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。
606 篇论文 · 按点赞排序
Zhiliang Peng, Jianwei Yu, Wenhui Wang +10 authors
VibeVoice synthesizes long-form multi-speaker speech using next-token diffusion and a highly efficient continuous speech tokenizer, achieving superior performance and fidelity.
Jin Xu, Zhifang Guo, Jinzheng He +11 authors
Qwen2.5-Omni is a multimodal model that processes text, images, audio, and video in a streaming fashion and generates text and speech using a dual-track architecture, achieving state-of-the-art performance on multimodal benchmarks.
Jiachen Zhu, Xinlei Chen, Kaiming He +2 authors
Dynamic Tanh (DyT) replaces normalization layers in Transformers, achieving equivalent or superior performance without hyperparameter tuning across various tasks.
B. Y. Yan, Chaofan Li, Hongjin Qian +2 authors
GAM, a novel framework that employs JIT compilation principles, improves memory efficiency and task completion by leveraging a lightweight memorizer and researcher in conjunction with reinforcement learning.
Boyang Zheng, Nanye Ma, Shengbang Tong +1 authors
Replacing VAEs with pretrained representation encoders in Diffusion Transformers enhances generative quality and convergence speed without auxiliary losses.
Jingyang Yuan, Huazuo Gao, Damai Dai +12 authors
NSA, a trainable sparse attention mechanism, enhances long-context modeling efficiency without sacrificing performance, achieving improvements in speed and accuracy over full attention models.
Zhiyu Li, Shichao Song, Chenyang Xi +36 authors
MemOS, a memory operating system for Large Language Models, addresses memory management challenges by unifying plaintext, activation-based, and parameter-level memories, enabling efficient storage, retrieval, and continual learning.
Michael Tschannen, Alexey Gritsenko, Xiao Wang +11 authors
SigLIP 2, a multilingual vision-language encoder, improves upon SigLIP with unified training techniques, enhancing performance in zero-shot classification, image-text retrieval, localization, and dense prediction across various model sizes and data diversity.
Ahmed Nassar, Andres Marafioti, Matteo Omenetti +10 authors
SmolDocling is a compact vision-language model that performs end-to-end document conversion with robust performance across various document types using 256M parameters and a new markup format.
Shuai Bai, Yuxuan Cai, Ruizhe Chen +61 authors
Qwen3-VL, a vision-language model, excels in text and multimodal understanding through advanced architectures and larger contexts, achieving superior performance across benchmarks.
Shunyu Liu, Minghao Liu, Huichi Zhou +29 authors
VeriGUI is a novel dataset for evaluating GUI agents in long-horizon tasks, emphasizing long-chain complexity and subtask-level verifiability.
Mustafa Shukor, Dana Aubakirova, Francesco Capuano +11 authors
SmolVLA is a compact, efficient vision-language-action model that achieves competitive performance at reduced computational costs and can be deployed on consumer-grade hardware.
Huichi Zhou, Yihang Chen, Siyuan Guo +8 authors
A novel memory-augmented reinforcement learning paradigm enables adaptive LLM agents to continually learn without fine-tuning, using episodic memory and a neural case-selection policy.
Yukang Chen, Wei Huang, Baifeng Shi +11 authors
A framework scales vision-language models for long video reasoning using reinforcement learning, achieving strong performance on benchmarks and demonstrating consistent gains with increased video frames.
Guanting Dong, Hangyu Mao, Kai Ma +11 authors
Agentic Reinforced Policy Optimization (ARPO) enhances multi-turn reasoning in large language models by balancing long-horizon capabilities and tool interactions, using entropy-based adaptive rollouts and advantage attribution.
Fangyu Lei, Jinxiang Meng, Yiming Huang +14 authors
DAComp is a benchmark of 210 tasks that evaluates the capabilities of agents in real-world data engineering and data analysis workflows, revealing significant deficiencies in both areas.
Yuyang Hu, Shichun Liu, Yanwei Yue +44 authors
This survey provides an updated overview of agent memory research, distinguishing its forms, functions, and dynamics, and highlights emerging research directions.
Jonas Geiping, Sean McLeish, Neel Jain +6 authors
A novel language model scales test-time computation by reasoning in latent space, outperforming token-based models on reasoning benchmarks.
Dong Guo, Faming Wu, Feida Zhu +194 authors
Seed1.5-VL, a vision-language foundation model combining a vision encoder and a large MoE LLM, achieves state-of-the-art performance across various benchmarks and excels in multimodal reasoning tasks such as visual puzzles.
Zhiqiu Lin, Siyuan Cen, Daniel Jiang +12 authors
CameraBench assesses and enhances camera motion understanding by providing a large-scale dataset, taxonomy, and benchmark for evaluating SfM and VLMs, emphasizing the need for both semantic and geometric information.
Junru Lu, Jiarui Qin, Lingfeng Qiao +35 authors
Youtu-LLM is a lightweight language model optimized for computational efficiency and agentic intelligence through a compact architecture, STEM-focused training curriculum, and scalable mid-training strategies for planning and reasoning tasks.
Jin Xu, Zhifang Guo, Hangrui Hu +35 authors
Qwen3-Omni, a multimodal model, achieves state-of-the-art performance across text, image, audio, and video, using a Thinker-Talker MoE architecture and a lightweight causal ConvNet for efficient streaming synthesis.
Bo Peng, Ruichong Zhang, Daniel Goldstein +12 authors
RWKV-7 "Goose" achieves state-of-the-art performance in multilingual tasks with optimal memory and inference efficiency, exceeding Transformer capabilities in complexity.
Runze Liu, Junqi Gao, Jian Zhao +5 authors
Test-Time Scaling improves Large Language Model performance by optimizing inference computation based on policy models, Process Reward Models, and problem difficulty.
Zhi Zhou, Yuhao Tan, Zenan Li +4 authors
A theoretical framework for sampling-based test-time scaling in large language models reveals limitations in self-consistency and perplexity, and introduces RPC to improve reasoning performance and reduce sampling costs.
Haozhe Wang, Haoran Que, Qixin Xu +9 authors
REER, a new paradigm for deep reasoning, uses reverse engineering to discover step-by-step reasoning processes, enabling a model to perform competitively on open-ended tasks.
Fuhao Li, Wenxuan Song, Han Zhao +5 authors
Vision-language-action (VLA) models have recently shown strong potential in enabling robots to follow language instructions and execute precise actions. However, most VLAs are built upon vision-language models pretrained solely on 2D data, which lack accurate spatial awareness and hinder their ability to operate in the 3D physical world. Existing solutions attempt to incorporate explicit 3D sensor inputs such as depth maps or point clouds, but these approaches face challenges due to sensor noise, hardware heterogeneity, and incomplete depth coverage in existing datasets. Alternative methods that estimate 3D cues from 2D images also suffer from the limited performance of depth estimators.We propose Spatial Forcing (SF), a simple yet effective alignment strategy that implicitly forces VLA models to develop spatial comprehension capabilities without relying on explicit 3D inputs or depth estimators. SF aligns intermediate visual embeddings of VLAs with geometric representations produced by pretrained 3D foundation models. By enforcing alignment at intermediate layers, SF guides VLAs to encode richer spatial representations that enhance action precision.Extensive experiments in simulation and real-world environments demonstrate that SF achieves state-of-the-art results, surpassing both 2D- and 3D-based VLAs. SF further accelerates training by up to 3.8x and improves data efficiency across diverse robotic tasks. Project page is at https://spatial-forcing.github.io/
Heejun Lee, Geon Park, Jaduk Suh +1 authors
InfiniteHiP is a novel LLM inference framework that enhances long-context processing with token pruning, RoPE adjustment, and host memory offloading, enabling faster and more memory-efficient operation.
Jianhong Bai, Menghan Xia, Xiao Fu +8 authors
ReCamMaster uses pre-trained text-to-video models to render dynamic scenes of input videos from novel camera trajectories, leveraging a unique video conditioning mechanism and a custom dataset created with Unreal Engine 5.
Suwhan Choi, Jaeyoon Jung, Haebin Seong +7 authors
D2E framework uses desktop interactions to pretrain embodied AI, achieving high success rates in physical manipulation and navigation 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