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
Eduard Allakhverdov, Dmitrii Tarasov, Elizaveta Goncharova +1 authors
Image reconstruction reveals that vision encoders retain more image information after image-based tasks and that orthogonal rotations in feature space control color encoding.
SiXiang Chen, Jianyu Lai, Jialin Gao +11 authors
PosterCraft improves aesthetic poster generation through a unified, modular pipeline with enhanced text rendering, region-aware fine-tuning, aesthetic reinforcement learning, and joint vision-language refinement.
Sheng Chen, Peiyu He, Jiaxin Hu +68 authors
Astra, a dual-model architecture for mobile robot navigation, uses a multimodal LLM for global localization and a multitask network for local path planning and odometry estimation, achieving high success rates in diverse indoor environments.
Jiatao Gu, Tianrong Chen, David Berthelot +7 authors
STARFlow, a generative model combining normalizing flows with autoregressive Transformers, achieves competitive image synthesis performance with innovations in architecture and latent space modeling.
C. Opus, A. Lawsen
Evaluation artifacts, particularly token limits and impractical instances in benchmarks, lead to misreported failures in Large Reasoning Models on planning puzzles.
Aditi Sundararaman, Amogh Adishesha, Andrew Jaegle +10 authors
Mirage generates realistic video from audio inputs, integrating with speech synthesis to create compelling multimodal content through a unified, self-attention-based training approach.
Donghoon Ahn, Jiwon Kang, Sanghyun Lee +7 authors
The paper proposes HeadHunter, a systematic framework for selecting attention heads in Diffusion Transformer architectures to enable precise control over image generation quality and style, outperforming existing methods.
Yijia Dai, Zhaolin Gao, Yahya Satter +2 authors
In-context learning in large language models can effectively model sequences generated by hidden Markov models, achieving predictive accuracy and uncovering scaling trends, thus demonstrating its potential as a diagnostic tool for complex scientific data.
Tinghui Zhu, Kai Zhang, Muhao Chen +1 authors
Research investigates the impact of extending modality and model merging on maintaining language abilities and generalization in omni-modal language models.
Yizhao Gao, Shuming Guo, Shijie Cao +12 authors
SeerAttention-R is a sparse attention framework for reasoning models that maintains high accuracy and achieves significant speedups through optimized sparse decoding kernels.
Sangwon Jang, Taekyung Ki, Jaehyeong Jo +4 authors
Frame Guidance offers a training-free method for controlling video generation using frame-level signals, reducing memory usage and enhancing globally coherent video output.
Shangshang Wang, Julian Asilis, Ömer Faruk Akgül +4 authors
SAE-Tuning efficiently elicits strong reasoning in language models by leveraging sparse autoencoders, enabling cost-effective performance gains without extensive retraining.
Mozhi Zhang, Howe Tissue, Lu Wang +1 authors
Domain2Vec decomposes datasets into meta-domains to optimize language model pretraining and downstream performance with reduced computational cost.
Nick Jiang, Amil Dravid, Alexei Efros +1 authors
A training-free method shifts high-norm activations in Vision Transformers to an untrained token, enhancing attention maps and performance across visual tasks, and improving interpretability in vision-language models.
Dror Ivry, Oran Nahum
Sentinel, a detection model based on ModernBERT-large, effectively identifies prompt injection attacks with high accuracy and outperforms existing baselines.
Xing Han Lù, Gaurav Kamath, Marius Mosbach +1 authors
A new Agentic Web Interface (AWI) design paradigm is proposed to optimize web agents for navigating websites, focusing on safety, efficiency, and standardization to address fundamental interface mismatches.
Zaijing Li, Yuquan Xie, Rui Shao +4 authors
Optimus-3, a multimodal large language model agent, uses knowledge-enhanced data generation, a Mixture-of-Experts architecture, and multimodal reasoning-augmented reinforcement learning to achieve superior performance across various tasks in Minecraft.
Or Tal, Felix Kreuk, Yossi Adi
A systematic comparison of Auto-Regressive decoding and Conditional Flow-Matching in text-to-music generation highlights distinct strengths and limitations of each modeling paradigm.
Aadarsh Sahoo, Vansh Tibrewal, Georgia Gkioxari
A unified language, image, and 3D scene model framework is proposed, achieving optimal training and performance across various 3D tasks and datasets.
Hongyu Wang, Chuyan Xiong, Ruiping Wang +1 authors
BitVLA, a 1-bit VLA model with ternary parameters, achieves comparable performance to OpenVLA-OFT on LIBERO while using 29.8% less memory through distillation-aware training.
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