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
Meituan LongCat Team, Bairui Wang, Bayan +129 authors
LongCat-Flash-Omni, a 560 billion parameter omni-modal model, achieves real-time audio-visual interaction through curriculum-inspired training and modality-decoupled parallelism.
Zhuoran Zhang, Tengyue Wang, Xilin Gong +4 authors
A framework decomposes modality following in multimodal large language models into relative reasoning uncertainty and inherent modality preference, providing insights into how models resolve conflicting information.
Xiangyu Fan, Zesong Qiu, Zhuguanyu Wu +6 authors
Phased DMD enhances multi-step distillation of score-based generative models by dividing SNR ranges and using progressive distribution matching, improving diversity and generative capabilities.
Jiayu Liu, Cheng Qian, Zhaochen Su +4 authors
CostBench evaluates Large Language Model agents' cost-aware planning and adaptability in response to dynamic changes, revealing significant gaps in current models' performance.
Tim R. Davidson, Adam Fourney, Saleema Amershi +3 authors
Evaluation of agent-based systems reveals a collaboration gap where solo-performing models degrade in pairings, suggesting the need for collaboration-aware evaluation and training strategies.
Mengjie Deng, Guanting Dong, Zhicheng Dou
ToolScope, an agentic framework for multimodal large language models, enhances visual question answering by integrating external tools and achieving significant performance improvements across various benchmarks.
Jie Huang, Xuejing Liu, Sibo Song +4 authors
A comprehensive analysis of multimodal Rotary Positional Embedding (RoPE) leads to the proposal of Multi-Head RoPE (MHRoPE) and MRoPE-Interleave (MRoPE-I), which improve multimodal understanding in vision-language models.
Wai-Chung Kwan, Joshua Ong Jun Leang, Pavlos Vougiouklis +3 authors
OpenSIR is a self-play framework that enables large language models to improve their reasoning abilities through open-ended problem generation and solving without external supervision.
Shaojie Zhang, Pei Fu, Ruoceng Zhang +8 authors
HyperClick enhances GUI automation by calibrating confidence and reducing overconfidence through a dual reward mechanism and spatial confidence modeling.
Zhuoning Guo, Mingxin Li, Yanzhao Zhang +3 authors
A framework combining a diagnostic benchmark, data synthesis, and a modality pyramid curriculum achieves state-of-the-art zero-shot generalization in video retrieval.
Yinhan He, Wendy Zheng, Yaochen Zhu +6 authors
SemCoT enhances CoT efficiency by optimizing token-level generation speed and preserving semantic alignment with ground-truth reasoning using contrastive training and knowledge distillation.
Jian Mu, Chaoyun Zhang, Chiming Ni +14 authors
GUI-360° is a large-scale dataset and benchmark suite for computer-using agents, addressing gaps in real-world tasks, automated data collection, and unified evaluation of GUI grounding, screen parsing, and action prediction.
Hikari Otsuka, Daiki Chijiwa, Yasuyuki Okoshi +3 authors
Theoretical analysis proves the existence of strong lottery tickets within multi-head attention mechanisms and extends the strong lottery ticket hypothesis to transformers without normalization layers.
Aditya Tanna, Pratinav Seth, Mohamed Bouadi +2 authors
TabTune is a unified library that standardizes the workflow for tabular foundation models, supporting various adaptation strategies and evaluation metrics.
Ming Li, Jike Zhong, Shitian Zhao +6 authors
TIR-Bench evaluates advanced visual reasoning capabilities in multimodal models through diverse tasks requiring tool use and chain-of-thought, demonstrating the need for genuine thinking-with-images.
Abdelaziz Bounhar, Hadi Abdine, Evan Dufraisse +5 authors
Retaining and up-weighting moderately easy problems in RLVR pipelines for LLMs reduces output verbosity without explicit length penalization.
Roman Beliy, Amit Zalcher, Jonathan Kogman +2 authors
Brain-IT uses a Brain Interaction Transformer to reconstruct images from fMRI data with high fidelity, surpassing current methods and requiring less training data.
Mohamed Bouadi, Pratinav Seth, Aditya Tanna +1 authors
Orion-MSP, a tabular in-context learning architecture, addresses limitations in current models by incorporating multi-scale processing, block-sparse attention, and a Perceiver-style memory, achieving state-of-the-art performance on diverse benchmarks.
Yifan Zhang, Zhen Qin, Quanquan Gu
Higher-order Linear Attention (HLA) is a scalable, causal, and efficient mechanism for long-context autoregressive language models, combining attention-like mixing with recurrent architecture efficiency.
Jiawei Wang, Dingyou Wang, Jiaming Hu +3 authors
Kinematify is an automated framework that synthesizes articulated objects from RGB images or textual descriptions, addressing challenges in inferring kinematic topologies and estimating joint parameters.
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