Page 2 / 30
349 posts in total. Keep on posting.
Showing posts 13–24 of 349. Each entry opens locally on this site; legacy Hexo posts link back to their original article at the bottom for reference.
2026
- EN
FACTS and CoRS: Which Curvature Should Low-Rank Compression Preserve?
A technical reading of token-local Fisher weighting, constrained rank allocation, and the gap between an accurate curvature approximation and useful compression.
- 中
FACTS 与 CoRS 阅读笔记：低秩压缩究竟该保留哪一种曲率？
从加权 SVD 的推导出发，理解 token 局部 Fisher 统计、跨层秩分配，以及曲率近似精度与压缩质量之间的距离。
- EN
Deep Delta Learning: Editing the Residual Stream, and Accounting for the Cost
A technical review of depth-wise delta updates, their conditional geometry, expanded residual states, and the unresolved quality-versus-compute tradeoff.
- 中
Deep Delta Learning 阅读笔记：怎样改写残差流，以及这次改写要付出什么代价
从读—比较—写的局部几何出发，推导 DDL 的深度方向 delta 更新，再结合扩展状态、压缩器和实测吞吐，讨论质量收益的来源与边界。
- EN
Gated Attention: Controlling What a Softmax Head Writes
A technical review of query-dependent output gating, its algebra and training evidence, and the limits of claims about sparsity, attention sinks, and long-context efficiency.
- 中
Gated Attention：让注意力头决定写入多少信息
从输出门控的数学结构出发，理解非线性、激活稀疏、训练稳定性和长上下文实验，并区分论文证据与尚待验证的机制解释。
- EN
Progressive Point Matching: Dense Credit for Long-Horizon Language Model RL
A technical review of PPM: reasoning graphs, shortcutting, gradient signal-to-noise, long-budget evaluation, and the gap between ideal reward guarantees and practical judges.
- 中
PPM 阅读笔记：长程强化学习，怎样给尚未成功的推理记分
从推理点、捷径补分和梯度信噪比推导 PPM，分析长预算数学实验、奖励配方，以及理想定理与实际评审模型之间的边界。
- EN
GEPA: Learning Better Agent Instructions from Execution Feedback
A technical reading of GEPA's reflective prompt evolution, instance-wise selection and module merging, with careful treatment of rollout budgets, transfer and failure cases.
- 中
GEPA 阅读笔记：把 Agent 的失败轨迹变成更好的提示词
从反馈信息、候选选择和模块合并理解 GEPA，并细看 35 倍 rollout、跨模型迁移与失败案例各自成立的条件。
- EN
mHC: Stable Residual Mixing, Its Guarantees, and Its Costs
A technical review of manifold-constrained hyper-connections, from doubly stochastic mixing and finite Sinkhorn iterations to memory traffic, experiments, and unresolved stability questions.
- 中
mHC 阅读笔记：残差混合如何稳定，保证又止于哪里
从双随机矩阵与 Sinkhorn 迭代出发，推导 mHC 的残差传播性质，分析内存与通信代价，并讨论信息收缩、动态路由梯度和实验边界。