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
Abdelrahman Shaker, Ahmed Heakl, Jaseel Muhammad +8 authors
A compact vision-language-diffusion model called Mobile-O enables efficient unified multimodal understanding and generation on mobile devices through specialized architecture design and optimized training methodology.
Hanxiang Qin, Alexander Martin, Rohan Jha +3 authors
Attention-guided clustering method for compressing multi-vector document representations in late interaction retrieval tasks shows superior performance compared to other compression techniques across text, visual-document, and video modalities.
Boyuan An, Zhexiong Wang, Yipeng Wang +4 authors
EgoPush enables robot manipulation in cluttered environments through perception-driven policy learning that uses object-centric latent spaces and stage-decomposed rewards for long-horizon tasks.
Kaiyu Yue, Menglin Jia, Ji Hou +1 authors
The Sphere Encoder is an efficient generative model that produces images in a single forward pass by mapping images to a spherical latent space and decoding from random points on that sphere, achieving diffusion-like quality with significantly reduced inference costs.
Sven Elflein, Ruilong Li, Sérgio Agostinho +4 authors
VGG-T³ addresses scalability issues in 3D reconstruction by transforming variable-length key-value representations into fixed-size MLPs through test-time training, enabling linear scaling with input views and achieving significant speedup over traditional softmax attention methods.
Kushal Kedia, Tyler Ga Wei Lum, Jeannette Bohg +1 authors
SimToolReal enables generalizable robot manipulation of diverse tools through procedural simulation and universal reinforcement learning policies without task-specific training.
Rui Yang, Qianhui Wu, Zhaoyang Wang +8 authors
GUI-Libra addresses limitations in open-source GUI agents through specialized training methods that improve reasoning-grounding alignment and reinforcement learning under partial verifiability, demonstrating enhanced task completion across web and mobile platforms.
Yue Su, Sijin Chen, Haixin Shi +7 authors
World Guidance framework enhances Vision-Language-Action models by mapping future observations into compact conditions for improved action generation and generalization.
Liangbing Zhao, Le Zhuo, Sayak Paul +2 authors
PhysicEdit enhances image editing by incorporating physical state transitions through a dual-thinking mechanism combining frozen vision-language models with learnable transition queries in a diffusion framework.
Jingxuan Zhang, Yunta Hsieh, Zhongwei Wang +5 authors
QuantVLA is a post-training quantization framework for vision-language-action models that enables efficient deployment through selective quantization, attention temperature matching, and output head balancing while maintaining performance and reducing memory and latency.
Shannan Yan, Leqi Zheng, Keyu Lv +7 authors
A conditional binary segmentation framework with cycle-consistency training enables robust object correspondence across egocentric and exocentric viewpoints without ground-truth annotations.
Kai Liu, Yanhao Zheng, Kai Wang +7 authors
JavisDiT++ presents a unified framework for joint audio-video generation using modality-specific mixture-of-experts, temporal-aligned RoPE, and audio-video direct preference optimization to achieve high-quality, synchronized multimedia synthesis.
Yukang Feng, Jianwen Sun, Zelai Yang +16 authors
LongCLI-Bench evaluates AI agents' ability to complete complex, multi-step programming tasks through command-line interfaces with detailed failure analysis and human-agent collaboration insights.
Zhongwei Wan, Yun Shen, Zhihao Dou +9 authors
DSDR is a reinforcement learning framework that enhances large language model reasoning by promoting diversity at both global and local levels through dual-scale regularization techniques.
Elad Ben Avraham, Changhao Li, Ron Dorfman +8 authors
Deep Research Agents generate analyst-grade reports, yet evaluating them remains challenging due to the absence of a single ground truth and the multidimensional nature of research quality. Recent benchmarks propose distinct methodologies, yet they suffer from the Mirage of Synthesis, where strong surface-level fluency and citation alignment can obscure underlying factual and reasoning defects. We characterize this gap by introducing a taxonomy across four verticals that exposes a critical capability mismatch: static evaluators inherently lack the tool-use capabilities required to assess temporal validity and factual correctness. To address this, we propose DREAM (Deep Research Evaluation with Agentic Metrics), a framework that instantiates the principle of capability parity by making evaluation itself agentic. DREAM structures assessment through an evaluation protocol combining query-agnostic metrics with adaptive metrics generated by a tool-calling agent, enabling temporally aware coverage, grounded verification, and systematic reasoning probes. Controlled evaluations demonstrate DREAM is significantly more sensitive to factual and temporal decay than existing benchmarks, offering a scalable, reference-free evaluation paradigm.
Wenjia Wang, Liang Pan, Huaijin Pi +8 authors
A portable dual-iPhone system enables metric-scale human-scene reconstruction and supports embodied AI tasks including physics-based animation and robot motion control.
Elron Bandel, Asaf Yehudai, Lilach Eden +12 authors
General-purpose agents remain underdeveloped despite promising implementations, necessitating systematic evaluation frameworks and benchmarks to assess their true versatility across diverse environments.
Xiaotong Ji, Rasul Tutunov, Matthieu Zimmer +1 authors
Decoding is reinterpreted as a principled optimization layer that balances model scores with structural preferences, recovering existing methods as special cases and enabling the creation of new decoders like Best-of-K that improve accuracy in mathematical reasoning tasks.
Hila Manor, Rinon Gal, Haggai Maron +2 authors
Visual analogy learning via dynamic composition of learned LoRA transformation primitives enables flexible image manipulation with improved generalization over fixed adaptation modules.
Xiaoke Huang, Bhavul Gauri, Kam Woh Ng +12 authors
VecGlypher is a multimodal language model that generates high-fidelity vector glyphs directly from text or image inputs, bypassing traditional raster-to-vector processes and enabling direct SVG path generation.
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