331 posts in total. Keep on posting.

2026

• 08-19 EN

  KV Cache Compression Through the Lens of Transform Coding: What AATC Gets Right About Attention-Aware Quantization

  AATC reframes KV cache quantization as a classical transform-coding problem, deriving a closed-form attention-aware distortion measure that unifies token eviction, rotation, and mixed-precision methods, then allocates bits by reverse waterfilling to reach near-lossless 5.8x compression.

• 08-19 中

  AATC 阅读笔记：用变换编码的视角重新理解 KV Cache 量化

  AATC 把 KV Cache 量化重新表述为经典变换编码问题，推导出一个闭式的注意力感知失真度量，统一了驱逐、旋转、混合精度等多条技术路线，再用逆水填算法分配比特，在约5.8倍压缩比下做到近乎无损。

• 08-18 EN

  GCPO: Why Some GRPO Updates Quietly Break Your Model, and How to Structurally Forbid Them

  GCPO diagnoses a geometric warning sign in rollout RL — transient overlap between policy updates and a pretrained model's dominant singular subspaces — then structurally forbids it with a projected low-rank parameterization, beating GRPO/GSPO/DAPO/GMPO on accuracy, capability retention, training stability, and response-length control.

• 08-18 中

  GCPO 阅读笔记：为什么某些 GRPO 更新会悄悄搞垮模型，以及如何从结构上禁止它们

  GCPO 发现了 rollout RL 中一个几何层面的预警信号——策略更新与预训练模型主导奇异子空间之间的瞬时重叠——并用一种投影式低秩参数化方法把它从结构上彻底禁止，在准确率、能力保持、训练稳定性和输出长度控制上全面超过 GRPO/GSPO/DAPO/GMPO。

• 08-17 EN

  AgentRewind: Giving Long-Horizon LLM Agents an Undo Button That Actually Undoes Things

  AgentRewind adds a runtime recovery layer that checkpoints and jointly restores both agent context and workspace filesystem state, letting agents rewind past an unrecoverable error rather than only patching forward — improving task success by 25.6 points on a new long-horizon benchmark, MettleBench.

• 08-17 中

  AgentRewind 阅读笔记：给长程 LLM Agent 装一个真正能撤销的悔棋键

  AgentRewind 在 agent 与环境之间加了一层运行时恢复层，把 agent 的对话上下文和工作区文件系统一起打包成检查点，让 agent 能悔棋回到早先某个节点，而不是只能在已经出错的轨迹上继续硬撑——在新提出的长程基准 MettleBench 上把任务成功率提升了 25.6 个百分点。

• 08-17 中

  0 train-infer mismatch

  说明如何在开源 MoE 上把训练与推理的 logprob 对齐到 bitwise-identical，并在 XoRL 中以 0 train-infer mismatch 进行强化学习。

• 08-16 EN

  vToken: Turning Token-Level KV Eviction into Actually Reusable GPU Memory

  vToken adds a thin token-level virtualization layer between KV eviction policies and PagedAttention's block manager, closing a granularity mismatch that traps 40-60% of allocated KV memory in partially-live blocks — reclaiming 27.2-72.3% of retained blocks and doubling feasible concurrency under memory pressure.

• 08-16 中

  vToken 阅读笔记：如何把 Token 级 KV 淘汰真正变成可用显存

  vToken 在 KV 淘汰策略与 PagedAttention 块管理器之间插入一层轻量级的 token 级虚拟化，弥合了导致 40%-60% 已分配 KV 内存被'困'在半空块中的粒度错配问题——回收了 27.2%-72.3% 的保留块，并在显存压力下将可行并发数提升近一倍。

• 08-15 EN

  SwiftQK: When Normalization Becomes the Bottleneck of Tensor Parallelism

  SwiftQK replaces full-vector Q/K activation exchange under tensor parallelism with scalar partial-sum aggregation, overlapping the residual peer-to-peer reduction with independent computation inside a deadlock-safe persistent kernel — cutting QK-Norm latency by up to 93.9% and end-to-end TPOT by 29.5%.

• 08-15 中

  SwiftQK：当归一化层成为张量并行的瓶颈

  SwiftQK 用标量偏和聚合取代张量并行下 Query-Key 归一化所需的整向量 All-Gather 通信，并在一个避免死锁的持久化 kernel 内把剩余的点对点同步延迟与计算重叠，QK-Norm 延迟最多降低 93.9%，端到端 TPOT 平均降低 29.5%。

• 08-14 EN

  Chasing a Moving Target: Calibration and Rank-Allocation Drift in Training-Free Low-Rank LLM Compression

  A close read of a COLM 2026 paper showing that training-free low-rank LLM compression pipelines silently drift away from the model they think they are compressing, and how two lightweight corrections close much of that gap.