331 posts in total. Keep on posting.

2026

• 07-27 中

  Regression Tax（回归税）：给 LLM Agent 加 Skill 为什么既能帮忙又能拆台

  一篇基于 5832 次配对运行的实证研究：Agent Skill 库带来的收益里，有 59% 会被它自己制造的新失败悄悄抵消，而这种破坏可以归因到三种可以在执行轨迹里指认出来的具体机制。

• 07-26 EN

  KV-Fold: Turning the KV Cache Into a Left Fold for Long-Context Inference

  A training-free protocol that lets a frozen transformer carry its own KV cache across chunks like a left-fold accumulator, holding onto exact long-range facts where sliding-window streaming quietly forgets them.

• 07-26 中

  KV-Fold：把 KV 缓存当作长上下文推理的左折叠累加器

  一种无需训练的推理协议：让冻结的 transformer 像函数式编程里的左折叠一样，把 KV 缓存作为累加器在 chunk 间传递，在滑窗流式方法悄悄遗忘的地方保持精确的长距离事实检索。

• 07-25 EN

  When Does Adding More Agents Actually Help? An Information-Bottleneck Answer

  This paper shows that multi-agent LLM systems only beat single-agent systems when their inter-worker relay messages compress away irrelevant context without discarding what the next worker actually needs, formalizing the trade-off as an information bottleneck controlled by a capability-dependent parameter.

• 07-25 中

  多智能体到底什么时候才有用？一个信息瓶颈视角的答案

  本文证明多智能体 LLM 系统只有当中间转发消息（relay）能压缩掉无关上下文、同时又不丢失下游真正需要的信息时才能超越单智能体系统，并把这个权衡形式化为一个由模型能力决定权重的信息瓶颈问题。

• 07-24 EN

  SVD-Surgeon: Bringing Optimal Brain Surgeon to the Singular-Value Basis

  SVD-Surgeon treats the singular values of an SVD-compressed weight matrix as first-class parameters and derives a closed-form, training-free update that repairs the survivors after truncation, turning a badly-degraded low-rank approximation into a much more accurate one with a single linear solve.

• 07-24 中

  SVD-Surgeon：把「最优脑外科手术」搬进奇异值空间

  SVD-Surgeon把SVD压缩后保留下来的奇异值当作一等公民参数，推导出一个闭式的、免训练的修正公式：在裁剪掉一部分奇异值之后，让剩下的奇异值主动「移动」去吸收被删除部分带来的损失，用一次线性求解就能把原本严重劣化的低秩近似修复到更精确的水平。

• 07-23 EN

  PHOENIX: Recovering LLM Training in 40 Seconds Instead of Restarting the Whole Job

  PHOENIX replaces the checkpoint-restart cycle for large-scale LLM training with per-step in-memory replication and online topology repair, delivering zero measurable overhead during normal training and under-40-second hot-swap recovery from permanent node failures.

• 07-23 中

  PHOENIX：让大模型训练在故障后 40 秒内复活，而不是整个作业重启

  PHOENIX 用每步在内存中的异步复制和在线拓扑修复,取代了大规模LLM训练中传统的checkpoint-restart机制,实现了正常训练零可测量开销、故障后40秒内热替换恢复。

• 07-22 EN

  AdaFlash: Taming Variance in Diffusion Speculative Decoding

  AdaFlash finds that diffusion-model drafters for speculative decoding suffer from two hidden variance problems — across domains and across token positions — and fixes both with an online on-policy distillation loop and an adaptive verification-length head, delivering up to 66% higher throughput than prior state of the art at high concurrency.

• 07-22 中

  AdaFlash 阅读笔记：如何驯服扩散式投机解码中的方差问题

  AdaFlash 发现用扩散模型做投机解码的草稿模型，存在跨领域和跨token位置两种隐藏的方差问题；论文用一套在线策略蒸馏和一个自适应验证长度头同时解决这两个问题，在高并发场景下相比此前最优方法吞吐量提升最高约66%。

• 07-21 EN

  PS-PPO: Skipping the Boring Parts of a Trajectory Without Breaking PPO's Math

  PS-PPO observes that reasoning-trace prefixes often already determine the final reward, then builds a convex-optimization-derived, provably unbiased scheme to backpropagate through only a prompt-dependent random prefix of each rollout — cutting RLHF update-stage compute by up to 45% with no accuracy loss.