331 posts in total. Keep on posting.

2026

• 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.

• 07-17 中

  LACE-SVD 阅读笔记：均匀秩预算和局部重建为什么不够用了

  LACE-SVD 指出，当 SVD 压缩比例足够激进之后，问题就不再是「怎么把每个矩阵近似好」，而变成了「怎么把有限的参数预算分给不同层」和「怎么防止误差沿残差流累积放大」；把秩分配与校准损失挂钩、再对残差流误差做门控修正，把 LLaMA-7B 在 60% 压缩比下的 WikiText-2 困惑度从 Dobi-SVD 的 46.18 降到了 32.57。

• 07-16 EN

  GIFT: Why the Coordinate System You Quantize In Matters More Than the Quantizer

  GIFT shows that FP8 gradient communication error in LLM pretraining is partly a coordinate-system problem, not just a numerical-format problem, and fixes it by quantizing gradients in a cheap K-FAC-whitened space instead of raw Euclidean coordinates.

• 07-16 中

  GIFT 阅读笔记:量化用的坐标系,比量化器本身更重要

  GIFT 指出 LLM 预训练里 FP8 梯度通信的误差,有一部分根本不是数值格式的问题,而是坐标系的问题——把梯度先变换到一个用 K-FAC 曲率近似「白化」过的坐标系里再量化,比直接在原始欧氏坐标里量化更保真。

• 07-15 EN

  COBS: What Block-Sparse Attention Selectors Are Actually Missing (A Second-Order Fix)

  COBS shows that every existing cacheable block selector for sparse attention (NSA, CSA, mean-pool, Quest) scores blocks using only a first-order approximation of the true attention mass, and recovers most of the gap to dense attention by caching a compressed second-order (covariance) term instead.

• 07-15 中

  COBS 阅读笔记:块稀疏注意力的选择器到底漏掉了什么(一个二阶修正)

  COBS 指出现有所有可缓存的块选择器(NSA、CSA、mean-pool、Quest)对注意力质量的估计都只用到了一阶信息;论文把块内 key 协方差的二阶项找回来,在长上下文检索基准上补齐了大部分与稠密注意力的差距。

• 07-14 EN

  RIPO: Why PPO's Ratio Clipping Is the Wrong Metric, and What Riemannian Geometry Fixes

  RIPO shows PPO-Clip silently measures policy change with the wrong (Euclidean) metric, so it over-constrains rare exploration tokens and under-constrains dominant ones — and fixes this with a clipping rule derived directly from the policy's Riemannian (KL-induced) geometry, beating GRPO by up to 60% on AIME24.

• 07-14 中

  RIPO 阅读笔记：PPO 的比率裁剪为什么量错了尺度,黎曼几何给出的修正

  RIPO 指出 PPO-Clip 隐含地用欧氏距离衡量策略变化,导致稀有探索 token 被过度压制、主导 token 又被约束不足;论文用策略的黎曼(KL 诱导)几何重新推导出裁剪边界,在 AIME24 上比 GRPO 最高提升 60%。

• 07-13 EN

  Long-Horizon-Terminal-Bench: Dense Reward Grading Exposes Agent Long-Horizon Execution Limits

  LHTB is a 46-task benchmark that replaces binary pass/fail with subtask-level dense rewards, revealing that even the strongest frontier models (GPT-5.5 at 15.2%) cannot reliably finish long-horizon terminal workflows — with 79% of failures caused by time-budget exhaustion rather than local reasoning errors.