1
Rethinking the Divergence Regularization in LLM RL
重新审视LLM RL中的散度正则化，提出改进方案，提升模型对齐效率
arXiv:2606.09821v1 Announce Type: new Abstract: Reinforcement learning (RL) has become a key component of post-training large language models (LLMs). …
2
Approximating $f$-Divergences with Rank Statistics
提出用秩统计量近似f-散度的新方法，理论简洁且计算高效，为信息论和机器学习提供实用工具。
arXiv:2601.22784v2 Announce Type: replace-cross Abstract: We introduce a rank-statistic approximation of $f$-divergences that avoids explicit density-…
3
Calibeating for general proper losses: A Bregman divergence approach
将校准方法从Brier与对数损失推广至通用适当损失函数，基于Bregman散度与遗憾最小化框架创新。
arXiv:2605.17269v1 Announce Type: new Abstract: This work introduces a general framework for calibeating based on regret minimization. As compared to …
4
Decoupling KL and Trajectories: A Unified Perspective for SFT, DAgger, Offline RL, and OPD in LLM Distillation
一篇统一SFT、DAgger、离线RL和OPD视角的LLM蒸馏论文，解耦KL与轨迹，为模型优化提供新理论框架。
arXiv:2605.16826v1 Announce Type: new Abstract: Knowledge distillation is central to LLM post-training, yet its design space remains poorly understood…
5
AMiD: Knowledge Distillation for LLMs with $\alpha$-mixture Assistant Distribution
AMiD提出了一种统一的知识蒸馏框架，通过α-混合辅助分布系统性地桥接了教师与学生的容量鸿沟，解决了因高维输出近零概率引发的训练不稳定问题——这是LLM蒸馏中关键却长期碎片化的挑战。
arXiv:2510.15982v3 Announce Type: replace-cross Abstract: Autoregressive large language models (LLMs) have achieved remarkable improvement across many…