https://yiwen-cai.github.io/notes/papers/Recent content in 论文阅读 on Yiwen CaiHugo -- gohugo.iozh-cncaiyiwen.cs@foxmail.com (Yiwen Cai)caiyiwen.cs@foxmail.com (Yiwen Cai)© 2026 Yiwen CaiTue, 04 Aug 2026 15:52:37 +0800https://yiwen-cai.github.io/notes/papers/rl-meets-llms-survey/Tue, 04 Aug 2026 00:00:00 +0000caiyiwen.cs@foxmail.com (Yiwen Cai)https://yiwen-cai.github.io/notes/papers/rl-meets-llms-survey/梳理 RL 如何贯穿 LLM 全生命周期——预训练/Mid-training、RLHF 对齐微调、RLVR 强化推理三条主线；从策略梯度、PPO 到 GRPO 的算法演进，以及 RLVR 是否真正扩展推理能力、熵坍缩与性能上限等争议。https://yiwen-cai.github.io/notes/papers/sparsec-speculative-decoding/Sat, 20 Jun 2026 00:00:00 +0000caiyiwen.cs@foxmail.com (Yiwen Cai)https://yiwen-cai.github.io/notes/papers/sparsec-speculative-decoding/针对推理语言模型长输出的 memory-bound 瓶颈，用同一模型做 self-speculative decoding——verification 阶段顺手 dump 出 attention scores 做 Top-K，作为后续 draft 的动态稀疏模式，零训练、无损、最高 2.13× 加速。https://yiwen-cai.github.io/notes/papers/flashattention/Thu, 18 Jun 2026 00:00:00 +0000caiyiwen.cs@foxmail.com (Yiwen Cai)https://yiwen-cai.github.io/notes/papers/flashattention/通过 IO-aware 的 tiling 和重计算，在不改变 attention 数学定义的前提下大幅减少 HBM 读写，实现 2-4× 加速与 5-20× 内存节省——attention 优化的基础构件。