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
Ling Yang, Zixiang Zhang, Zhilong Zhang +6 authors
Consistency Flow Matching improves the efficiency and quality of sample generation by enforcing self-consistency in velocity fields and using multi-segment training for flow-based generative models.
Chenchen Ye, Ziniu Hu, Yihe Deng +4 authors
MIRAI introduces a benchmark to evaluate LLM agents' forecasting capabilities for international events using historical databases and APIs.
Dustin Wright, Arnav Arora, Nadav Borenstein +3 authors
Analysis of LLM responses to the Political Compass Test reveals bias through diverse prompt variations and recurring tropes in justifications.
Bozhong Tian, Xiaozhuan Liang, Siyuan Cheng +6 authors
MemFlex, a method leveraging gradient information, improves precise knowledge unlearning and retains essential knowledge in Large Language Models without excessive data erasure.
Omer Goldman, Alon Jacovi, Aviv Slobodkin +3 authors
A position paper argues for a more precise taxonomy of long-context tasks based on their diffusion and scope of necessary information, noting that highly diffused and lengthy information in long contexts is under-explored.
Chirui Chang, Zhengzhe Liu, Xiaoyang Lyu +1 authors
An ensemble of 3D convolutional network classifiers detects fake videos by analyzing appearance, motion, and geometry, revealing persistent gaps between real and synthetic videos.
Yiming Zhang, Yicheng Gu, Yanhong Zeng +4 authors
FoleyCrafter, a novel framework, generates high-quality, semantically aligned, and temporally synchronized sound effects for videos using parallel cross-attention layers and a timestamp-based adapter.
Yilun Xu, Gabriele Corso, Tommi Jaakkola +2 authors
Discrete-Continuous Latent Variable Diffusion Models simplify generative learning by introducing discrete latent variables, enhancing performance across various tasks.
Yicheng Chen, Xiangtai Li, Yining Li +4 authors
Auto Cherry-Picker, a framework for generating high-quality multi-modal training examples using language models and a novel metric, enhances performance in perception tasks, particularly addressing long-tailed distributions and imbalanced datasets.
Zihao Wang, Shaofei Cai, Zhancun Mu +7 authors
OmniJARVIS, a Vision-Language-Action model for Minecraft, uses unified tokenization of multimodal interaction data with a behavior encoder and autoregressive transformers to perform reasoning, planning, question answering, and action execution.
Yifei Zhang, Xintao Wang, Jiaqing Liang +3 authors
Chain-of-Knowledge framework enhances large language models with knowledge reasoning capabilities through rule mining and trial-and-error mechanisms.
Taoran Yi, Jiemin Fang, Zanwei Zhou +7 authors
GaussianDreamerPro enhances the quality of 3D-Gaussian asset generation by binding Gaussians to evolving geometry, improving details and facilitating integration into downstream applications.
Jianyu Wei, Shijie Cao, Ting Cao +4 authors
T-MAC is a lookup table-based method for efficient low-bit LLM inference on CPUs, supporting mixed precision matrix multiplication without dequantization and reducing computational overhead.
William Chen, Wangyou Zhang, Yifan Peng +7 authors
XEUS, a cross-lingual encoder for universal speech, enhances language coverage and robustness through novel dereverberation, outperforming existing models on various benchmarks with less data and parameters.
Peng Dai, Feitong Tan, Qiangeng Xu +5 authors
A pose-free, training-free method uses monocular video generation to produce high-quality 3D stereoscopic videos through warping, frame matrix inpainting, and disocclusion handling.
Yuxuan Zhang, Tianheng Cheng, Rui Hu +6 authors
EVF-SAM combines image and text prompts with early vision-language fusion for improved referring expression segmentation, achieving state-of-the-art results on RefCOCO/+/g with fewer parameters.
Jia Fu, Xiaoting Qin, Fangkai Yang +7 authors
AutoRAG-HP uses a Hierarchical MAB method to efficiently optimize hyperparameters in RAG systems, achieving better recall with reduced API calls compared to grid search.
Chao-Wei Huang, Hui Lu, Hongyu Gong +4 authors
Decoder-only large language models integrated with speech-to-text translation achieve state-of-the-art performance without proprietary data through parameter-efficient fine-tuning.
Vincent Micheli, Eloi Alonso, François Fleuret
A new agent, $\Delta$-IRIS, achieves state-of-the-art performance in model-based reinforcement learning by using a discrete autoencoder and autoregressive transformer architecture, resulting in faster training and improved simulation efficiency.
Ziqi Wang, Hanlin Zhang, Xiner Li +6 authors
The proposed Position-Invariant Inference (PINE) method eliminates position bias in language models by changing causal to bidirectional attention, improving performance and reliability in tasks like LM-as-a-judge and retrieval-augmented QA.
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