Page 1 / 36
432 posts in total. Keep on posting.
Showing posts 1–12 of 432. Each entry opens locally on this site; legacy Hexo posts link back to their original article at the bottom for reference.
2026
- EN
CAAS | Scheduling LLM agent batches across inference and tools
A technical review of hybrid-cache sizing, trajectory-aware inference, and online tool dispatch for reducing agent batch makespan.
- 中
CAAS 阅读笔记|怎样让 LLM 智能体的一整批任务更早完成?
从混合模型缓存、跨轮路由和工具队列三个切面,理解样本并发如何缩短或拉长整批任务的耗时。
- EN
Spectral Rank and Step-Direction Dynamics in Language-Model Post-Training: Technical Review
I work through why a small stable rank can hide a broad update spectrum and what checkpoint cosines reveal about post-training.
- 中
论文阅读笔记|语言模型后训练中的谱秩与步进方向
从 OLMo 的检查点出发,理解稳定秩为何低估截断预算,以及相邻参数位移转向时究竟测到了什么。
- EN
DictLLM: Shared Dictionaries for Post-Training Compression
How input-scaled SVD, sparse coefficients and NVFP4 change the trade-off between retained quality, stored bytes and decode speed.
- 中
DictLLM 阅读笔记:共享字典、稀疏系数与低比特压缩
从激活加权分解推导到实际解码,拆开低秩、剪枝与 FP4 各自节省的成本。
- EN
onóna: Giving a Tied Language Model Two Output Distributions
Our new arXiv paper studies a small change to tied embeddings: two shared vocabulary tables, one input-token gate, and an unchanged Transformer body.
- EN
Only Accepted Tokens Count: The Energy Cost of Speculation
An energy ledger for speculative decoding, from accepted-prefix length and verification work to GPU balance points and serving latency.
- 中
推测解码的能耗账:一次验证要产出多少 token 才划算
从接受前缀、验证宽度和 GPU 平衡点出发,理解批量和延迟约束下的推测解码能耗。
- EN
Library of Layers: Routing Whole Transformer Blocks at Runtime
A technical review of prefix-causal window routing, shared transformer blocks, and the evidence for controlling learned execution paths.
- 中
Library of Layers 阅读笔记:Transformer 的运行时选层
从因果窗口、整块路由和残差门控出发,拆解 LoL 如何复用参数,以及路径干预究竟验证了什么。
- EN
Training–Inference Mismatch in LLM RL: What DLC Corrects
A technical review of discrepancy-lag correction, from prefix-level score centering to an effective-sample-size trigger for refreshing the rollout policy.