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。
Lei Huang, Xiang Cheng, Chenxiao Zhao +6 authors
Language feedback is leveraged in reinforcement learning to improve exploration efficiency and sample utilization through grouped critique aggregation and joint generation-refinement optimization.
50 篇论文 · 按点赞排序
Yinjie Wang, Xuyang Chen, Xiaolong Jin +2 authors
OpenClaw-RL framework enables policy learning from diverse next-state signals across multiple interaction modalities using asynchronous training with PRM judges and hindsight-guided distillation.
Jiyuan Wang, Chunyu Lin, Lei Sun +8 authors
RL3DEdit uses reinforcement learning with rewards from a 3D foundation model to achieve multi-view consistent 3D editing from 2D editing priors.
Boqiang Zhang, Lei Ke, Ruihan Yang +5 authors
Penguin-VL demonstrates that text-only initialized vision encoders can achieve superior performance in multimodal understanding tasks compared to traditional contrastive pretraining methods, enabling efficient deployment on resource-constrained devices.
Junjie Li, Xinrui Guo, Yuhao Wu +3 authors
Large language models struggle with maintaining narrative consistency over long-form storytelling, exhibiting predictable patterns of contradictions that this benchmark identifies and categorizes.
Fangfu Liu, Diankun Wu, Jiawei Chi +7 authors
Spatial-TTT enables streaming visual-based spatial intelligence through test-time training that adapts parameters to capture spatial evidence over long video sequences using hybrid architecture and 3D spatiotemporal convolution.
Yuanyuan Gao, Hao Li, Yifei Liu +14 authors
Holi-Spatial presents the first fully automated, large-scale, spatially-aware multimodal dataset constructed from raw video inputs, supporting multi-level spatial supervision for 3D scene understanding and spatial reasoning tasks.
Shuo Yang, Haocheng Xi, Yilong Zhao +10 authors
Flash-kmeans enables efficient online k-means clustering on GPUs through novel kernel-level optimizations that eliminate I/O bottlenecks and reduce atomic write contention.
Zorik Gekhman, Roee Aharoni, Eran Ofek +3 authors
Reasoning in large language models enhances parametric knowledge recall through computational buffer and factual priming mechanisms, though it carries risks of hallucination that can be mitigated by prioritizing accurate reasoning trajectories.
Yushi Bai, Qian Dong, Ting Jiang +5 authors
IndexCache reduces sparse attention computation in large language models by reusing top-k token selections across layers, achieving significant speedups with minimal quality loss.
Łukasz Borchmann, Jordy Van Landeghem, Michał Turski +12 authors
MADQA benchmark evaluates multimodal agents' strategic reasoning capabilities through diverse PDF document questions, revealing gaps between human-level accuracy and efficient reasoning performance.
Junyi Zhang, Charles Herrmann, Junhwa Hur +5 authors
LoGeR enables long-term 3D video reconstruction by combining bidirectional priors with a hybrid memory system that includes parametric Test-Time Training and non-parametric sliding window attention mechanisms.
Bingxiang He, Yuxin Zuo, Zeyuan Liu +18 authors
Unsupervised reinforcement learning with verifiable rewards faces fundamental limitations in scaling large language model training due to inherent convergence properties and confidence-correction misalignment, though external methods show promise for overcoming these barriers.
Zongxia Li, Hongyang Du, Chengsong Huang +8 authors
MM-Zero enables zero-data self-evolution of vision-language models through a multi-role framework with proposer, coder, and solver components trained via group relative policy optimization.
Lijiang Li, Zuwei Long, Yunhang Shen +6 authors
Omni-Diffusion introduces the first any-to-any multimodal language model based on mask-based discrete diffusion models, unifying text, speech, and image processing in a single framework.
Changyao Tian, Danni Yang, Guanzhou Chen +26 authors
InternVL-U is a 4-billion parameter unified multimodal model that combines advanced visual generation with robust semantic understanding through specialized modular design and reasoning-centric data synthesis.
Parishad BehnamGhader, Vaibhav Adlakha, Fabian David Schmidt +3 authors
LLM2Vec-Gen introduces a self-supervised method for text embedding that represents model responses through trainable special tokens, achieving superior performance on MTEB while reducing harmful content and improving reasoning capabilities.
Linxin Song, Jieyu Zhang, Huanxin Sheng +6 authors
Video-execution reward modeling enables scalable evaluation of computer-using agents by predicting task success from user instructions and execution videos, outperforming proprietary models across multiple operating systems.
Yaoqi Ye, Yiran Zhao, Keyu Duan +4 authors
In-Context Reinforcement Learning (ICRL) enables large language models to effectively use external tools through a reinforcement learning-only framework that eliminates the need for costly supervised fine-tuning by gradually reducing in-context examples during training.
Chen Yueh-Han, Robert McCarthy, Bruce W. Lee +5 authors
Chain-of-thought controllability measures how effectively models can be constrained to follow reasoning steps, with findings showing significantly lower controllability in reasoning versus output generation, and varying impacts from model size, training methods, and task complexity.
Dongwon Kim, Gawon Seo, Jinsung Lee +2 authors
CompACT, a discrete tokenizer that reduces observation encoding from hundreds to 8 tokens, enables faster and more efficient world model planning for real-time control applications.
Shijia Liao, Yuxuan Wang, Songting Liu +11 authors
Fish Audio S2 is an open-source text-to-speech system with multi-speaker capabilities, multi-turn generation, and instruction-following control through natural-language descriptions, utilizing a multi-stage training approach and production-ready inference engine.
Haodong Li, Chunmei Qing, Huanyu Zhang +10 authors
CoCo is a code-driven framework for text-to-image generation that uses executable code for precise spatial layout planning and structured image creation, achieving superior performance over traditional CoT-based methods.
Xizhong Yang, Haotian Zhang, Huiming Wang +1 authors
Large reasoning models enhance prediction accuracy through test-time scaling techniques that generate multiple candidate responses, with the proposed DistriVoting method utilizing distributional priors and confidence scores to improve answer selection.
Yucheng Wang, Zedong Wang, Yuetong Wu +2 authors
CARE-Edit introduces a condition-aware routing mechanism that dynamically allocates diffusion model computation to specialized experts for improved contextual image editing tasks.
Qin Guo, Tianyu Yang, Xuanhua He +5 authors
WildActor generates consistent human videos with full-body identity preservation across varying viewpoints and motions using a large-scale dataset and novel attention mechanisms.
Tianhao Chen, Xin Xu, Lu Yin +4 authors
Progressive Residual Warmup (ProRes) improves transformer language model pretraining by implementing gradual layer activation that stabilizes training and accelerates convergence.
Guanyu Jiang, Zhaochen Su, Xiaoye Qu +1 authors
XSkill is a dual-stream framework that enables multimodal agents to learn from experience and skills through visual grounding, improving tool use efficiency and reasoning capabilities in open-ended environments.
Songlin Yang, Zhe Wang, Xuyi Yang +7 authors
ShotVerse enables cinematic video generation through a plan-then-control framework that decouples textual planning from camera control using vision-language models and trajectory alignment.
Yujie Wei, Xinyu Liu, Shiwei Zhang +12 authors
DreamVideo-Omni is a unified framework for video synthesis that enables precise multi-subject identity control and multi-granularity motion manipulation through a two-stage training approach combining condition-aware 3D positional embeddings, hierarchical motion injection, and latent identity reward learning.
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