Page 7 / 36
432 posts in total. Keep on posting.
Showing posts 73–84 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
SAES-SVD: Compensating for the Errors That Earlier Layers Leave Behind
A technical review of cumulative-error-aware low-rank compression, its closed-form solution, and the limits of spectral coefficient selection.
- 中
SAES-SVD 阅读笔记:让后面的层补偿前面的压缩误差
从双目标回归理解累积误差补偿,推导白化 SVD 与自适应系数,并辨清局部最优、模型质量和推理加速的边界。
- EN
Pie: Who Should Control the LLM Generation Loop?
A technical reading of programmable LLM serving, explicit KV ownership, API-level batching, and the costs of moving generation policy into applications.
- 中
Pie 阅读笔记:谁来控制大模型的生成循环
从 KV 缓存、命令队列和工具调用理解可编程推理服务,并分析应用控制权带来的性能收益、调度成本与语义边界。
- EN
TTT-E2E: Learning a Long Context into Weights
A technical reading of causal test-time updates, meta-gradients, memory capacity, and the gap between predictive compression and exact recall.
- 中
TTT-E2E 阅读笔记:把长上下文写进权重
从因果更新、元梯度和局部注意力出发,理解测试时训练的预测收益、精确检索边界及真实系统成本。
- EN
SICA: When a Coding Agent Becomes Its Own Next Engineer
A technical reading of self-referential agent improvement, with utility derivations, context design, benchmark reconstruction, and the limits of recursive gains.
- 中
SICA 阅读笔记:让代码 Agent 成为自己的下一位工程师
从效用函数、上下文组织和逐轮实验出发,理解 SICA 如何改进自己的 harness,以及自我改进离递归能力增长还有多远。
- EN
Darwin Gödel Machine: How a Coding Agent Learns to Improve Its Own Harness
A technical reading of DGM, from archive-based self-modification and parent selection to evaluation leakage, transfer, and the limits of recursive improvement.
- 中
Darwin Gödel Machine 阅读笔记:让智能体改进自己的 Harness
从档案搜索、父代采样和分阶段评估理解 DGM,并分析自我改进的证据、泛化边界与评分失真。
- EN
SVDQuant: How a Small High-Precision Branch Makes 4-Bit Diffusion Practical
A technical review of outlier migration, low-rank protection, quantization error, and the memory traffic behind Nunchaku.
- 中
SVDQuant 阅读笔记:用少量高精度计算保护 4-bit 扩散模型
从异常值迁移、低秩分解和量化误差推导出发,理解 Nunchaku 为什么必须同时处理计算与显存搬运。