291 posts in total. Keep on posting.

2026

• 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：给开源 MoE RL 训练栈做一次字节级对齐

  Ashwinee Panda 执笔的这篇技术长文，把训练引擎 XoRL 和推理引擎 SGLang fork 之间的 logprob 不一致问题，从浮点舍入顺序一路拆到 MoE 路由、稀疏注意力和 MXFP4 专家权重，最终做到字节级完全一致——在 Wordle 任务上把 solve rate 从 63.9% 提到 77.4%，也顺带揭示了 Tinker 上两次训练灾难性发散的真实原因。

• 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.

• 08-14 中

  追着一个不断移动的目标：训练无关低秩压缩中的校准漂移与秩分配漂移

  阅读笔记：COLM 2026 一篇论文指出，主流的训练无关低秩压缩流水线在压缩过程中会悄悄偏离它本应对齐的模型状态，本文提出两个轻量修正来缓解这一问题。

• 08-13 EN

  ZeroLock: Breaking Update Locking in Pipeline-Parallel LLM Fine-Tuning

  ZeroLock replaces backpropagation's update-locking pipeline with per-chunk local objectives, cutting per-stage peak memory by up to 55% and lifting throughput by up to 63% over GPipe/1F1B/PipeDream, backed by the first convergence analysis for local-objective pipeline training under general chunk division.

• 08-13 中

  ZeroLock 阅读笔记：用模块化解耦打破流水线并行训练的更新锁定

  ZeroLock 用逐块本地目标替代反向传播的全链路更新依赖，单卡峰值显存最多降低约55%，吞吐最多提升约63%，并首次给出了任意分块数量下本地目标构造类算法的收敛性证明。

• 08-12 EN

  OasisKV: Scaling In-Decode KV Cache Beyond HBM with Lookahead Sparse Prefetching

  OasisKV turns speculative-decoding draft tokens into a free, one-step-ahead KV-cache prefetch signal, letting sparse decoding scale batch size and throughput up to 2.1x over dense vLLM while staying within 0.7 accuracy points of full attention.