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
Shanghaoran Quan, Tianyi Tang, Bowen Yu +7 authors
The Self-Lengthen framework iteratively trains LLMs to generate longer, coherent outputs using only intrinsic model capabilities without requiring additional data or proprietary models.
Chulin Xie, Yangsibo Huang, Chiyuan Zhang +6 authors
LLMs demonstrate both near-perfect memorization of training puzzles and improved generalization after fine-tuning, indicating a complex interplay between memorization and genuine reasoning.
Yongchang Hao, Yanshuai Cao, Lili Mou
NeuZip compresses neural network weights using entropy for memory-efficient training and inference without performance loss.
Renze Lou, Hanzi Xu, Sijia Wang +15 authors
A benchmark dataset evaluates LLMs for research tasks requiring deep expertise, highlighting their potential and limitations.
Xueyang Yu, Xinlei Chen, Yossi Gandelsman
A synthetic video dataset enables effective pre-training of VideoMAE models for video tasks without using natural videos, achieving high performance on various classification and out-of-distribution datasets.
Itay Yona, Ilia Shumailov, Jamie Hayes +1 authors
Attackers can exploit architectural flaws in Mixture-of-Experts models to extract user prompts by manipulating query batches in a specific manner.
Jianlan Luo, Charles Xu, Jeffrey Wu +1 authors
A human-in-the-loop vision-based RL system achieves high performance on diverse robotic manipulation tasks with near-perfect success rates and fast training times.
Ruisi Cai, Yeonju Ro, Geon-Woo Kim +4 authors
Read-ME framework transforms pre-trained dense LLMs into efficient MoE models using activation sparsity, a pre-gating router, and enhanced batching, improving inference performance and scalability.
Omer Nahum, Nitay Calderon, Orgad Keller +2 authors
LLMs used as judges can identify label errors in datasets, improving annotation quality and model performance across various tasks and domains.
Phillip Y. Lee, Taehoon Yoon, Minhyuk Sung
A novel training-free method using Diffusion Transformers achieves state-of-the-art spatial grounding in text-to-image generation by generating and denoising patches corresponding to each bounding box.
Ryan Liu, Jiayi Geng, Addison J. Wu +3 authors
Chain-of-thought prompting reduces performance in tasks such as implicit statistical learning, visual recognition, and classifying with patterns containing exceptions, while it retains or increases performance in tasks where humans benefit from verbal thinking.
Anurag Bagchi, Zhipeng Bao, Yu-Xiong Wang +2 authors
REM, a video segmentation framework using visual-language models, excels in segmenting and tracking unseen and dynamic concepts by leveraging Internet-scale pre-training.
Catherine Arnett, Eliot Jones, Ivan P. Yamshchikov +1 authors
A new fully open-source pipeline for toxicity filtering is introduced to detect and reduce harmful outputs in language models trained on public domain data.
Grace Luo, Trevor Darrell, Amir Bar
Vision-and-language models encode conceptually similar tasks into similar vector representations across different modalities and specifications, undergoing an input-task-answer process that benefits from ensemble methods.
Hanshi Sun, Li-Wen Chang, Wenlei Bao +6 authors
ShadowKV is a high-throughput inference system for long-context LLMs that reduces memory footprint and minimizes decoding latency by using a low-rank key cache and minimal sparse KV pair reconstruction.
Zhixin Zhang, Yiyuan Zhang, Xiaohan Ding +1 authors
Vision Search Assistant enhances VLMs by collaborating with web agents for retrieving and generating information about unseen visual content, improving performance on image-based QA tasks.
Xiang Zhang, Juntai Cao, Chenyu You
Research explores how tokenization affects the counting abilities of large language models, revealing that byte-level tokenization can impair their theoretical computability and suggesting the need for new tokenization techniques to improve reasoning.
Lester James V. Miranda, Yizhong Wang, Yanai Elazar +6 authors
A routing framework that combines human and synthetic annotations using a performance prediction model improves reward model performance and reduces human annotation costs.
Yingzhe Peng, Xiaoting Qin, Zhiyang Zhang +6 authors
CARE, a multi-agent LLM framework with a structured user interface, enhances personalization in exploratory tasks by improving user interaction, reducing cognitive load, and providing tailored solutions.
Guangqi Jiang, Yifei Sun, Tao Huang +3 authors
A new representation learning framework, MCR, enhances manipulation-centric visual representations by integrating both visual features and dynamic information, leading to improved performance in robotic manipulation tasks.
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