291 posts in total. Keep on posting.

2026

• 08-12 中

  OasisKV 阅读笔记：把投机采样的草稿 token 变成免费的 KV Cache 预取信号

  OasisKV 把投机采样早已算出来的草稿 token 直接拿来当 KV cache 预取的『提前一步』信号，让稀疏解码把批大小和吞吐量提升到稠密 vLLM 的 2.1 倍，同时精度损失控制在 0.7 个百分点以内。

• 08-11 EN

  CVPO: Curriculum-Guided Value-Variance Policy Optimization for LLM Reasoning

  CVPO turns the variance of a value model's token-level estimates into a signal that both stabilizes correct trajectories and encourages exploration on incorrect ones, then layers a Bayesian curriculum on top to fight difficulty drift during long RL runs.

• 08-11 中

  CVPO 阅读笔记：用价值方差驱动的课程式策略优化

  CVPO 把价值模型对同一条轨迹内部的方差波动，转化成一个能区分对错、动态调节探索强度的优势加权信号，再叠加一套基于贝叶斯后验的难度课程机制来对抗训练中的'难度漂移'。

• 08-10 EN

  GRAFT: Global Optimization and Inference-Time Region Grafting for Agentic Workflows

  GRAFT keeps a globally-searched agentic workflow frozen as structural scaffolding, then locally swaps out only the operators that a label-free quality signal flags as weak for the current input — beating supernet- and search-based workflow optimizers without any training or per-query global re-search.

• 08-10 中

  阅读笔记：GRAFT——给智能体工作流做「全局定骨架、局部动手术」的推理时嫁接

  GRAFT 把离线搜好的智能体工作流当作冻结的结构骨架，推理时只用无标签质量信号局部替换表现不佳的算子——不训练、不重搜全局，就在多个基准上超过 supernet 与搜索式工作流优化方法。

• 08-09 EN

  Bole: Efficient Tree Speculation for Hybrid-Attention Language Models

  Bole reformulates tree speculative decoding for hybrid full-attention/linear-attention models as one closed-form parallel solve, cutting transient recurrent-state memory by up to 99x and linear-attention verification time by up to 7.7x.

• 08-09 中

  Bole 阅读笔记：为混合注意力大模型重新设计树状投机解码

  Bole 把混合全注意力/线性注意力模型的树状投机验证重写成一次并行闭式求解，让线性注意力验证提速最高7.7倍、瞬时状态显存开销降低最高99倍。

• 08-08 EN

  QEvict: Recoverable Quantized KV Eviction for Attention-Drift-Robust Long-Context Decoding

  QEvict replaces binary KV-cache eviction with a three-tier recoverable hierarchy — full precision, INT2, and evicted — so windows that look unimportant now can be dequantized and promoted back once the model's attention returns to them.

• 08-08 中

  QEvict 阅读笔记：把 KV Cache 淘汰从「一锤子买卖」变成可逆的三层调度

  QEvict 用全精度/INT2 可恢复量化/永久淘汰三层结构取代传统二元淘汰，让曾经「看起来不重要」的 KV 窗口在注意力回流时能被反量化并重新晋升，而不是永久丢弃。

• 08-07 EN

  SALT: Subspace-Aligned Centroid-Residual Training for Efficient Ultra-LoRA Serving

  SALT pins one shared, subspace-aligned domain centroid in GPU memory and swaps only a rank-1 residual per tenant, recovering high-rank LoRA accuracy while cutting per-adapter memory by up to 16x.

• 08-07 中

  SALT 阅读笔记：子空间对齐质心-残差训练如何让超低秩 LoRA 服务成为可能

  SALT 把一份子空间对齐的质心永久驻留在显存里，每个租户只需要热插拔一个 r≤2 的残差，就能追回接近满秩 LoRA 的精度，同时把单适配器显存开销压缩到原来的 1/16。

• 08-06 EN

  Fine-Grained Compute-Communication Overlap for MoE: Hiding All-to-All Behind GEMM with Tile-Level Signaling

  A producer-consumer GEMM/all-to-all co-design that hides the MoE return communication behind expert compute using per-tile device-resident signals, a remote-owner-aligned row layout, and disjoint SM partitions — up to 2.74x MoE-layer speedup on 4 A100 GPUs without touching NCCL or fusing kernels.