291 posts in total. Keep on posting.

2026

• 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.

• 07-21 中

  PS-PPO 阅读笔记:跳过轨迹里「已经确定」的部分,但不破坏 PPO 的数学

  PS-PPO 发现推理轨迹的前缀往往已经能决定最终奖励,于是构造了一套基于凸优化、可证明无偏的方案,让每次梯度更新只对轨迹中一个与 prompt 相关的随机前缀做反向传播——在不损失准确率的前提下,把 RLHF 更新阶段的算力开销削减最多 45%。

• 07-20 EN

  SEED: When Hindsight Becomes a Training Signal, Not Just a Better Prompt

  SEED turns completed agentic-RL trajectories into self-generated, self-evolving hindsight skills and distills their behavioral effect back into the policy at the token level, lifting ALFWorld success from 75.0 (GRPO) to 91.8 while removing all skill prompts at inference time.

• 07-20 中

  SEED 阅读笔记：当「事后诸葛亮」变成训练信号，而不只是更好的提示词

  SEED 把已完成的智能体强化学习轨迹转化为自我生成、自我进化的「事后经验技能」，并把这些技能对策略的行为影响蒸馏回策略本身，在 ALFWorld 上把成功率从 GRPO 的 75.0 提升到 91.8，同时推理阶段完全不再需要任何技能提示。

• 07-17 EN

  LACE-SVD: Why Uniform Rank Budgets and Local Reconstruction Are Not Enough for SVD Compression

  LACE-SVD shows that SVD-based LLM compression stops being about matrix approximation and becomes a budget-allocation and error-propagation problem once compression gets aggressive, and that tying rank choice to calibration loss plus correcting for cumulative residual-stream drift cuts WikiText-2 perplexity at 60% compression from 46.18 (Dobi-SVD) to 32.57 on LLaMA-7B.