349 posts in total. Keep on posting.

2026

• 09-03 EN

  HIERA: Teaching LLM Agents Where to Optimize, Not Just How — Workload-Aware Planning Across GPU Implementation Spaces

  A technical deep dive into HIERA, a training-free LLM-agent framework that treats the choice of implementation space (custom CUDA vs. libraries vs. PyTorch operators) itself as a search decision, and why that reframing beats deeper-but-narrower agentic CUDA search.

• 09-03 中

  HIERA 阅读笔记：让 LLM Agent 先学会「在哪一层优化」，再谈「怎么优化」——GPU 算子实现空间的分层规划

  深入解读 HIERA：一个训练无关（training-free）的 LLM Agent 框架，把「用原生 CUDA、还是调库、还是直接用 PyTorch 算子」这个实现空间的选择，本身当作一个需要搜索/规划的决策，而不是事先固定死。

• 09-02 EN

  Verification-Aware Training: Teaching Speculative Decoding Drafters to Care About the Rejection Point

  A technical review of Verification-Aware Training (VAT), a plug-in objective for speculative-decoding drafters that supervises acceptance outcomes directly and re-anchors per-position loss weighting to each sample's first-rejection point, improving EAGLE-3 and DFlash acceptance length by up to 11.4% with no change to inference.

• 09-02 中

  验证感知训练（VAT）：让投机解码的草稿模型学会关心「拒绝点」

  这篇阅读笔记介绍 Verification-Aware Training（VAT）——一种即插即用的投机解码草稿模型训练目标，通过直接监督验证结果并把逐位置权重重新锚定到每个样本自己的首次拒绝点，在完全不改变推理流程的前提下，把 EAGLE-3 和 DFlash 的接受长度提升最多 11.4%。

• 09-01 EN

  RRC: Unlocking Generative Reward Models in LLM Reinforcement Learning via Ranking-Based Reward Construction

  A close technical read of RRC, a reward-construction scheme that stops forcing generative reward models to output scalar scores and instead builds RL rewards directly from pairwise-ranking win counts, closing most of the gap between generative reward models' strong ranking ability and their weak RL performance.

• 09-01 中

  RRC 阅读笔记：如何真正把生成式奖励模型的判卷能力用到 RL 里

  生成式奖励模型（GRM）在两两排序任务上明显强于判别式奖励模型（DRM），但接入 GRPO 之后收益却大幅缩水。这篇笔记梳理 RRC 提出的排序式奖励构造方法：不再逼 GRM 输出一个标量概率，而是让它只做「谁更好」的二元判断，再用胜场数把这些判断汇聚成分数。

• 08-31 EN

  ContextPilot: Teaching LLM Agents to Manage Their Own Context with Fine-Grained RL

  A deep technical read of ContextPilot, a Tencent/Tsinghua framework that extends agent context-editing toolsets with planning, long-term memory and soft offloading, then trains them with a snapshot-level, variance-reduced credit assignment scheme that beats trajectory-level RL by 3-5 points on long-context QA and deep search.

• 08-31 中

  ContextPilot 阅读笔记：让 Agent 用细粒度强化学习学会自己管理上下文

  细读 ContextPilot——腾讯优图/清华团队提出的一套让 Agent 自主编辑工作上下文的方法：扩充了规划、长期记忆、软卸载工具集，并设计了一套方差更小的片段级信用分配训练方案，在长上下文问答和深度搜索任务上比已有方法提升 2-5 分。

• 08-30 EN

  VPP: Virtual Pipeline Parallelism for Efficient Chunked Prefill in Long-Context LLM Inference

  A close technical read of VPP, a fixed-chunk-size pipeline layout technique that folds prefill chunks back across pipeline ranks in a V-shape to absorb the near-linear latency growth caused by causal attention, cutting bubble ratio from 6.4% to 0.1% on a 512K-token DeepSeek-V3.1 workload.

• 08-30 中

  VPP：面向长上下文 LLM 推理的虚拟流水线并行 —— 用 V 形折返调度吸收分块预填充的延迟增长

  这是一篇关于 VPP 的阅读笔记：一种保持分块预填充 chunk 大小固定、转而重塑流水线布局的方法，通过 V 形折返调度把因果注意力带来的近似线性延迟增长直接吸收进并行执行，在 512K token 的 DeepSeek-V3.1 工作负载上把流水线气泡比例从 6.4% 降到 0.1%。

• 08-28 EN

  LowRankArena: What Happens When You Force SVD Compression Papers to Play by the Same Rules

  A technical review of LowRankArena, a standardized evaluation platform that re-audits five popular SVD-based LLM compression methods under matched budgets and finds that most published rankings do not survive standardization.

• 08-28 中

  LowRankArena 阅读笔记：把 SVD 压缩论文放到同一套规则下会发生什么

  这篇笔记详细拆解 LowRankArena——一个把五种主流 SVD 低秩压缩方法放到统一预算和统一推理链路下重新评测的标准化平台，发现大多数已发表的排名结论在标准化之后并不成立。