Trends · 研究趋势
数据来自 Hugging Face 论文的 AI 提取关键词,按月统计研究方向的增长与热度。
Yuntai Bao, Qinfeng Li, Guoqing Jiang +5 authors
On-policy distillation of reinforcement learning improves reasoning across model scales with predictable scaling laws, showing that smaller teachers can outperform larger ones at matched scores and that students often exceed their teachers.
Shi-Yu Tian, Yong Liu, Ximan Liu +10 authors
Road MOPD reduces gradient conflicts in multi-teacher distillation by using cross-teacher agreement for token-level credit assignment and auxiliary gradients to constrain conflicting updates.
Guanting Dong, Xinyu Yang, Jiajie Jin +9 authors
Agent-Reflex trains interactive agents to reflect and recover from noisy tool feedback and user corrections through skill discovery, supervised fine-tuning, and on-policy optimization.
Bingxi Hou, Guochao Jiang, Guofeng Quan +4 authors
On-policy distillation with strict token alignment outperforms broader cross-tokenizer supervision by avoiding weak or conflicting gradient signals.
Hao Li, MeiJia Chen, Weijie Ren +4 authors
RIDE improves on-policy distillation by extrapolating reinforcement-learning-induced representation shifts directly in hidden states, yielding stable gains over RL-trained teachers.
Tencent Robotics X, HY Vision Team, Xumin Yu +19 authors
HY-Embodied-0.5 is a foundation model family for embodied agents featuring Mixture-of-Transformers architecture and iterative post-training for enhanced visual perception and reasoning capabilities.
Chenxu Yang, Chuanyu Qin, Qingyi Si +7 authors
RLSD combines reinforcement learning with verifiable rewards and self-distillation to achieve stable training with fine-grained updates and reliable policy direction from environmental feedback.
Yuxiao Yang, Tianrun Yu, Shangzhe Li +6 authors
On-policy distillation causes student responses to grow excessively long due to termination-token mismatch with teachers, which is reduced by treating equivalent end-of-sequence tokens as shared stopping actions.
Xinlei Yu, Gen Li, Qingyi Si +13 authors
DOPD addresses privilege illusion in on-policy distillation by dynamically routing token-level supervision between teacher and student policies based on advantage gaps and probabilities, improving capability transfer in large and vision-language models.
Yaxuan Li, Yuxin Zuo, Bingxiang He +8 authors
On-policy distillation dynamics in large language models depend on compatible thinking patterns between teacher and student models, with successful distillation characterized by alignment on high-probability tokens and requiring teachers to provide novel capabilities beyond student training data.
Jinyang Wu, Shuo Yang, Zhengxi Lu +8 authors
SEED improves long-horizon agent training by extracting hindsight skills from completed trajectories and distilling them into dense token-level supervision alongside outcome-based reinforcement learning.
Yuchao Gu, Guian Fang, Yuxin Jiang +4 authors
AnyFlow introduces a novel any-step video diffusion distillation framework that improves upon consistency distillation by optimizing full ODE sampling trajectories through flow-map transition learning and backward simulation techniques.
Zhen Fang, Wenxuan Huang, Yu Zeng +8 authors
Flow-OPD addresses limitations in Flow Matching text-to-image models through a two-stage alignment approach combining on-policy distillation and manifold anchor regularization, achieving significant improvements in generation quality and alignment metrics.
Zixuan Fu, Bingxiang He, Yuxin Zuo +10 authors
On-policy distillation improves over hundreds of steps from a single query by rapidly covering teacher states, yet student alignment remains slow, indicating the method is algorithm-starved rather than data-starved.
Yi Ding, Ruqi Zhang
On-policy distillation relies mainly on suppressing low-probability tokens rather than teacher guidance, motivating a supervision-free entropy-adaptive method that substantially improves reasoning performance.
Bingnan Li, Haozhe Wang, Haozhong Xiong +5 authors
On-policy diffusion distillation under classifier-free guidance suffers from branch-level error compensation, which is resolved by separately matching positive predictions and guidance directions.
Lei Bai, Jiaqi Cao, Chiyu Chen +122 authors
Intern-S2-Preview is a scientific agentic foundation model series that integrates multimodal pre-training, multi-task reinforcement learning, and memory-augmented extensions to support long-horizon scientific reasoning and forecasting.
Zhennan Shen, Yanshu Li, Qingyu Yin +6 authors
On-policy distillation exhibits distinct parameter space dynamics characterized by relaxed off-principal updates and subspace locking, forming a unique geometric pattern separate from supervised fine-tuning and reinforcement learning with verifiable rewards.
Wenkai Yang, Weijie Liu, Ruobing Xie +3 authors
On-policy distillation is extended through a generalized framework that introduces flexible reference models and reward scaling factors, demonstrating improved performance through reward extrapolation and reward correction techniques.
Ethan Chern, Zhulin Hu, Bohao Tang +4 authors
Real-time multimodal video generation via diffusion is enabled by an improved distillation approach with multimodal conditioning and optimized scheduling, reducing inference latency while maintaining quality for interactive systems.
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