Page 1 / 28
331 posts in total. Keep on posting.
Showing posts 1–12 of 331. Each entry opens locally on this site; legacy Hexo posts link back to their original article at the bottom for reference.
2026
- EN
LOCUS: Task-Aware Low-Rank Post-Training for Token-Efficient Language Generation
A technical review of how low-rank adaptation can become a search space for concise generation while preserving the original preference objective.
- 中
LOCUS 阅读笔记：用任务感知低秩后训练减少生成 Token
这篇笔记拆解如何把低秩更新空间变成简洁生成的搜索变量，并分析效用约束、实验边界与部署含义。
- EN
Jacap: Preserving Nonlinear Information in KV-Cache Eviction
A derivation-first review of Jacobian capacity, softmax sensitivity, leverage-score selection, and the evidence behind Jacap.
- 中
Jacap 阅读笔记：用雅可比信息容量指导 KV Cache 淘汰
从注意力前置知识出发，逐步推导雅可比容量、Softmax 敏感度与杠杆分数，并分析 Jacap 的实验边界。
- EN
Cliff: Learning Process Rewards from the First Mistake
A close read of Cliff, a reward-shaping method that turns an LLM teacher's judgment of where a rollout first goes wrong into a two-segment, token-level advantage signal for GRPO-style RLVR training.
- 中
Cliff 阅读笔记：只找第一个错误就够了——用教师模型给 RLVR 做过程监督
细读 Cliff：一种奖励塑形方法,让 LLM 教师模型判断学生 rollout 第一次出错的位置,把这个判断转化成 GRPO 式 RLVR 训练里的两段式 token 级 advantage 信号,不需要训练额外的奖励模型。
- EN
Strong Drafts Need Compact Memories: A Technical Review of Memory-Augmented Sliding-Window (MASW) Drafting for Long-Context Speculative Decoding
A close technical read of MASW, which equips a strong independent draft model with a compact three-part working memory (sink tokens, exact local window, learned memory slots) so long-context speculative decoding keeps high acceptance length without paying full-KV-cache draft latency.
- 中
强推理草稿也要瘦身：长上下文投机解码的记忆增强滑窗（MASW）阅读笔记
读 MASW 这篇论文：给一个独立的强草稿模型配上三段式紧凑工作记忆（哨兵token、精确局部窗口、可学习记忆槽），让长上下文投机解码既保住高接受率、又不用支付全量KV访存的代价。
- EN
TreeWY: Removing the Memory Wall in Speculative Decoding for Gated DeltaNet Hybrids
A close technical read of TreeWY, which rewrites the gated delta rule as a tree-structured WY transform so that a Gated DeltaNet hybrid model can verify a wide speculative-decoding draft tree with one triangular solve instead of snapshotting a full recurrent state at every draft node.
- 中
TreeWY 阅读笔记：用树结构 WY 变换拆掉 Gated DeltaNet 混合模型投机解码里的内存墙
这篇论文把 Gated DeltaNet 的门控 delta 规则重新写成树结构的 WY 变换，让投机解码在验证一整棵很宽的草稿树时只需要做一次三角方程求解，而不必在每个草稿节点上都完整保存一份循环状态快照。
- 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.
- 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。