aligns 2D images via a shared 3D canonical space, optimizing for geometry and semantics without pre-assumed shapes or camera parameters.
Detai Xin, Xu Tan, Kai Shen +8 authors
RALL-E, a robust text-to-speech method using chain-of-thought prompting, enhances prosody prediction and reduces word error rate in zero-shot TTS compared to VALL-E.
Ruyang Liu, Chen Li, Haoran Tang +3 authors
ST-LLM leverages large language models to effectively encode and understand videos for dialogue systems by integrating spatial-temporal modeling with dynamic masking and global-local input strategies, achieving state-of-the-art performance on VideoChatGPT-Bench and MVBench.
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