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
StepFun, Bin Wang, Bojun Wang +196 authors
Step-3, a 321B-parameter VLM, optimizes decoding costs through Multi-Matrix Factorization Attention and Attention-FFN Disaggregation, achieving high efficiency and throughput compared to other models.
Yin Xie, Kaicheng Yang, Xiang An +9 authors
RICE, a novel method using a Region Transformer and region cluster discrimination loss, enhances region-level visual and OCR capabilities, outperforming previous methods in tasks like segmentation and dense detection.
Asaf Yehudai, Lilach Eden, Yotam Perlitz +2 authors
CLEAR is an interactive, open-source package for LLM-based error analysis that provides detailed feedback and visualizations to understand specific performance issues.
Junzhe Li, Yutao Cui, Tao Huang +4 authors
MixGRPO, a novel framework integrating SDE and ODE, enhances flow matching models for image generation by optimizing only within a sliding window, improving efficiency and performance.
Guangchen Lan, Sipeng Zhang, Tianle Wang +7 authors
MaPPO, a framework for preference optimization, enhances alignment of large language models with human preferences by integrating prior reward knowledge into a Maximum a Posteriori objective, improving performance across various benchmarks.
Shaohan Li, Hao Yang, Min Chen +1 authors
An implicit two-stage training method with separate encoders, decoders, and a translator, enhanced by self-attention, improves weather prediction accuracy.
Shuangkang Fang, I-Chao Shen, Takeo Igarashi +5 authors
NeRF-GS combines Neural Radiance Fields and 3D Gaussian Splatting to enhance 3D scene representation and performance through joint optimization and shared spatial information.
Chong Xia, Shengjun Zhang, Fangfu Liu +3 authors
ScenePainter addresses semantic drift in 3D scene generation by aligning scene-specific priors with current scene comprehension using a hierarchical graph structure.
Peng Cai, Qiang Li, Kaicheng Yang +6 authors
ForCenNet, a foreground-centric network, improves document image rectification by enhancing foreground element distinction and using curvature consistency loss, achieving state-of-the-art results on real-world benchmarks.
Kejia Zhang, Keda Tao, Zhiming Luo +3 authors
TARS, a token-adaptive preference strategy, improves multimodal large language models by reducing hallucinations through min-max optimization under semantic constraints.
Shaofei Cai, Zhancun Mu, Haiwen Xia +3 authors
Reinforcement Learning enhances generalizable spatial reasoning and interaction in 3D environments through cross-view goal specification and automated task synthesis, achieving zero-shot generalization and improved interaction success rates.
Xiao Fang, Minhyek Jeon, Zheyang Qin +5 authors
A multi-stage, multi-modal knowledge transfer framework using fine-tuned latent diffusion models improves vehicle detection in aerial imagery across different domains.
Yonghyun Kim, Wayne Chi, Anastasios N. Angelopoulos +5 authors
Music Arena is an open platform for real-time human preference evaluation of text-to-music models, featuring a standardized protocol, detailed feedback, and privacy guarantees.
Renhang Liu, Chia-Yu Hung, Navonil Majumder +5 authors
JAM, a flow-matching-based model, introduces word-level timing and duration control in song generation and uses Direct Preference Optimization to enhance audio quality, outperforming existing models in music-specific attributes.
Haichuan Hu, Xiaochen Xie, Quanjun Zhang
Repair-R1 enhances automated program repair by integrating test cases into the training phase and prioritizing test generation before repair, improving repair success, test generation success, and test coverage.
Kaining Ying, Henghui Ding, Guanquan Jie +1 authors
Omnimodal Referring Audio-Visual Segmentation (OmniAVS) and Omnimodal Instructed Segmentation Assistant (OISA) advance audio-visual segmentation by integrating complex multimodal expressions and leveraging MLLM for reasoning-based segmentation.
Risa Shinoda, Nakamasa Inoue, Iro Laina +2 authors
AnimalClue is a large-scale dataset for species identification from indirect evidence images, addressing challenges in classification and segmentation tasks.
Afshin Khadangi, Amir Sartipi, Igor Tchappi +2 authors
RLDP, a deep reinforcement learning framework, optimizes differentially private training by dynamically adjusting gradient clipping and noise, enhancing model utility and speed while maintaining privacy.
Xingjian Zhang, Siwei Wen, Wenjun Wu +1 authors
The EDGE-GRPO algorithm addresses the advantage collapse problem in Group Relative Policy Optimization by incorporating entropy-driven advantage and guided error correction, enhancing response diversity and training signal.
Shanghai AI Lab, Xiaoyang Chen, Yunhao Chen +34 authors
The report assesses frontier risks of AI models using the E-T-C analysis framework, categorizing risks into green, yellow, and red zones based on intolerable thresholds and early warning indicators.
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