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
Qifan Yu, Xinyu Ma, Zhijian Zhuo +7 authors
SPARKLING is a framework for mid-stage width expansion in deep learning models that maintains signal preservation and breaks symmetry to stabilize training and reduce computational costs.
Zehong Ma, Ruihan Xu, Shiliang Zhang
PixelGen is a pixel-space diffusion framework that uses perceptual supervision through LPIPS and DINO-based losses to generate high-quality images without requiring VAEs or latent representations.
Tianhe Wu, Ruibin Li, Lei Zhang +1 authors
A novel distillation framework called DP-DMD is introduced that preserves sample diversity in text-to-image generation by separating the roles of distilled steps, using v-prediction for diversity and standard DMD loss for quality refinement without additional computational overhead.
Dylan Zhang, Yufeng Xu, Haojin Wang +2 authors
Post-training of reasoning large language models can be improved by correcting distribution mismatches between supervised fine-tuning and reinforcement learning stages through importance sampling reweighting of the SFT loss.
Yu Liang, Zhongjin Zhang, Yuxuan Zhu +10 authors
ReSID presents a novel recommendation-native framework that improves sequential recommendation by learning predictive item representations and optimizing quantization for information preservation and sequential predictability.
Shuang Sun, Huatong Song, Lisheng Huang +11 authors
A Docker-free framework replaces physical execution environments with learned surrogates for training software engineering agents, enabling efficient training and test-time scaling without costly container setup.
Woosung Koh, Sungjun Han, Segyu Lee +2 authors
Visual world models for mobile GUI agents are improved through renderable code generation using vision-language models, achieving better performance with reduced model size compared to existing approaches.
Pengyu Wang, Benfeng Xu, Licheng Zhang +4 authors
WildGraphBench evaluates GraphRAG performance in realistic scenarios using Wikipedia's structured content to assess multi-fact aggregation and summarization capabilities across diverse document types.
Yu Bai, MingMing Yu, Chaojie Li +3 authors
EgoActor is a unified vision-language model that translates high-level instructions into precise humanoid robot actions through integrated perception and execution across simulated and real-world environments.
Huatong Song, Lisheng Huang, Shuang Sun +10 authors
SWE-Master presents a reproducible framework for developing software engineering agents through systematic optimization across multiple stages of agent development, achieving superior performance on software task resolution benchmarks.
Seanie Lee, Sangwoo Park, Yumin Choi +6 authors
ThinkSafe is a self-aligned framework that enhances safety in large reasoning models through lightweight refusal steering and fine-tuning on self-generated responses, maintaining reasoning performance while reducing computational costs.
Penghui Qi, Xiangxin Zhou, Zichen Liu +4 authors
DPPO addresses limitations in PPO for LLM fine-tuning by replacing ratio clipping with direct policy divergence constraints, improving training stability and efficiency.
Yuezhou Hu, Harman Singh, Monishwaran Maheswaran +10 authors
Residual Context Diffusion (RCD) enhances diffusion large language models by recycling discarded token information through contextual residuals, improving accuracy with minimal computational overhead.
Shuo Chen, Cong Wei, Sun Sun +5 authors
Context Forcing addresses student-teacher mismatch in long video generation by using a long-context teacher to guide long-rollout students through a Slow-Fast Memory architecture that extends context length beyond 20 seconds.
Yinjie Wang, Tianbao Xie, Ke Shen +2 authors
RLAnything enhances reinforcement learning for LLMs and agents through dynamic model optimization and closed-loop feedback mechanisms that improve policy and reward model training.
Andong Chen, Wenxin Zhu, Qiuyu Ding +3 authors
Thinking with Comics emerges as an effective visual reasoning approach that bridges images and videos by leveraging comic structures for improved multimodal reasoning efficiency and performance.
Haocheng Xi, Shuo Yang, Yilong Zhao +13 authors
Quant VideoGen addresses KV cache memory limitations in autoregressive video diffusion models through semantic-aware smoothing and progressive residual quantization, achieving significant memory reduction with minimal latency impact.
Hang Yan, Xinyu Che, Fangzhi Xu +7 authors
Test-Time Improvement (TTI) in autonomous LLM agents involves iterative environmental interaction that enhances performance, but current evaluation methods inadequately capture task optimization efficiency and memory utilization.
Fanmeng Wang, Haotian Liu, Guojiang Zhao +2 authors
ReGuLaR introduces a variational auto-encoding framework that compresses reasoning processes into latent space while maintaining performance through image-rendered explicit reasoning chains for guidance.
Zhiyuan Yao, Yi-Kai Zhang, Yuxin Chen +7 authors
CoBA-RL adapts rollout budget allocation for LLM training by evaluating sample training value and optimizing resource distribution through a capability-oriented value function and greedy strategy.
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