331 posts in total. Keep on posting.

2026

• 08-02 中

  FEPLB 阅读笔记：用 Copy Engine 几乎零成本地做 MoE 负载均衡

  即使不加辅助负载均衡损失，MoE 路由不均也会白白浪费 18.6% 的 GPU 时间——FEPLB 的答案不是设计更聪明的调度算法，而是发现 Hopper 架构上一直闲置的 NVLink Copy Engine 硬件。

• 08-01 EN

  Back from the Future: KV Cache Management by Counter-Causal Surprise

  A training-free KV-cache eviction rule that asks a strange but powerful question -- can the future predict this token? -- and in doing so escapes the self-reinforcing bias that quietly poisons every attention-score-based eviction method.

• 08-01 中

  来自未来的信号：用反因果惊奇度做 KV Cache 淘汰

  一个不需要训练的 KV Cache 淘汰策略，用一个反直觉的问题——这个 token 能被它后面的内容预测出来吗——绕开了几乎所有基于注意力分数的淘汰方法都存在的自我强化偏差。

• 07-31 EN

  DynaCalKV: Rethinking Fixed Head Grouping in Low-Rank KV Cache Compression

  Fixed group sizes for attention-head clustering in low-rank KV cache compression turn out to be a needless assumption — letting similarity structure itself decide the groups saves parameters, but only if you already have enough independent heads to cluster.

• 07-31 中

  DynaCalKV 阅读笔记：低秩 KV Cache 压缩里，固定分组到底有多不合理

  低秩 KV cache 压缩里对注意力头做固定大小分组这件事,原来只是个没必要的假设——让相似度结构自己决定分组,确实能省参数,但前提是这个模型本来就有足够多可聚类的独立头。

• 07-30 EN

  Libra: Taming Attention Workload Skew in Long-Context LLM Training with Bounded Sequence Pool

  Sequence packing balances memory, not attention FLOPs — Libra shows that the right fix is to keep the load-balancing pool a fixed size, tuned by the law of large numbers, rather than letting it grow with the cluster.

• 07-30 中

  Libra 阅读笔记:用有界序列池驯服长上下文训练中的注意力负载偏斜

  序列打包只能均衡内存,均衡不了注意力算力——Libra 告诉我们,真正该做的是把负载均衡池的规模固定住,用大数定律去算这个固定值该多大,而不是让它随集群一起膨胀。

• 07-29 EN

  LOCKS: Page-Local Compact Key Summaries for Efficient Long-Context Decoding

  A page-by-page eigenbasis, not a shared low-rank sketch, is what lets a resident KV-cache summary track the exact top-k oracle down to the smallest attention budgets — here is why the scope of a summary matters more than its size.

• 07-29 中

  LOCKS 阅读笔记：为什么 KV Cache 摘要要按「页」而不是按「序列」或「全局」来建

  同样大小的摘要，只要拟合的范围（scope）从全局或整段序列缩小到单个 page，选中关键 token 的能力就有质的差别——这篇笔记梳理 LOCKS 论文里这条推理链的每一环。

• 07-28 EN

  Branching Policy Optimization: Turning Sandbox Snapshots Into a Variance-Reduction Primitive

  BPO forks sibling rollouts from snapshotted mid-trajectory states instead of only the prompt, turning a sandbox's checkpoint-restore capability into a provably lower-variance advantage estimator than GRPO or RLOO.

• 07-28 中

  Branching Policy Optimization：把沙盒快照变成一个方差削减原语

  BPO 从沙盒中途的快照状态而不只是从 prompt 出发去分叉兄弟 rollout，把沙盒的 checkpoint-restore 能力变成了一个可证明比 GRPO、RLOO 更低方差的 advantage 估计器。

• 07-27 EN

  The Regression Tax: Why Adding Skills to an LLM Agent Also Breaks Tasks It Used to Solve

  A paired-trajectory study of 5,832 agent runs showing that skill libraries buy their gains at a steep, usually invisible cost: 59% of the tasks they newly solve are offset by tasks they newly break, and the breakage traces to three concrete, nameable mechanisms.